null  null
On looking into
words (and beyond)
Structures, Relations, Analyses
Edited by
Claire Bowern
Laurence Horn
Raffaella Zanuttini
Empirically Oriented Theoretical
Morphology and Syntax 3
Empirically Oriented Theoretical Morphology and Syntax
Chief Editor: Stefan Müller
Consulting Editors: Berthold Crysmann, Laura Kallmeyer
In this series:
1. Lichte, Timm. Syntax und Valenz: Zur Modellierung kohärenter und elliptischer Strukturen
mit Baumadjunktionsgrammatiken
2. Bîlbîie, Gabriela. Grammaire des constructions elliptiques: Une étude comparative des
phrases sans verbe en roumain et en français
3. Bowern, Claire, Laurence Horn & Raffaella Zanuttini (eds.). On looking into words (and
beyond): Structures, Relations, Analyses
ISSN: 2366-3529
On looking into
words (and beyond)
Structures, Relations, Analyses
Edited by
Claire Bowern
Laurence Horn
Raffaella Zanuttini
Claire Bowern, Laurence Horn & Raffaella Zanuttini (eds.). 2017. On looking into words
(and beyond): Structures, Relations, Analyses (Empirically Oriented Theoretical
Morphology and Syntax 3). Berlin: Language Science Press.
This title can be downloaded at:
© 2017, the authors
Published under the Creative Commons Attribution 4.0 Licence (CC BY 4.0):
ISBN: 978-3-946234-92-0 (Digital)
978-3-946234-97-5 (Hardcover)
978-3-946234-81-4 (Softcover)
ISSN: 2366-3529
Cover and concept of design: Ulrike Harbort
Typesetting: Claire Bowern, Parker Brody, Jun Chen, Dolly Goldenberg, Maggie
Kandel, and Jason Zentz
Proofreading: Claire Bowern, Laurence Horn, and Sharon Klein
Editors: Claire Bowern, Laurence Horn, and Raffaella Zanuttini, in collaboration with
Ryan Bennett, Bob Frank, Maria Piñango, Jason Shaw, and Jim Wood
Fonts: Linux Libertine, Arimo, DejaVu Sans Mono
Typesetting software: XƎLATEX
Language Science Press
Unter den Linden 6
10099 Berlin, Germany
Storage and cataloguing done by FU Berlin
Language Science Press has no responsibility for the persistence or accuracy of URLs
for external or third-party Internet websites referred to in this publication, and does not
guarantee that any content on such websites is, or will remain, accurate or appropriate.
Between natural and unnatural phonology: The case of cluster-splitting
Juliette Blevins
The domain of stress assignment: Word-boundedness and frequent
Ellen Kaisse
Compensatory lengthening and structure preservation revisited yet again
Darya Kavitskaya
4 Phonological exceptionality is localized to phonological elements: the
argument from learnability and Yidiny word-final deletion
Erich R. Round
U-umlaut in Icelandic and Faroese: Survival and death
Höskuldur Thráinsson
II Morphology
Word-based Items-and-processes (WoBIP): Evidence from Hebrew
Outi Bat-El
Root-based syntax and Japanese derivational morphology
Brent de Chene
8 Morphological complexity and Input Optimization
Michael Hammond
Multiple exponence in the Lusoga verb stem
Larry M. Hyman & Sharon Inkelas with Fred Jenga
10 Romansh allomorphy (Again!)
Martin Maiden
11 How to wake up irregular (and speechless)
Charles Yang
III Syntax and Morphosyntax
12 Special clitics and the right periphery in Tsotsil
Judith Aissen
13 Another way around causatives in Chamorro
Sandra Chung
14 Preliminaries to the investigation of clitic sequencing in Greek and
Mark Hale
15 Nominal verbs and transitive nouns: Vindicating lexicalism
Paul Kiparsky
16 On mechanisms by which languages become [nominative-]accusative
Ashwini Deo
17 OF as a phrasal affix in the English determiner system
Randall Hendrick
18 Split-morphology and lexicalist morphosyntax: The case of transpositions
Andrew Spencer
19 Rules and blocks
Gregory Stump
IV Language and Linguistic Theory
20 Darwinism tested by the science of language
Mark Aronoff
21 Temporal patterning in speech and birdsong
Louis Goldstein
22 “Constructions” and grammar: Evidence from idioms
Julia Horvath & Tal Siloni
23 A-morphous iconicity
Ryan Lepic & Carol Padden
24 Iconicity chains in sign languages
Donna Jo Napoli
25 Where, if anywhere, are parameters? A critical historical overview of
parametric theory
Frederick J. Newmeyer
26 Saussure’s Dilemma: Parole and its potential
Alan Timberlake
While linguistic theory is in continual flux as progress is made in our ability to understand the structure and function of language, one constant has always been the central
role of the word. On Looking into Words is a wide-ranging volume spanning current
research into word-based morphology, morphosyntax, the phonology-morphology interface, and related areas of theoretical and empirical linguistics. The 26 papers that constitute this volume extend morphological and grammatical theory to signed as well as
spoken language, to diachronic as well as synchronic evidence, and to birdsong as well
as human language. Central concerns of the volume’s authors include morphological
gaps, learnability, increases and declines in productivity, and the interaction of different
components of the grammar.
The contributions in the volume explore linked synchronic and diachronic topics in
phonology, morphology, and syntax (in particular, inflection and cliticization) and their
implications for linguistic theory, and are informed by data from a wide range of language families. Languages discussed in the contributions include Ancient and Modern
Greek (Hale, Kavitskaya), Macedonian (Kaisse), Sanskrit, Finnish, and Sakha (Kiparsky),
Middle Indo-Aryan (Deo), Rumantsch (Maiden), Latin and French (Timberlake), Russian
(Spencer), Icelandic and Faroese (Thráinsson), Welsh (Hammond), Hebrew (Bat-El, Horvath & Siloni), Limbu and Southern Sotho (Stump), Lusoga (Hyman & Inkelas), Yidiny
(Round), Japanese (de Chene), and Tsotsil Mayan (Aissen), as well as ASL and other
signed languages (Lepic & Padden, Napoli), as well as English.
The papers are offered as a tribute to Stephen R. Anderson, the Dorothy R. Diebold
Professor of Linguistics at Yale, who is retiring at the end of the 2016–2017 academic
year and who has long been a major figure in several of the fields under investigation
in the volume: phonology and morphology (and the history of scholarship on those
topics), the of morphology-syntax interface, the relation of theory to data, the evolution
of human language and its relation to animal communication systems. The contributors
– friends, colleagues, and former students of Professor Anderson – are all important
contributors to linguistics in their own right. The central contributions that Anderson
has made to so many areas of linguistics and cognitive science, drawing on synchronic
and diachronic phenomena in diverse linguistic systems, are represented through the
papers in the volume.
The papers in Part I of the volume focus on phonology at the level of the word and
below. Ellen Kaisse draws on evidence from literary Macedonian to investigate why
stress assignment is word-bounded, despite the existence of higher-level prosodic interactions – that is, she discusses the reasons why lexical stress is specifically lexical. Dasha
Kavitskaya revisits a robust generalization from work by de Chene and Anderson that
predicts compensatory lengthening should be structure preserving. She argues that a
class of apparent exceptions to this generalization have independently motivated analyses consistent with the generalization. Juliette Blevins looks at the role of language contact in determining cross-linguistic strategies for the resolution of stop + liquid clusters
through vowel epenthesis as a recurrent feature of loan-word phonology. Erich Round’s
contribution addresses issues of learnability in determining the class of possible grammars. He discusses the implications of word-final deletion in Yidiny (Pama-Nyungan) for
the theory of exceptionality, which he shows cannot be represented in purely phonological or purely diacritic terms. Like several papers in the collection, Höski Thráinsson’s
examines the diachronic underpinnings of synchronic linguistic structure. He uses contrastive evidence from alternations in Modern Icelandic and Modern Faroese to argue for
different trajectories in historical change in u-umlaut and accounts for differing degrees
of synchronic productivity and transparency of the umlaut rules in the two languages.
The papers in Part II concentrate on issues in morphological theory, in several cases
also drawing on the interaction of synchronic and diachronic factors. Martin Maiden
argues for the use of comparative and historical evidence in deciding between possible synchronic interpretations of phenomena. He revisits paradigmatic phenomena in
Rumantsch and investigates whether allomorphy in verb stems is phonologically conditioned or is evidence for paradigm autonomy. Brent de Chene’s paper looks at an
apparent argument for a syntactic approach to derivational stem morphology (as proposed in the theory of Distributed Morphology) in languages like Japanese and shows
that a closer examination of the facts supports the approach defended by Anderson on
which it is just inflectional (and not derivational) processes that lend themselves to a syntactic analysis. Anderson’s A-Morphous Morphology is an item-and-process approach
that takes words to be related to each other by morphological processes; Semitic languages, with their celebrated triconsonantal roots, were explicitly excluded from the
purview of Anderson’s analysis. In her contribution, Outi Bat-El – invoking data from
both adult and child language – argues that in fact the Anderson-style item-and-process
framework can be naturally extended to provide an empirically grounded account of
word relations in Modern Hebrew, with the word/stem taken as basic and the putative
triconsonantal root seen as a mere residue of phonological elements. Mike Hammond
applies data from English and Welsh to the examination and measurement of morphological complexity, applying the statistical framework of Input Optimization developed
by Hammond himself in earlier work to questions arising in morphological investigation.
Larry Hyman and Sharon Inkelas (with Fred Jenga) address a different type of question
for morphological theory, the nature of and motivation for multiple exponence. Lusoga,
a Bantu language spoken in Uganda, exhibits multiple exponence of causative, applicative, subjunctive and perfective suffixes and the apparent intermingling of derivational
and inflectional affixation processes, properties which Hyman & Inkelas see as posing
challenges for theories of morphology that seek to minimize redundancy and treat derivation as distinct from (and invariably ordered before) inflection. Finally, Charles Yang’s
paper examines a different issue for morphological theory, that of morphological defectiveness arising when a productive process fails to be recognized by the language learner.
He reconsiders the *amn’t problem – the absence (in most dialects of English) of an inflected negation for the first person singular copula – through the lens of his Tolerance
Principle, a corollary of the Elsewhere Principle motivated by the behavior of speakers
with respect to paradigm gaps in English, Russian, Spanish, and other languages. As
in other papers in this section (and in the volume more generally), Yang extends his
view to consider the implications of the processes he describes for language acquisition,
variation, and change.
Part III encompasses a set of issues relating morphology to syntax, drawing on complex synchronic and diachronic evidence from a diverse set of languages. Sandy Chung’s
paper returns to the interaction of causative marking and number agreement in Chamorro, which has been recognized as a challenge for the traditional view that inflectional
affixes invariably attach outside derivational ones. While Anderson proposed taking
number agreement to be derivational, Chung argues that the apparent causative prefix
is actually a prosodically deficient verb, thus dissolving the apparent counterexample to
the inflection-outside-derivation while preserving the intuition that number agreement
is inflectional. Another contribution to morphosyntactic investigation in this section is
Randy Hendrick’s argument for analyzing the English preposition of (appearing in degree inversion expressions like more complex of a theory) as a phrasal affix rather than
a syntactic head, thus predicting certain distributional features of the construction and
allowing an explanation of the appearance of of in fractional expressions (one third of a
Several of the papers in this section examine the nature of clitics. Mark Hale investigates the syntax underlying the distribution of clitics that fall under Wackernagel’s Law,
which describes the tendency of clitics to appear in second (or, now, “Wackernagel”)
position in the sentence. Utilizing diachronic data from Ancient Greek and Indo-Iranian,
Hale provides a detailed examination of cases in which there are multiple demands on a
Wackernagel position and of the devices for resolving those conflicts, ultimately taking
Wackernagel’s Law to be epiphenomenal on other phenomena. Judith Aissen similarly
considers exceptional clitic placement, but in Tsotsil (Mayan); while Hale argues for
syntactic constraints on exceptionality based primarily on Attic Greek, Aissen focuses
on the role of phonology, and in particular prosodic constraints, in the determination
of the placement of clitics on the right periphery (rather than in the syntactically conditioned second position). Like Hale, Deo and Kiparsky examine diachronic shifts in
Indo-Aryan, albeit toward different ends. Ashwini Deo’s article traces the morphosyntax of ergative alignment loss in Indo-Aryan. Anderson and others have While much
work, including a seminal paper by Anderson, has sought to describe how languages
develop ergative case marking, Deo turns the tables by seeking to develop an understanding of how nominative-accusative marking can emerge in historically ergative systems, as in the gradual de-ergativization of Middle Indo-Aryan based on independent
shifts in argument realization in that language. Paul Kiparsky leverages a synchronic
analysis of gerunds and nominalizations across a variety of languages (Vedic Sanskrit,
Finnish, Sakha (Yakut), English) to argue for a lexicalist rather than syntactically derived
analysis of agent nominalizations; gerunds, on the other hand, are argued to be verbal
at all levels of the syntactic derivation. Thus, we have a neatly symmetrical system:
while gerunds are verbs that bear a Case feature (via their INFL head), transitive nominalizations (pace Baker) are nouns with an Aspect feature. The final two papers in the
(Morpho)syntax component of the volume bear on the status of inflectional morphology
within the general theory of grammar. Like other authors in the volume, Andy Spencer
takes as a jumping off point Anderson’s “split morphology” hypothesis – the view that
derivational morphology is lexically mediated while inflectional morphology is syntactically mediated. Along with the gerunds and nominalizations explored in Kiparsky’s
contribution, another test case for this view is posed by participles. Spencer investigates
Russian participles and other transpositions utilizing a generalized version of Stump’s
Paradigm Function Morphology. He concludes that while the participle’s actual lexical
entry is that of an adjective (as lexicalist theory predicts), it realizes the verbal properties of voice/aspect, sharing the semantic properties and lexemic index of its verbal base,
just as in the case of verb inflection. In his paper, Greg Stump focuses on the viability of Anderson’s conception of rule block for inflectional processes within a grammar
and on rule interaction. His evidence is provided by multiple exponence in Limbu, a
Tibeto-Burman language of Nepal, and by polyfunctionality in Southern Sotho, a Bantu
language of southern Africa. He develops principles of rule conflation that are independently motivated and converge to explain the outcome when a single rule participates
in more than one block.
The volume concludes with a set of papers that explore linguistic theory from a broad
perspective, while also extending the issues of words and rules from the structure and
evolution of spoken language to the properties of signed languages and birdsong. Mark
Aronoff, who is (along with Anderson) a key figure in the development of word-based
morphology, turns his attention here to the mismatch between the nature of linguistic
rules and that of biological evolution (despite the influence of the results of neogrammarians on Darwin and other prime movers in the establishment of the principles of
natural selection). This mismatch led Saussure to abandon his earlier productive focus on diachrony in favor of a single-minded characterization of synchronic linguistics.
Aronoff borrows Gause’s principle of competitive exclusion from evolutionary ecology
as a means for reviving the evolutionary approaches of Saussure and his 19th century
predecessors. Besides the dichotomy of diachrony and synchrony, Saussure is also celebrated for his distinction between langue and parole, the latter notion often dismissed
by generative linguists as a matter of peripheral interest. Alan Timberlake revisits Saussure’s treatment of the shift in stress rules from Latin to French and points to the crucial
role of parole in accounting for the relevant shifts not only in that case but in the alternations dictating the appearance of intrusive /r/ (the idear of it) and reanalysis (a
napron > an apron) in English. The papers by Newmeyer and Horvath & Siloni reconsider theoretical issues in generative grammar and its rivals. While parameters have
long been adduced in grammatical theory as a means to account for cross-linguistic
variation, Fritz Newmeyer’s paper re-examines the role of parametric explanation and
argues that despite its success in leading generative scholars to the discovery of crosslinguistic generalizations, macro- and micro-parameters rest on shaky ground empiri-
cally and conceptually in current theoretical scholarship. Newmeyer goes on to survey
some promising alternative approaches for capturing the relevant insights. Julia Horvath and Tal Siloni compare two theoretically distinct approaches to idioms. They argue
that the distributional differences between phrasal and clausal idioms with respect to the
diatheses with which they co-occur represent a serious problem for the approach of Construction Grammar, which treats linguistic knowledge as a network of stored constructions, while generative grammar, which allows for a computational system (providing
unstored derivational outputs) as well as a storage module, is more capable of offering
an empirically adequate account of these phenomena.
The of iconicity in sign languages representations is explored in the papers by Lepic &
Padden and Napoli. On the a-morphous theory of word structure proposed by Anderson,
the word and not the morpheme is the fundamental building block of morphology. As
earlier posited by Jackendoff and Aronoff, the task of Word Formation Rules is analysis
rather than creation; the speaker’s knowledge encompasses the complex relationships
among the words in her lexicon. Ryan Lepic and Carol Padden provide further support
for this theory by drawing on the properties of lexical iconicity (and the gradual loss of
iconicity) in American Sign Language. Donna Jo Napoli’s paper also deals with iconicity in non-spoken linguistic modalities, although she extends her domain from ASL to a
wide range of the world’s signed languages. She uncovers biologically motivated “chains
of iconicity” consisting of mappings from perceptual systems into visual realizations that
evolve along semantic lines, a natural development that helps explains why unrelated
signed languages exhibit similar signs for abstract concepts. The final contribution is
by Louis Goldstein, a pioneer in the development of articulatory phonology, a gesturebased approach to the problem of how messages in the mind of the sender are transferred
to the mind of the receiver through actions of the body. In his paper, Goldstein explores
the ways such messages are conveyed by humans (through spoken or signed words) and
by birds; his study demonstrates that despite their real and instructive differences, birdsong, like speech, is a complex motor behavior with an essential role played by temporal
The word has always been at the crux of descriptive and theoretical research on phonology, morphology, syntax, and semantics. By looking into words, the contributors to
this volume shed light on both specific subfields and the interfaces among them, on how
languages work and how they change over time. The papers in this volume, in addressing linguistic structure and linguistic relationships, help further our understanding of
individual languages and of language as a unified phenomenon.
Part I
Chapter 1
Between natural and unnatural
phonology: The case of cluster-splitting
Juliette Blevins
The Graduate Center, CUNY
A widely recognized feature of loan-word phonology is the resolution of clusters by vowel
epenthesis. When a language lacking word-initial clusters borrows words from a language
with initial #TRV sequences, T an oral stop and R a liquid, it is common to find vowel
epenthesis, most typically vowel-copy, as in, for example: Basque <gurutze> ‘cross’ from
Latin <cruce(m)>; Q’eqchi’ <kurus> ‘cross’ from Spanish <cruz> ‘cross’, or Fijian <kolosi>
‘cross’ from English <cross>. The phonological rule or sound change responsible for this pattern is sometimes called “cluster-splitting epenthesis”: #TRVi > #TV(i) RVi . The most widely
accepted explanation for this pattern is that vowel epenthesis between the oral stop and the
following sonorant is due to the vowel-like nature of the TR transition, since #TRVi is perceptually similar to #TV(i) RVi . A fact not often appreciated, however, is that cluster-splitting
epenthesis is extremely rare as a language-internal development. The central premise of
this chapter is that #TRVi in a non-native language is heard or perceived as #TV(i) RVi when
phonotactics of the native language demand TV transitions. Without this cognitive component, cluster-splitting epenthesis is rare and, as argued here, decidedly unnatural.
1 Introduction
Diachronic explanations have been offered for both natural and unnatural sound patterns in human spoken languages. Building on the Neogrammarian tradition, as well
as the experimental research program of Ohala (e.g. 1971; 1974; 1993), it is argued that
natural sound patterns, like final obstruent devoicing, nasal place assimilation, vowel
harmony, consonant lenition, and many others, arise from regular sound changes with
clear phonetic bases (Blevins 2004, 2006, 2008, 2015; Anderson 2016). Unnatural sound
patterns, in contrast, cannot be explained in terms of simple phonetic processes. Case
studies of unnatural patterns show more complex histories: in some cases more than one
natural change has been collapsed; in others, a natural change is reanalyzed or inverted;
Juliette Blevins. 2017. Between natural and unnatural phonology: The case of
cluster-splitting epenthesis. In Claire Bowern, Laurence Horn & Raffaella Zanuttini (eds.), On looking into words (and beyond), 3–16. Berlin: Language Science Press.
Juliette Blevins
in still others, analogy has extended a sound pattern whose origins are morphological,
not phonological; and combinations of all of these paths can also be observed (Bach
& Harms 1972; Anderson 1981; Buckley 2000; Vaux 2002; Blevins 2008b,a,c; Garrett &
Blevins 2009; Anderson 2016).
However, typological study of regular sound change reveals certain kinds of sound
change that are neither wholly natural nor wholly unnatural. For example, the shift
of *#kl > #tl, documented in at least three different language families, appears to have
a natural basis in perception, since [kl] and [tl] clusters are acoustically similar and
confused with each other. However, the rarity of this sound change is associated with
structural factors: misperception of [kl] as [tl] is strongly associated with the absence
of phonological /tl/ clusters in a language. In this case, the absence of a sound pattern,
/tl/, influences cognition, making listeners more likely to perceive [kl] as [tl] (Blevins &
Grawunder 2009). The presence of a contrast can also facilitate particular types of sound
change. As first argued by de Chene & Anderson (1979), compensatory lengthening
sound changes are strongly associated with pre-existing vowel length contrasts. This
statistical tendency is argued to arise from phonetically natural vowel lengthening in
pre-sonorant and open syllable contexts (Kavitskaya 2002), combined with the cognitive
effects of structural analogy, where pre-existing categories act as attractors in the course
of language acquisition (Blevins 2004: 150–155 and Kavitskaya, this volume).
In this contribution, I offer another example of regular sound change that is neither
wholly natural nor wholly unnatural, and highlight the role of human cognition in cases
where it has occurred. Cluster-splitting epenthesis is of interest, not only because of its
rarity as a regular sound change, but in how it advances our understanding of sound
patterns compared to 20th century models (Anderson 1985), and emphasizes the extent
to which historical linguists, phonologists, and phoneticians still need the cognitive scientist (Anderson 2001).
2 Cluster-splitting epenthesis in loanword phonology
A widely recognized feature of loanword phonology is the resolution of clusters by vowel
epenthesis. When a language lacking word-initial clusters borrows words from a language with initial #TRV sequences, T an oral stop and R a liquid, it is common to find
vowel epenthesis, most typically vowel-copy, as illustrated in (1).1
If loanword phonology is taken as evidence for properties of phonological grammars
(Hyman 1970), a phonological rule describing this pattern can be stated as in (2).
For these and other examples, see: Blevins & Egurtzegi (2016) on Latin loans in Basque; Casparis (1997)
on Sanskrit loans in Indonesian; Campbell (2013) on Colonial Spanish loans into Mayan languages; and
Kenstowicz (2007) on Fijian loanword phonology. A reviewer notes that some varieties of Q’eqchi’ permit
initial /CR/ clusters. Proto-Mayan had only *CVC syllables, and Mayan kurus arguably reflects borrowing
of Colonial Spanish cruz into a language which lacked initial CR clusters.
1 Between natural and unnatural phonology: The case of cluster-splitting epenthesis
Cluster-splitting epenthesis in loanword phonology
a. Source language Latin
Target language Basque
gurutze ‘cross’
Source language
Target language
Source language
Target language
Source language
Target language
(2) Cluster-splitting vowel-epenthesis
#TRVi → #TV(i )RVi
Within the structuralist and generative traditions detailed in Anderson (1985), the locus of explanation for this type of epenthesis lies in phonotactic differences between the
source language and the target language. Under this general account, the speaker of the
target language hears a word pronounced in the source language, constructs a phonological representation with an initial #TR cluster based on this hearing, but then alters
this phonological representation in line with the phonotactics of the speaker’s native
language which lacks initial #TR clusters (e.g. Broselow 1987; Itô 1989).
Typological studies of loanword phonology and advances in our understanding of
speech perception have given rise to 21st century treatments of these patterns that are
more explanatory in accounting not only for a “repair” but for the specific type of sound
pattern that results. At present, the most widely accepted explanation for the sound
pattern in (2) combines two new findings in speech perception, one related to perceptual similarity, and the other related to perceptual illusions. A first component of the
analysis is that vowel-epenthesis between the oral stop and following sonorant is due to
the vowel-like nature of the TR transition (Fleischhacker 2001; 2005; Kang 2011; Berent
2013; Broselow 2015). Fleischhacker (2001; 2005) argues that the general pattern is determined by perceptual similarity: initial TR clusters are more perceptually similar to
TVR than VTR. An important aspect of her work is the distinction between initial #TR
clusters and initial #sT clusters, which rarely show vowel-splitting epenthesis, and defy
purely phonotactic accounts. A second component of the analysis relates to specific
structural differences between the source and target languages. Under the perceptual account, perception of #TR by native speakers of languages that lack initial #TR is biased:
these speakers will tend to hear a vowel between the oral stop and the following liquid,
even if no vowel is present. Experimental work supporting the existence of illusionary
vowels for Japanese speakers presented with CC clusters was presented in Dupoux et al.
(1999), and has been supported by much subsequent work (see Peperkamp & Dupoux
Juliette Blevins
2002; Kang 2003; 2011; Kabak & Idsardi 2007; Davidson & Shaw 2012), including a range
of studies showing vowel percepts in TR clusters (Berent 2013, and works cited there).2
Given this evidence, one might conclude that the sound pattern described in (2) is
both natural and common, having a clear phonetic explanation (Blevins 2004; 2008b;
2015). As a natural, phonetically-based sound pattern one might expect many instances
of reconstructed word-initial *TR clusters to be resolved by a sound change parallel to
the synchronic rule in (2). A sound change of this kind might be even more common
than expected on phonetic grounds due to markedness proposals stating that complex
onsets are less preferred than simple onsets (Prince & Smolensky 1993; Kager 1999). As I
show below, these expectations are not borne out, suggesting that cognitive bias in the
context of novel stimuli plays a central role in cluster-splitting epenthesis.
3 Cluster-splitting epenthesis as regular sound change
Very few well-studied and widely agreed upon proto-languages are reconstructed with
initial *TR clusters at the oldest stages. One exception is Proto-Indo-European, reconstructed with *TR clusters as well as other initial cluster types (Fortson 2010: 64–65).
Some widely agreed upon Proto-Indo-European reconstructions with initial *TR clusters are shown in (3).
Word-initial *TR in Proto-Indo-European
a. *gras- ‘eat’
Cf. Vedic grásate ‘eats, feeds’, Greek grástis ‘green
fodder, grass’, Latin grāmen (< *gras-men) ‘grass, fodder’
b. *prekj - ‘ask’
Cf. Vedic pṛccháti ‘asks’, Latin precor ‘I entreat’, German fragen, Tocharian B prek-
c. *trejes ‘three’
Cf. Lycian tri-, Vedic tráyas, Greek treĩs, Avestan
θrayō, Latin trēs
The Indo-European language family is relatively large, relatively diversified, and relatively well-studied in comparison with other language families of the world. According
to Ethnologue, there are approximately 445 living Indo-European languages at present,
and linguists agree that the major subgroups of Anatolian, Indo-Iranian, Greek, Italic,
Celtic, Armenian, Tocharian, and Balto-Slavic have had long independent developments.
If cluster-splitting vowel-epenthesis in (2) has a natural phonetic basis in perceptual similarity as outlined above, then a sound change like (3) might be expected to have occurred
numerous times in the Indo-European language family.
(4) Cluster-splitting vowel-epenthesis as sound change
#TRVi > #TV(i )RVi
This approach is not agreed upon by all researchers. See Uffmann (2007) and Hall (2011) for discussion.
1 Between natural and unnatural phonology: The case of cluster-splitting epenthesis
However, cluster-splitting vowel epenthesis as a regular sound change is rare in the
Indo-European language family. *TR clusters are inherited intact in all of the major
subgroups, and sound changes affecting these clusters at later stages of development are
of distinct types (e.g. palatalization of *l in Romance *Tl clusters; loss of *p in Celtic).
Indeed, within the entire Indo-European language family, there appears to be only
one clear instance of a regular sound change like (3).3 The sound change in question
appears to have occurred in relatively recent times, in the transition from Middle Persian to Modern Persian (aka New Persian, Farsi, Dari, Tajiki), or, perhaps more generally,
from Middle to Early New Iranian.4 While the specific sound change is rarely stated as
in (3), it is implied. For example in their chapter on modern Persian and Tajik phonology,
Windfuhr & Perry (2009: 427–428) describe the language as having syllable onsets consisting of only a single consonant, and note that “The inherited initial clusters have been
resolved by prothetic or epenthetic vowels, either of which could become standardized,
e.g. st-: stār ‘star’ > setāre/sitora, br: brādar ‘brother’ > barādar/barodar…” (Windfuhr &
Perry 2009: 428). The epenthesis process described is identical to that schematized in (3),
and it is also characteristic in loanword phonology, on which much more has been written (see, e.g. Strain 1968, Karimi 1987). Illustrative examples comparing Middle Persian
inherited clusters to Modern Persian #CVC sequences are shown in (5).
(5) One case of cluster-splitting epenthesis in Indo-European: Modern Persian
Middle Persian
a. brādar
b. griftan
c. draxt
d. griy-
Modern Persian
‘grab, take’
*bh réh2 ter*gh rebh2 -
‘to cry’
*drew- ‘wood’
*gh reh2 d-
A second case of cluster-splitting epenthesis sound change is found in the SiouanCatawba language family, a small group of languages in North America that includes
Crow, Hidatsa, Mandan, Lakota, Dakota, Assiniboine, Yanktonai, Stoney, Sioux Valley,
Chiwere (aka Iowa-Missouria-Otoe), Hoocąk (aka Winnebago), Omaha-Ponca, Ponca,
Kanza/Kaw, Osage, Quapaw, Biloxi, Ofo, Tutelo, Saponi, Catawba and Woccon. The
diachronic process known as Dorsey’s Law (Dorsey 1885) is a sound change taking Proto3
Fortson (2010: 302–303) mentions evidence for a sound change similar to cluster-splitting epenthesis in
Oscan, an extinct Italic language known from inscriptions from approximately 400 BC - 100 CE, as in
aragetud ‘with money’ (cf. Lat. argentō) and sakarater ‘it is consecrated’ (cf. Lat. sacratūr). However, the
*rg cluster continued in aragetud was arguably heterosyllabic R.T (as opposed to tautosyllabic TR) and
initial TR clusters are continued intact in Oscan as in trístaa, trííbúm, prúfatted (op cit.).
4 In his discussion of East and West Iranian dialectology, Windfuhr (2009: 21) states the reflexes of initial
#CC-clusters as showing a distinct areal distribution: “insertion of a short vowel, CVC-, along the Zagros,
including the NW tier I from Kurdish, Zazaki to the SW Fars and Larestan dialects, as opposed to initial
vowel, VCC-, elsewhere”, while in the east, Balochi and most East Iranian languages allow initial clusters.
Juliette Blevins
Siouan *TRV to #TVi RVi in Hoocąk (aka Winnebago).5 Examples from Rankin et al.
(2015) are shown in (3).
One case of cluster-splitting epenthesis in Siouan: Dorsey’s Law in Hoocąk
a. églųñį
b. glé
c. wa/brú
ru/purú ‘plough’
‘go back to’
While the time-depth of Siouan-Catawba is thought to be 2,000–3,000 years (Parks &
Rankin 2001), Hoocąk and Chiwere are considered to be closely related and even sometimes treated as dialects of a single language (Miner 1979).6 Given this, Dorsey’s Law
must be a relatively recent development.
Outside of the Persian and Hoocąk cases, it is difficult to find convincing cases of
cluster-splitting epenthesis as a diachronic development. And here lies the central point
of interest. Given that cluster-splitting epenthesis is common in loanword phonology
(2), and appears to be a natural phonetically-motivated process, why is it rarely attested
as a regular sound change? Why, out of more than 440 Indo-European languages, is
there only one clear case of a #TRVi > #TV(i )RVi sound change? And how should we
understand the Siouan sister-languages Chiwere and Hoocąk, where Chiwere continues
#TRV, but Hoocąk does not?
I suggest that cluster-splitting epenthesis is neither wholly natural nor wholly unnatural: non-phonetic structural and cognitive factors are involved. The structural condition
is that cluster-splitting epenthesis occurs only when speakers of a language that lacks
initial TR clusters begin to acquire a language that has initial TR clusters. It is only under
this circumstance that the perceptual illusion of #TRV as #TVRV arises (cf. Dupoux et al.
1999), with this perceptual illusion constituting the cognitive catalyst for phonological
change. An important component of this model is that regular sound changes of this
kind will only occur under special types of language contact, where speakers dominant
in a language that lacks initial consonant clusters suddenly (or without extensive exposure) acquire a language with #CR-clusters.7 If extensive exposure occurs, perceptual
illusions of phantom vowels will weaken, lowering the probability of epenthesis as regular sound change. Let us now evaluate this proposal with respect to the two cases of
diachronic cluster-splitting epenthesis documented above.
Dorsey’s Law also refers to the resulting synchronic sound pattern in Hoocąk. It also applies to medial
clusters. Since the syllabification of medial TR is ambiguous cross-linguistically, discussion is limited here
to initial #TR where, at least utterance-initially, sequences constitute unambiguous complex onsets.
Miner (1979: 25) begins his article with the statement that: “Winnebago and Chiwere … are, in spite of
their geographical separation in historical times, very closely related and enjoy a high degree of mutual
comprehensibility.” He also notes on the same page (footnote 1) that “Winnebago-Chiwere is sometimes
referred to in the literature simply as Chiwere.”
For a similar proposal regarding paragoge (final vowel insertion), see Ng (2015), a dissertation supervised
by Steve Anderson.
1 Between natural and unnatural phonology: The case of cluster-splitting epenthesis
Modern Persian phonology has had significant influence from Arabic and Turkic. Arabic loans constitute about half of the lexicon, and some estimate that of the most frequent
vocabulary, at least 25% is Arabic (Perry 2004; 2005). Turkic loans also exist and there is
a long history of Persian-Turkic bilingualism as well as Turkic “Persianization”. Could
acquisition of Persian by Arabic or Turkic speakers be the source of Modern Persian
cluster-splitting epenthesis? I believe the answer is yes. More specifically, I suggest
that the Persianization of Turkic people, such as the one occurring during the Ghaznavid dynasty (977–1186), and extending over large parts of Iran, was a critical factor in
the evolution of cluster-splitting epenthesis in Modern Persian. Turkic languages have
phonotactics that appears to be most important in triggering cluster-splitting-epenthesis:
they disallow complex onsets in word-initial position (and elsewhere). Under this scenario, Middle Persian underwent rapid phonological change, as it was acquired by native speakers of Turkic languages across Iran. How early the process began is unknown,
though it could have begun as early as the 10th century when Turkic speakers came to
the area, or in the 11th and 12th centuries, when a large migration of Oghuz Turks resulted in the gradual “Turkification” of Azerbaijan and Anatolia (Frye 2004). Key (2012),
who focuses on morphosyntactic effects of contact, suggests that Turkic influence may
date from the Safavid state (1501–1736) “the rulers of which were Persianized Turks who
spoke a variety of Middle Azerbaijanian that might actually have been a mixed language
incorporating Ottoman elements (Stein 2005: 228).”8 Frye (2004) presents a distinct view
of the Safavids as Turkicized Iranians, but most seem to agree that it was the post-Islamic
migration of Turks, as opposed to Arabic speakers, that had the most linguistic influence
in the area: “ …the Turks who came, especially beginning from the tenth century, moved
in sufficient numbers to change the linguistic map of the whole area. (op cit.)”
Though Classical Arabic also disallows onset clusters, there are several reasons to
doubt Arabic as the source of cluster-splitting epenthesis in Modern Persian. First, evidence from early loans into Classical Arabic shows common prothetic vowels, with
epenthesis the exception (cf. Arabic ʔiklīl ‘crown , wreath’ from Syriac klīlo, Arabic
ʔiqlīm ‘region’ from Greek klīma; but also Arabic dirham ‘money’ from Greek drakhmi;
Bueasa 2015). Second, the influence of Arabic on Middle Iranian languages came, primarily, through translation of religious texts into Arabic, and through acquisition of Arabic
by writers and thinkers who used it as a prestige language. This socialization process
was notably different from the Persianization of Turkic people referred to above, and
resulted in significant loans, but no obvious evidence of Arabic influence on Persian
I hypothesize that cluster-splitting epenthesis in the history of Persian arose as a result
of contact between speakers of Turkic languages, which did not allow complex onsets,
and speakers of Middle Iranian languages with initial #TR-clusters. As Turks became
Persianized, they acquired Persian (and, perhaps, other Middle Iranian languages). In
this process, cognitive effects of CV(C) syllable structure resulted in the perception of
illusory vowels in #TR-initial words (cf. Dupoux et al. 1999), giving rise to the change in
Key’s (2012) study of differential object marking in Turkic and Persian identifies Iranian Azerbaijan as an
isogloss for this feature.
Juliette Blevins
pronunciation schematized in (3). Under this account, the rarity of sound changes like
(3) is attributed to three factors: first, initial #TR clusters are relatively stable over time,
so (3) is unexpected as a language-internal development; second, a sound change of this
kind requires contact between two distinct language types, one language which lacks
complex onsets and another which has word-initial #TR; a third factor is the nature of
the language contact involved, which must include social factors that demand rapid and
full acquisition of the language with #TR clusters despite minimal previous exposure.9
Only when these last two conditions are met will cluster-splitting epenthesis occur as a
regular sound change.10
Can the same hypothesis involving language contact of a very specific type account
for the evolution of Dorsey’s Law in Hoocąk (Winnebago)? I believe so. Oral histories
suggest that the split between Hoocąk, traditionally spoken between Green Bay and
Lake Winnebago in present-day Wisconsin, and Chiwere, once spoken south and west
of Hoocąk territory, occurred sometime in the mid-16th century, a time-line consistent
with the great similarity between the two languages.11 This would make the mid-16th century the earliest time at which Hoocąk could have developed cluster-splitting epenthesis,
an innovation not found in Chiwere (3). By the time Jean Nicolet made contact with the
“Ho-Chunk” in 1634, with an estimated population of 8,000 or more, their culture was
very similar to that of surrounding Algonquian tribes, they were completely encircled by
speakers of Algonquian languages, and the language had a significant number of borrowings from Central Algonquian languages (Radin 1990; Pfister 2009: 17).12 I suggest that
sometime between the mid-16th and mid-17th centuries, (pre-)Hoocąk was acquired by
speakers of neighboring Algonquian languages. Since none of the Central Algonquian
languages had initial #TR clusters, cognitive effects of #CV(C) syllable structure resulted
in the perception of illusory vowels in #TR-initial words (cf. Dupoux et al. 1999), giving
rise to Dorsey’s Law. As with the contact scenario sketched for Modern Persian above,
the evolution of cluster-splitting epenthesis is associated not only with these structuralcognitive factors, but also with a specific type of language contact: external social factors
demanding rapid and full acquisition of a language, (pre-)Hoocąk, with initial #TR clusters by speakers of a language Central Algonquian language with only simple #C-onsets
This process is distinct from creolization, since the starting point here is not a pidgin. Interestingly, many
Creoles show initial complex onsets (Klein 2013), consistent with the view here, that they are relatively
stable, and not particularly “marked”.
An anonymous reviewer notes that if future generations have access to the donor language, and that
language is prestigious, one may see a shift involving adoption of the donor phonotactics.
Though the homeland of the Siouan-Catawba language family is widely debated, oral histories and archeological remains are consistent with (pre)-Hoocąk occupation of land between Green Bay and Lake Winnebago (in present-day northeast Wisconsin) in pre-contact times.
By the late 1650s, the Hoocąk population may have been as few as 500 people, with great cultural devastation. This drastic decrease in population is attributed to a storm-related accident, epidemics (due to
European contact), and/or battle losses to neighboring tribes (Edmunds 1978; Radin 1990).
1 Between natural and unnatural phonology: The case of cluster-splitting epenthesis
4 Concluding remarks
The typology of sound change may seem like an odd place to uncover significant evidence of cognitive forces that are independent of universal phonetics, or evidence against
widely assumed markedness constraints.13 Yet, this study of cluster-splitting epenthesis
as regular sound change suggests that typological studies of this kind may illuminate
our understanding of the role of human cognition in shaping sound patterns, and the
extent to which general aspects of memory, category formation, similarity metrics, and
analogy contribute to their evolution (Blevins & Blevins 2009). Contrary to widely assumed markedness constraints treating all complex onsets as marked or dispreferred,
the typology of sound change suggests that word-initial #TR clusters are phonotactically stable. On the other hand, in the rare cases where these clusters undergo regular cluster-splitting epenthesis, this epenthesis is not a simple case of ”syllable-repair”.
Rather, native-language #CV-structure in language-contact situations results in the perception of phantom vowels which take on phonological status when speakers of #CVinitial languages must quickly, and with little earlier familiarity, acquire a language with
#TR clusters. This, I suggest, was the original situation of Turkic speakers acquiring Persian, and of Central Algonquians acquiring Hoocąk. Unlike many other common sound
patterns, regular cluster-splitting epenthesis does not have a simple phonetic explanation, and is not known as a purely language-internal development. By examining other
sound changes with this profile, we may, unexpectedly, learn even more about the human mind.
Anderson, Stephen R. 1981. Why phonology isn’t “natural”. Linguistic Inquiry 12(4). 493–
Anderson, Stephen R. 1985. Phonology in the Twentieth century: Theories of Rules and
Theories of Representations. Chicago: University of Chicago Press.
Anderson, Stephen R. 2001. Why linguistics needs the cognitive scientist. Psychiatry
64(1). 11–13.
Anderson, Stephen R. 2016. Synchronic versus diachronic explanation and the nature of
the language faculty. Annual Review of Linguistics 2. 11–31.
Bach, Emmon & Robert Harms. 1972. How do languages get crazy rules? In Robert Stockwell & Ronald Macaulay (eds.), Linguistic change and generative theory. Bloomington:
Indiana University Press.
Berent, Iris. 2013. The phonological mind. Cambridge: Cambridge University Press.
Blevins, James & Juliette Blevins. 2009. Analogy in grammar: Form and acquisition. Oxford: Oxford University Press.
Though other instances of non-phonetic cognitive forces have been suggested. See, for example, Blevins &
Wedel (2009), where lexical competition is argued to play an active role in shaping the typology of sound
Juliette Blevins
Blevins, Juliette. 2004. Evolutionary phonology: The emergence of sound patterns. Cambridge: Cambridge University Press.
Blevins, Juliette. 2006. New perspectives on English sound patterns: ‘Natural’ and ‘unnatural’ in Evolutionary Phonology. Journal of English Linguistics 34. 6–25.
Blevins, Juliette. 2008a. Consonant epenthesis: Natural and unnatural histories. In Jeff
Good (ed.), Language universals and language change, 79–107. Oxford: Oxford University Press.
Blevins, Juliette. 2008b. Natural and unnatural sound patterns: A pocket field guide. In
Klaas Willems & Ludovic De Cuypere (eds.), Naturalness and iconicity in language,
121–48. Amsterdam: John Benjamins.
Blevins, Juliette. 2008c. Phonetic explanation without compromise: The evolution of Mussau syncope. Diachronica 25. 1–19.
Blevins, Juliette. 2015. Evolutionary phonology: A holistic approach to sound change
typology. In Patrick Honeybone & Joseph Salmons (eds.), Handbook of historical phonology, 485–500. Oxford: Oxford University Press.
Blevins, Juliette & Ander Egurtzegi. 2016. The nature of obstruent loss in Basque borrowings with initial OR clusters. In CUNY Phonology Forum. The Graduate Center, New
Blevins, Juliette & Sven Grawunder. 2009. *K1 > T1 sound change in Germanic and elsewhere: Descriptions, explanations, and implications. Linguistic Typology 13. 267–303.
Blevins, Juliette & Andrew Wedel. 2009. Inhibited sound change: An evolutionary approach to lexical competition. Diachronica 26. 143–183.
Broselow, Ellen. 1987. Non-obvious transfer: on predicting epenthesis errors. In Georgette Ioup & Steven H. Weinberger (eds.), Interlanguage phonology: the acquisition of
a second language sound system, 269–280. Rowely, MA: Newbury House Publishers.
Broselow, Ellen. 2015. The typology of position-quality interactions in loanword vowel
insertion. In Y. Hsiao & L-H. Wee (eds.), Capturing phonological shades, 292–319. Cambridge: Cambridge Scholars Publishing.
Buckley, Eugene. 2000. On the naturalness of unnatural rules. Santa Barbara Papers in
Linguistics 9. 16–29.
Bueasa, Noor Mohammed. 2015. The adaptation of loanwords in Classical Arabic: The governing factors. University of Kentucky MA thesis.
Campbell, Lyle. 2013. Historical linguistics: An introduction. Third Edition. Edinburgh: Edinburgh University Press.
Casparis, Johannes de. 1997. Sanskrit loan-words in Indonesian: An annotated check-list
of words from Sanskrit in Indonesian and traditional Malay. Badan Penyelenggara Seri
NUSA, Universitas Katolik Indonesia Atma Jaya.
Davidson, Lisa & Jason Shaw. 2012. Sources of illusion in consonant cluster perception.
Journal of Phonetics. 234–248.
de Chene, Brent E. & Stephen R. Anderson. 1979. Compensatory lengthening. Language
55(3). 505–535.
Dorsey, J. Owen. 1885. On the comparative phonology of four Siouan languages. Annual
Report of the Board of Regents of the Smithsonian Institution 919-929.
1 Between natural and unnatural phonology: The case of cluster-splitting epenthesis
Dupoux, Emmanuel, Kazuhiko Kakehi, Yuki Hirose, Christophe Pallier & Jacques Mehler.
1999. Epenthetic vowels in Japanese: A perceptual illusion? Journal of Experimental
Psychology: Human Perception and Performance 25. 1568–1578.
Edmunds, David. 1978. The Potawatomis: Keepers of the fire. Norman: University of Oklahoma Press.
Fleischhacker, Heidi. 2001. Cluster-dependent epenthesis asymmetries. In Adam Albright & Taehong Cho (eds.), UCLA working papers in linguistics 7, Papers in phonology
5, 71–116.
Fleischhacker, Heidi. 2005. Similarity in phonology: Evidence from reduplication and loanword adaptation. UCLA PhD thesis.
Fortson Benjamin W., IV. 2010. Indo-European language and culture: An introduction. Second Edition. London: John Wiley & Sons.
Frye, Richard. 2004. Peoples of Iran (1): A general survey. Encyclopaedia Iranica XIII(3).
Garrett, Andrew & Juliette Blevins. 2009. Analogical morphophonology. In Sharon Inkelas & Kristin Hanson (eds.), The Nature of the Word: Essays in honor of Paul Kiparsky,
527–46. MIT Press.
Hall, Nancy. 2011. Vowel epenthesis. In Marc van Oostendorp, Colin Ewen, Elizabeth
Hume & Keren Rice (eds.), The Blackwell companion to phonology, 1576–1596. Oxford:
Hyman, Larry. 1970. The role of borrowing in the justification of phonological grammars.
Studies in African Linguistics 1. 1–48.
Itô, Junko. 1989. A prosodic theory of epenthesis. Natural Language and Linguistic Theory
7(2). 217–259.
Kabak, Barış & William J. Idsardi. 2007. Perceptual distortions in the adaptation of English consonant clusters: syllable structure or consonantal contact constraints? Language and Speech 50(1). 23–52.
Kager, René. 1999. Optimality theory. Cambridge: Cambridge University Press.
Kang, Yoonjung. 2003. Perceptual similarity in loanword adaptation: English postvocalic
word-final stops in Korean. Phonology 20. 219–273.
Kang, Yoonjung. 2011. Loanword phonology. In Marc van Oostendorp, Colin J. Ewen,
Elizabeth Hume & Keren Rice (eds.), The Blackwell companion to phonology, 2258–2282.
Karimi, Simin. 1987. Farsi speakers and the initial consonant cluster in English. In Georgette Ioup & Steven H. Weinberger (eds.), Interlanguage phonology: the acquisition of
a second language sound system, 305–318. Rowely, MA: Newbury House Publishers.
Kavitskaya, Darya. 2002. Compensatory lengthening: Phonetics, phonology, diachrony.
New York: Routledge.
Kenstowicz, Michael. 2007. Salience and similarity in loanword adaptation: A case study
from Fijian. Language Sciences 29. 316–340.
Key, Greg. 2012. Differential object marking in Turkic and Persian as a contact phenomenon. In Proceedings of the Annual Meeting of the Berkeley Linguistics Society,
vol. 38, 239–252.
Juliette Blevins
Klein, Thomas. 2013. Typology of creole phonology: Phoneme inventories and syllable
templates. In Parth Bhatt & Tonjes Veenstra (eds.), Creole languages and linguistic typology, 207–244. Amsterdam: John Benjamins.
Miner, Kenneth. 1979. Dorsey’s Law in Winnebago-Chiwere and Winnebago accent. International Journal of American Linguistics 45. 25–33.
Ng, E-Ching. 2015. The phonology of contact: Creole sound change in context. Yale University PhD thesis.
Ohala, John. 1971. The role of physiological and acoustic models in explaining the direction of sound change. Project on Linguistic Analysis Reports (Berkeley) 15. 25–40.
Ohala, John. 1974. Experimental historical phonology. In J. M. Anderson & C. Jones (eds.),
Historical linguistics II. Theory and description in phonology [Proceedings of the 1st International Conference on Historical Linguistics], 353–389. Amsterdam: North Holland.
Ohala, John. 1993. The phonetics of sound change. In Charles Jones (ed.), Historical linguistics: Problems and perspectives, 237–278. London: Longman.
Parks, Douglas & Robert Rankin. 2001. The Siouan languages. In Raymond DeMallie &
William Sturtevant (eds.), Handbook of North American Indians: Plains, vol. 13, 94–114.
Washington, D.C.: Smithsonian Institution.
Peperkamp, Sharon & Emmanuel Dupoux. 2002. A typological study of stress “deafness”.
Laboratory phonology 7. Carlos Gussenhoven & Natasha Warner (eds.). 203–240.
Perry, John. 2004. Arabic elements in Persian. Encyclopaedia Iranica 2. online version,, 229–36.
Perry, John. 2005. Lexical areas and semantic fields of Arabic loanwords in Persian and
beyond. In Carina Jahani, Bo Isaksson & Éva Ágnes Csató (eds.), Linguistic convergence
and areal diffusion: Case studies from Iranian, Semitic and Turkic, 97–110. New York:
Pfister, Joel. 2009. The Yale Indian: The education of Henry Roe Cloud. Chapel Hill: Duke
University Press.
Prince, Alan & Paul Smolensky. 1993. Optimality Theory: Constraint interaction in generative grammar. Oxford: Blackwell Publishing.
Radin, Paul. 1990. The Winnebago tribe. Lincoln: University of Nebraska Press.
Rankin, Robert, Richard Carter, A. Welsey Jones, John Koontz, David Rood & Iren Harmann (eds.). 2015. Comparative Siouan dictionary. Leipzig: Max Planck Institute for
Evolutionary Anthropology.
Stein, Heidi. 2005. Traces of Türki-yi Acemi in Pietro della Valle’s Turkish Grammar
(1620). In Éva Ágnes Csató, Bo Isaksson & Carina Jahani (eds.), Linguistic convergence
and areal diffusion: Case studies from Iranian, Semitic and Turkic. New York: Routledge
Strain, Jeris. 1968. A contrastive sketch of the Persian and English sound systems. International Review of Applied Linguistics 6. 55–62.
Uffmann, Christian. 2007. Vowel epenthesis in loanword adaptation. Tübingen: Max Niemeyer Verlag.
Vaux, Bert. 2002. Consonant epenthesis and the problem of unnatural phonology. Department of Linguistics, Harvard University MA thesis.
1 Between natural and unnatural phonology: The case of cluster-splitting epenthesis
Windfuhr, Gernot. 2009. Dialectology and topics. In Gernot Windfuhr (ed.), The Iranian
languages, 5–42. New York: Routledge.
Windfuhr, Gernot & John Perry. 2009. Persian and Tajik. In Gernot Windfuhr (ed.), The
Iranian languages, 416–544. New York: Routledge.
Chapter 2
The domain of stress assignment:
Word-boundedness and frequent
Ellen Kaisse
University of Washington
Phenomena that a theory of the human
language faculty ought to accommodate might well happen never to be
attested because there is no course of
change or borrowing by which they
could arise. (Anderson 1992: 336)
The linguistic literature treats hundreds of processes that apply between adjacent, open class
content words. Overwhelmingly, these processes are local, segmental adjustments applying
between the last segment of one word and the first segment of the next, such as voicing or
place assimilation. Most other kinds of processes are profoundly underrepresented. Notably,
iterative processes that eat their way across words such as vowel harmony, consonant harmony, or footing (assignment of rhythmic stress) typically do not extend beyond the word
or count material outside the word when locating a particular stressable syllable, such as
the penult. This result becomes more understandable when one considers how processes
are phonologized from their phonetic precursors. Precursors are articulatorily or perceptually motivated and are strongest in temporally adjacent segments; they lose their force
as the distance between segments increases, so there are no strong precursors for iteration
outside of a word. Furthermore, frequent repetition leads to phonologization (Bybee 2006).
Any content word in a lexicon occurs next to any other content word much less frequently
than do roots plus their affixes, hosts plus clitics, or any combination that includes at least
one closed class item. So we should expect roots and affixes or hosts and clitics to be much
more common as domains for stress assignment or other iterative rules than are strings of
independent lexical items. In this paper, I concentrate on the near non-existence of stress
assignment rules that span a domain larger than a morphological word plus clitics. We look
at one revealing case that does treat some pairs of content words as a single span: stress
assignment in literary Macedonian (Lunt 1952). The spans involve frequent collocations –
Ellen Kaisse. 2017. The domain of stress assignment: Word-boundedness and
frequent collocation. In Claire Bowern, Laurence Horn & Raffaella Zanuttini
(eds.), On looking into words (and beyond), 17–41. Berlin: Language Science Press.
Ellen Kaisse
groups of already frequent words that are frequently heard together, sometimes to the extent that they have developed a lexicalized meaning. The other interword Macedonian cases
involve closed class item such as prepositions and interrogatives. Finally, we consider evidence that rhythm can be evidenced statistically in the syntactic choices speakers make
(Anttila, Adams & Speriosu 2010; Shih 2016), concluding that there may be rhythmic pressures from the lexicon to make phrases eurhythmic as well as eurhythmic pressures from
phrases that can be phonologized to create rules of lexical stress assignment.
1 Introduction: what kinds of processes are postlexical
and where do stress rules fit in?
Some types of phonological processes seem always to apply solely within a single word,
not taking into account any material in adjacent words, while other types can apply
between words. In work to appear (Kaisse Forthcoming) I surveyed the literature on
lexical (word-bounded) and postlexical (non-word-bounded) rules, sampling about 70
careful descriptions of non-tonal processes that make up their focus and environment
from material that spans more than one content word.1 Only a few involve anything
other than strictly local adjustments between the last segment of one word and the first
segment of the next. They might require agreement in voicing, place of articulation, or
other features. Or they might avoid onset-less syllables by deleting or desyllabifying a
word final vowel when the next word begins with a vowel, or by moving a word final
consonant into the onset of the next, vowel-initial word. (1) contains some fairly familiar
examples from Spanish. (1a) illustrates postlexical voice assimilation of /s/ to [z] before
voiced consonants and assimilation of continuancy (/g/ → [ɣ] and /b/ → [β]) after that
/s/; (1b) shows place assimilation of a nasal to a following consonant and the retention
of an underlying stop after non-continuant /n/; (1c) shows reduction of hiatus; and (1d)
shows resyllabification of a word-final consonant:
Spanish (personal knowledge)
a. /los ˈgato-s ˈbjen-en/ → [loz ˈɣatoz ˈβjenen]
the cat-pl come-3pl
‘The cats are coming.’
b. /ˈbjen-en ˈgato-s/ → [ˈbjeneŋ ˈgatos]
come-3pl cat-pl
‘Cats are coming.’
Some tonal processes are well known to span large numbers of syllables, both within and between words
(Hyman 2011), though many are also word-bounded. In Kaisse (forthcoming) I endorse David Odden’s (p.c.
2015) speculation about the long- distance behavior of tonal adjustments in Bantu languages, which offer
the most numerous and robust examples of processes where several tones in one word can be affected by
a distant tone in an adjacent word. Odden cites the perceptual difficulty of locating tone in Bantu, with its
long words and subtle cues for tone, and tone’s low functional load there, since only H, not L tone needs
to be marked.
2 The domain of stress assignment: Word-boundedness and frequent collocation
c. /ˈteŋ-o ˈotro-s/ → [ˈteŋ ˈotros]
have-1sg other-pl
‘I have others.’
d. /ˈtjen-es ˈotro-s/ → [ˈˈso.tros]
have-2sg other-pl
‘You have others.’
Really, any local adjustment that is found within words can be found between words.
This exuberance of types is probably due to the fact that most phonologized processes
start life as natural local effects and these effects are not sensitive to grammatical information but rather to temporal adjacency (Kiparsky 1982 et seq.). However, I found
almost no vowel harmony processes that extended into an adjacent content word, no
consonant harmony processes (Hansson 2010), and, crucially for the current paper, no
stress rules with a domain larger than the morphological word or the morphological
word plus clitics. Compare the familiar types of examples in (1) with the fanciful ones in
(2), where something resembling the English rule that constructs moraic trochees from
the end of a word takes a whole noun phrase as its domain, resulting in feet that span
syllables belonging to different words and in wide-scale allomorphy:
(2) Fanciful English with noun phrase as initial domain of footing
( × )( ×
/ taj ni dag /
‘tiny dog’
( × . )( × . )
/ taj ni kɪ tɛn /
‘tiny kitten’
(×) (×
. ) (× . )
/ taj ni pɪ ɹa na /
‘tiny piranha’
In this paper, I will describe the continuum of stress behavior of cohering and noncohering affixes (i.e. affixes that do and do not interact phonologically with their bases),
clitics of varying degrees of stress-interaction with their hosts, compound words, and the
one detailed presentation of a stress rule I have encountered where initial stress assignment takes some strings of content words and treats them as a single domain, ignoring
any stress the component words might otherwise have in isolation. That case comes
from literary Macedonian (Lunt 1952; Franks 1987; 1989).
Ellen Kaisse
Before continuing, I should make it clear that there are many rhythmic adjustments
that do apply between content words – cases like the Rhythm Rule in English (Hayes
1984) which is responsible for the retraction of the secondary stress in ˌJapanese ˈlanguage
(vs. ˌJapaˈnese) or ˌMississippi ˈmud vs. ˌMissiˈssippi. These kinds of cases are postlexical
and involve prosodically self-sufficient content words that have had their own stresses assigned lexically, independent of the larger context in which they find themselves. There
is then a rhythmic adjustment that demotes or moves a nonprimary stress in order to
avoid stress clash. Like the invented example (2), the Macedonian case that I will look
at instead involves assigning a single antepenultimate stress to a string of two content
words which, in other syntactic contexts, would each receive a stress of their own. Often
the single stress does not fall on any of the syllables that would have been stressed in
isolation. This is clearly different from the way a rhythm rule works.
Another example of postlexical stress adjustment, as opposed to the first pass of stress
assignment, involves compound stress rules. Again, these are not relevant to my claim
that pairs of content words are almost never the domain of the first pass of stress assignment. The most well-known of these compound stress rules, like that of English,
also simply adjust the relative primary stresses of the members, rather than treating
them as a single unit for the lexical footing process. Thus, if we put together linˈguistics
and ˈscholar we get linˈguistics ˌscholar, with the primary stress of ˈscholar subordinated
to that of linˈguistics, but we do not stress the whole string as a single prosodic word,
which might result in antepenultimate stress falling on the last syllable of the first member, with the bizarre (for English) output linguiˈstics scholar. We shall see, however, that
occasionally languages (such as Modern Greek) do take compound words as the initial
domain of stress assignment. So while prosodically independent content words are not
commonly the domain for the first pass of stress assignment, some languages stretch
that domain to include both members of a compound word.
Yet another aspect of the postlexical adjustment of stress is offered by Newman’s (1944:
28–29) insightful early discussion of stress domains in Yokuts, pointed out to me by
an anonymous referee. Nouns and verbs maintain their lexical stresses in connected
speech but function words tend to lean on them and to lack stresses of their own, and
the faster the speech, the bigger the phrasing groups and the fewer the perceived stresses.
Newman’s description is perforce rather general and impressionistic, but it summarizes
well the plasticity of postlexical stress adjustments, which favor cliticization of function
words and variable rhythmic groupings dependent on tempo and number of syllables.2
While truly grammaticized rhythmic stress assignment almost never takes an initial
domain beyond a word and its affixes and clitics, there are now known to be gradient
rhythmic effects that extend beyond the word. They simply don’t seem to be able to
rise to the level of phonologization in that larger domain. Martin (2007), Anttila, Adams
& Speriosu (2010), and Shih (2016) have found that vowel harmony (Martin) and optimization of rhythm can have statistical reflections in Turkish compound words and in
I cannot tell from Newman’s description how content words that are not nouns or verbs might behave. I
tentatively conclude that he is referring only to lack of stress on function words in rapid connected speech,
not to complete loss of stress on open class modifiers or other content words.
2 The domain of stress assignment: Word-boundedness and frequent collocation
English word order. That is, to use Martin’s felicitous phrasing, lexical generalizations
that are part of the grammar can “leak” into larger domains, causing statistical preferences for outputs that are consonant with the phonology of a language’s words. I would
add that it is hard to know what the primary direction of leakage really is: the leakage
Martin and Shih posit from smaller domains to larger ones probably co-exists with the direction of larger-to-smaller domain phonologization, which I posit in Kaisse (Forthcoming) and which is the staple view of Lexical Phonology and its descendants (Kiparsky
1982; Bermudez-Otero 2015) and of most traditional approaches to sound change. In
Kaisse (Forthcoming) I used the example of the phonetic precursor of vowel harmony,
namely the vowel-to-vowel coarticulation that peters out as one gets farther away from
the source vowel (Öhman 1966; Ohala 1992; Flemming 1997), which can be grammaticized within words because stems and their affixes are in frequent collocation (Bybee
2006)3 . Classical Lexical Phonology postulates this one-way direction (from postlexical
and variable or gradient to lexical and regular and obligatory), but there is no reason
that once a rule is lexicalized, it cannot then generalize postlexically again (Kaisse 1993).
One can imagine a feedback loop, where small postlexical variation in favor of alternating stresses and avoidance of stress clash starts to be phonologized as iterative footing,
while iterative footing starts to make syntactic phrases that are eurhythmic more desirable as choices for speakers in real time.
There is a continuum of likely domains for foot-construction. Selkirk (1995), Peperkamp (1997) and Anderson (2011) demonstrate that there are various types of clitics that
range from more to less rhythmically cohesive (interactive) with their hosts. I will extend
the continuum, showing that in literary Macedonian (Lunt 1952; Franks 1987; 1989) the
lexical stress rule assigning antepenultimate stress has stretched its domain to include
even two content words, but only when they are in frequent collocation and, in some
of the cases, have taken on a more lexicalized, semantically less compositional meaning.
This is the “exception that proves the rule.” In general, only the supremely frequent collocation of a closed class bound morpheme – an affix – with another affix or the root
it can attach to provides the frequent collocation that allows for phonologization. However, occasionally even two content words can appear together so frequently that they
become subject to the first pass of the lexical stress rules of the language; they act like a
single word for the purposes of building a foot at the edge of a word. Clitics and function
words in Macedonian also form part of the initial domain for this foot-building. This accords with the general observation that the phonology of clitics is like the phonology of
affixes in many cases – they are ‘phrasal affixes’ (Anderson 1992). Clitics are usually less
cohering than stem-level affixes but, as we shall illustrate in §2, there are even a few that
act like they are inside the phonological word for the purpose of foot-construction. But
basic foot building algorithms do not typically extend beyond the morphological word.
Barnes (2006) cites Inkelas et al. (2001) for evidence that phonetic vowel-to-vowel coarticulation is problematic as a simple, unaided source for vowel harmony. Their argument comes from Turkish, where anticipatory phonetic effects are stronger than perseveratory ones, but the phonologized harmony system is perseveratory. He instead attributes the phonologization of vowel harmony to vowel-to-vowel coarticulation
coupled with lengthening of the trigger syllable and paradigm uniformity effects allowing longer-distance
effects on distant affixes.
Ellen Kaisse
Occasionally they extend even less far than that, as in the case of non-cohering (stress
neutral) suffixes of English, and occasionally they extend into the larger phonological
word, which can include clitics and other function words that can have stressless variants – i.e. be prosodically deficient, in Anderson’s (2011)’s terms. Indeed, the failure of
an affix to participate in the lexical footing domain is one of the main ways in which
non-cohering affixes have been defined. Note for instance that Chomsky & Halle (1968:
84ff) use the term “stress neutral” for the group of suffixes which, to anticipate the later
interpretation of their intentions, are outside the prosodic word. These suffixes, which
include all the inflectional affixes, as well as a number derivational affixes such as #ly,
#ness, agentive #er, #ful, and many others, not only fail to affect stress but also don’t
induce word-internal rules like Trisyllabic Shortening and Sonorant Syllabification. (See
Kiparsky 1982 for a full discussion of Trisyllabic Laxing and Kaisse 2005 for a summary
of the characteristics of English cohering suffixes.) Indeed, we might ask why stress
is one of the most typical diagnostics for cohering suffixes, and not only in English. I
would speculate that rhythm is hard to grammaticize as a phonological process without frequent repetition of the same strings. Like clitics, word-level suffixes have less
stringent subcategorization restrictions. Fabb (1988), which we will discuss in a bit more
detail in §2 discovered that the possible combinations of English stem-level suffixes with
other suffixes or with roots are very restricted. On the other hand, word-level suffixes
can combine freely with many suffixes and words, and therefore are not in as frequent
collocation as the stem level ones, which are heard over and over again with the same
preceding morpheme, be it an affix or a root. Clitics are even less demanding of the
preceding host – indeed in some cases, such as special clitics, the host can belong to any
part of speech and can even be a phrase – so they are less likely to be phonologized as
part of the stress domain. But because they are prosodically unable to stand on their
own, they must lean on a host and thus may sometimes be phonologized as part of the
stress domain.
2 Clitics and the stress domain continuum
There has been considerable attention paid to the various ways that clitics – prosodically
deficient items that must lean on a host in some fashion – interact with the lexical stress
assignment rules of a language and fall into their domain. I will summarize some recent
results here because I believe that the insights from clitics can help us understand how
compound words and even some phrases can come to behave in the same way as clitics
do with respect to their hosts.
It is worth reviewing some of the typical diagnostics of clitics (paraphrased from
Zwicky & Pullum 1983):
Characteristics of clitics
a. clitics have a low degrees of selection with respect to their hosts; affixes have
a higher degree of selection.
2 The domain of stress assignment: Word-boundedness and frequent collocation
b. affixed words are more likely to have idiosyncratic semantics than host +
clitic combinations
c. affixed words are treated as a unit by the syntax while hosts plus their clitics
are not.
To these we can add that while the boundary between a root and stem-level (cohering)
affix is the most favorable position for phonological interactions to occur, and the boundary between base and word-level (non-cohering) suffixes somewhat less favorable, clitics are sometimes phonologically even less connected – less cohesive – with their hosts;
they might fail, for instance, to undergo vowel harmony, the last segment of their hosts
might undergo word-final devoicing, as if there were no following segment, and as we
shall see, in some languages, the clitics might not be visible to the lexical stress rules or
at least to the first pass of those rules.
Beginning with Anderson’s (1992) classification of clitics as phrasal affixes as a foundation, Selkirk (1995) and Peperkamp (1997), elaborated in Anderson (2011), propose a
hierarchy of clitic types. Ranked from least-cohering to most cohering with respect to
phonological interaction with their host, we have the continuum in (4):
(4) The clitic continuum
prosodic word clitic > free clitic > affixal clitic > internal clitic
Free Clitic
Prosodic Word Clitic
Affixal Clitic
Internal Clitic
Intuitively speaking, we know that clitics usually fall somewhere in between the phonological cohesiveness of stem-level, highly cohering affixes and the phonological noncohesiveness of independent prosodic words. But in this elaborated hierarchy, not all clitics do. There are clitics termed ‘internal’ by Selkirk that act, at least for some processes,
exactly like stem-level affixes. We will extend this notion to the compound words of
Ellen Kaisse
Modern Greek and the extended stress domains of Macedonian – these involve normally
prosodically independent words that nonetheless act as a single domain and receive one
stress as if they were a single word. On the other end, there are clitics that Selkirk
terms ‘prosodic word clitics;’ these are forced in some cases to behave like independent
prosodic words despite their underlying prosodic deficiency. In (5), I illustrate the internal type with the example of verbal clitics in Formentera Catalan (Torres-Tamarit &
Pons-Moll 2015), where the whole verb-clitic complex is used to assign moraic trochees
from the right. This can result in no stress falling on the verb stem itself if there are
two monosyllabic clitics (5d) or one monosyllabic clitic followed by a clitic consisting
of a single consonant that makes a heavy syllable (5e). The raising of the root /o/ of the
imperative verb in (5b–5e), which only occurs to unstressed mid vowels, suggests that
stress is not assigned cyclically to the root and then again to the host-clitic complex, but
rather all in one go to the entire string:
Formentera Catalan (Torres-Tamarit & Pons-Moll 2015) internal clitics
a. ˈkompr-ə
buy -2sg.imp
b. kumˈpr-ə
buy -2sg.imp=3sg.m.acc
‘Buy it/him!’
c. kumˈpr-ə
buy -2sg.imp=part
‘Buy some!’
d. kumpr-ə
=ˈmə =lə
buy -2sg.imp=1sgdat=3sg.f.acc
‘Buy it/her for me!’
e. kumpr-ə
buy -2sg.imp=1sg.dat=3m.acc
‘Buy it/him for me!’
Internal clitics meet the criteria of low selection, predictable semantics, and so forth,
but phonologically they are unusual – they act like stem-level affixes. The situation in
Formentera is similar to that described for Lucanian Italian clitics by Peperkamp (1997).
Formentera and Lucanian internal clitics are wholly included within the phonological
word and count from the outset in the calculation of where a right edge trochaic foot
should be built.
More generally speaking, internal clitics are prosodically deficient words that act more
cohesively in their phonological interaction with their host than their other characteristics would suggest they should. This is captured in Selkirk’s framework by having them
form a single prosodic word in combination with their host. Their hallmark is that they
2 The domain of stress assignment: Word-boundedness and frequent collocation
are stressed and generally processed by the word-bounded on the first pass, acting just
like a stem-level affix. This means that they can never receive a second stress all their
own, but they may happen to be in position to take the stress of the whole host-clitic
string, as illustrated in the Formentera form (5d) [kumprə=ˈmə=lə], where the clitic /mə/
receives the penultimate stress for the whole host-clitic complex. I suggest that we can
profitably extend this insight onto bigger, more independent words than clitics. We will
see in §3 that Modern Greek treats the otherwise independent pieces of a compound
word as a single domain for stress assignment – a comparatively rare phenomenon. And
Macedonian (§4) treats prepositions, even polysyllabic ones, as a single stress domain
with their objects and treats certain common collocations of adjective + noun as a single
stress domain as well. To put it another way, one occasionally encounters situations
where a pair of words that should be expected to be prosodically independent can act
like ‘internal words,’ parallel to the notion of internal clitics. I will link this unusual phenomenon to frequency of collocation. The more exemplars there are of a string of words,
repeatedly heard together, the more likely they are to be internalized as a rhythmic group
and to be reified as a domain for the assignment of a single stress.
At the other end of the continuum of independence of clitics lie the Prosodic Word
clitics, which receive postlexical treatment as independent prosodic words even though
they are underlyingly prosodically deficient. They are illustrated by the prevocalic clitics
of Bilua (Obata 2003, reported in Anderson 2011), where preconsonantal proclitics are
normally free – they are outside the domain of stress assignment, which targets the first
syllable of the word:
(6) Bilua (Obata 2003, Anderson 2011: 2004) free clitic
‘He killed it.’
When a host word begins with a vowel, however, the usually stressless clitics instead
take on a derived prosodic word-hood of their own and receive an independent stress:
Bilua (Obata 2003, Anderson 2011) PWd clitic under duress
3sg.m=call =3sg.fem.obj=tns
‘He called her.’
In this case, the clitic is forced to act like content words do normally.
A referee has expressed some doubt about the proper analysis of Bilua clitics and, by
extension, whether the category of prosodic clitics needs to exist at all. But it is useful to
suspend our skepticism and contemplate such an entity because it illustrates the opposite side of a prosodic coin. The putative prosodic clitics are elements that are inherently
prosodically deficient, but that can be forced under certain circumstances to act as independent prosodic words with their own stress. Macedonian nouns and adjectives in
certain frequent collocations are inherently prosodically independent elements that can
Ellen Kaisse
be forced under certain circumstances to act as prosodically deficient pieces of a single
prosodic word, as are the words that make up a compound in Modern Greek.
To complete the hierarchy, there are clitics whose characteristics of independence
place them between internal and Prosodic Word clitics. The affixal clitics and the free
clitics are probably the ones that linguists are most used to encountering. An affixal
clitic shows some prosodic independence from its host: it is not included in the domain
of footing on the first pass. In Selkirk’s framework it forms a recursive prosodic word
with that host, and it therefore may receive or induce an additional stress or stress shift
once it is included in a second pass of stress assignment. When word stress falls on the
penult or antepenult of a content word, we find an additional stress added to a string
of clitics if they can form a foot, as shown in the third column of (8) (Binary footing is
indicated with parentheses.):
Neapolitan Italian (Peperkamp 1997) affixal clitics
a. ˈkonta
b. ˈpettina
ˈpettina =lə
comb.imp comb.imp=it comb.imp=you=it
Neapolitan Italian treats the morphological word, sans clitics, as the basic, initial domain
of footing, but clitics are partially cohering, falling into a larger domain that still follows
the basic principles of word-internal stress assignment. In Neapolitan, if the stress assigned to the clitic sequence clashes with the stress of the host, the host’s stress is lost:
Neapolitan Italian (Peperkamp 1997) affixal clitics with clash reduction
ˈfal =lə
=ˈtti =llə
The parallel for us here is that the pieces of compound words and larger phrases usually
receive stress treatment individually for their component parts, but a postlexical pass
of stress assignment can promote, move, or demote one of the stresses once the two
elements are considered together in a larger domain. In English compounding, as we
have mentioned, the component prosodic words each receive a lexical stress but the
rhythm of the two is adjusted when they come together, promoting the first stress to
English compound
linˈguistics deˈpartment → linˈguistics deˌpartment
And in English phrases, which typically have the most prominent stress on the final
content word, the Rhythm Rule (Hayes 1984) adjusts the tertiary and secondary stresses
of the first word in that phrase to avoid stress clash with the primary stress.
(11) English phrase
(× . ) (×) (× . )
(× . ) (×) (× . )
[Japanese] [language] → [Japanese] [language]
2 The domain of stress assignment: Word-boundedness and frequent collocation
The last type of clitic we need to discuss is ‘free.’ Such clitics simply do not interact
with their host for the purposes of stress assignment. They are not themselves independent prosodic words but do not fall inside of any prosodic word and thus are invisible
to every pass of footing. For instance, Barcelona Catalan verbs (Torres-Tamarit & PonsMoll 2015) show the same stress whether or not they have enclitics, and the stress can
fall three or even four syllables from the end of the word, in contravention to the generalization that there is a three-syllable final stress window and that stress is typically
Barcelona Catalan (Torres-Tamarit & Pons-Moll 2015)
a. ˈkompr-ə
buy -2sg.imp
b. ˈkompr-ə
buy -2sg.imp=1sgdat=3sg.f.acc
‘Buy it/her for me!’
c. ˈkompr-ə
buy -2sg.imp=1sgdat=3sg.m.acc
‘Buy it/him for me!’
The parallel for larger units is this: in §4.4 we will see that while some Macedonian
prepositions are inherently stressless and are included in a stress-assignment domain
with their objects (i.e. they are ‘internal’), others always remain stressless and do not
form a domain with their objects (i.e. they are free).
Even though stress is very often iterative within a word and its stem-level affixes, it
can be unable to cross even the relatively weak boundary of a word level suffix. Consider
for instance the lack of movement of stress from the base in English words suffixed by
word-level suffixes such as #hood and #ly, and the concomitant location of stress outside
the lexically mandated window of the last three syllables:
(13) English
a. ˈdialect#hood (c.f. ˈdialect, ˈdiaˈlect-al)
b. ˈmanifestly (c.f. ˈmanifest, ˌmanifesˈtation)
If even an affix can be outside the domain of stress assignment and other phonological
processes, it is not surprising that stress assignment often does not count clitics on the
initial pass (or ever), let alone treat them as single domain compound words, words
plus function items leaning on them, or the strings of prosodically independent content
Let us tie these increasingly unlikely domains for stress assignment to frequency of
collocation. Cohesive affixes are typically less productive and more selective with respect
Ellen Kaisse
to the stems they combine with than are non-cohesive affixes, which are somewhat cliticlike. There is thus a hierarchy of selectiveness tied to frequency of collocation:4
Hierarchy of selectiveness
• stem-level cohering affixes
• word-level non-cohering affixes
• clitics
• compound words
• words in fixed expressions
• collocations involving closed class items, such as prepositions, even if they
are polysyllabic
• truly novel or unpredictable collocations that are clearly not listed in the
For instance, Fabb (1988) discovered that the English stem-level adjective-forming suffix -ous has very narrow sectional restrictions. It cannot attach after any other affix at
all. It’s clear that -ous is stem level because it affects the stress of its base (ˈ moment,
moˈ mentous), triggers Trisyllabic Laxing (ˈ ōmen, ˈ ŏminous) and meets the various other
tests for cohesion in the literature. Similarly, the stem-level adjective-forming suffix ic attaches only after the suffix -ist (capitalistic), and the stem-level adjective-forming
suffix -ary attaches only after the suffix -ion (revolutionary). On the other hand, wordlevel adjective-forming #y can attach productively to virtually any noun (chocolate#y, or
protein#y, recently heard on a yoghurt commercial), and it can occur in novel coinages
after most other noun-forming suffixes (musi-cian#y, profess-or#y.) Like clitic-host semantics, the meanings of words with word-level affixes are usually transparent. Finally,
notice that ˈchocolate#y has primary stress in the same place as the base ˈchocolate, even
though, if ˈchocolate is pronounced with three syllables, stress in ˈchocolate#y falls in an
otherwise impossible pre-antepenultimate position. The same is true for ˈmanifestly (13b).
Non-cohesive affixes are in a sense invisible to the lexical phonology of the language and
can violate various of its requirements.
Clitics are even less selective than word-level affixes. For example, many Romance,
Greek and Slavic clitics attach to any verb and after any set of affixes. Similarly, the
English contracted auxiliary =s can be found leaning leftward on words of almost any
a. That man’s a linguist;
b. The thing you are leaning on’s not safe;
c. What you think’s not important;
A referee raises the excellent question of “whether this hierarchy predicts an implicational typology of
stress domains, e.g. if a language treats compound words as a single stress domain, then clitics and all
affixes should be parsed within the same stress domain as their host.” I do not know the answer to this
question, though it does seem to be true for Modern Greek (discussed for instance in Anderson 2011 as
having affixal clitics which affect the stress of their hosts) and for Macedonian.
2 The domain of stress assignment: Word-boundedness and frequent collocation
d. Skin that looks pink’s an indication of good circulation.
And both within compounds and within sentences almost any word can be followed by
almost any other word – there are almost no selectional requirements. However, some
words belong to closed classes – prepositions, relative and interrogative elements, and
other function words and therefore will recur sequences more frequently. This cline lines
up with the likelihood of two elements being taken into a single stress domain.
The hierarchy in (14) will take us through the extended domain cases of Macedonian,
where prepositions – even polysyllabic and semantically rich ones – are unstressable on
their own, always forming a stress domain with their complements or being unstressable,
and where frequent collocations of content words, particularly involving frequent words
or collocations with unpredictable or frozen semantics, are stressed as if they were single
3 Compounds and the stress domain continuum
The continuum of domain size we have observed for clitics continues outward into compounds. To review, there are clitics of various sorts: the ‘internal’ type which is considered in the first pass of stress assignment, the more familiar type which figures in a
second pass that takes into account previously assigned stresses, and a third, free type,
which is never counted for stress. My impression is that it is even more uncommon
to find compound words – i.e. words made of otherwise prosodically independent elements – forming a single domain for the first pass of footing. Rather, as we noted earlier,
members of a compound are rather like affixal clitics, where a postlexical instantiation
of rhythm may adjust the stresses on the basis of the newly available material but does
not erase earlier, lexically assigned footing. But this is not always the case.
Modern Greek demonstrates the comparatively rare type of compounding where a
compound word is stressed as a single domain. It has been shown instrumentally by
Athanasopoulou & Vogel (2014) that the well-known traditional description, sharpened
in Ralli (2013), of a single, usually antepenultimate primary stress, is correct. The stress
is placed without regard to where the stresses would fall in the individual members in
isolation. A compounding morpheme -o is often inserted at the end of the first member
(replacing any final inflectional ending) and stress falls on one of the last three syllables
of the whole compound, most commonly the antepenult:
(16) Modern Greek (Athanasopoulou & Vogel 2014; Ralli 2013)
a. ˈlik -os
b. ˈskil-os
c. liˈk -oskil-o
Ellen Kaisse
d. ˈriz -i
e. ˈɣala
f. riˈz -oɣal -o
‘rice pudding’
Tokyo Japanese (Poser 1990; Kubozono 2008) has a similar phenomenon whereby a single
pitch accent is assigned within a compound based, they argue, on foot structure.
A referee points out that English compounds, especially those ending in the element
-man, can sometimes have only a single stress, reminiscent of the Greek case here and
of some Macedonian cases to come. While I have not found a published study on this
phenomenon, Language Log (Liberman 2015) has an informative post that discusses the
unstressed, hence reduced, final vowels in such words as fireman, clansman, gentleman
versus the full final vowels in words like caveman, handyman, and weatherman. A lively
set of reader responses about which words have a schwa versus the expected compound
stress and full final vowel seems to lead to the conclusion that the longer a -man compound has been in English, the more likely it is to have a single stress. The situation here
may be the demotion of a compound to a single word over time rather than a Greek-like
treatment of a compound word as a single stress domain, but it falls under the general
rubric of familiarity breeding unitary stress domains. However, the reader consensus is
that there is no simple connection to actual contemporary frequency in the behavior of
I had believed that the Greek case was as far as regular stress domains ever extend.
However, there is at least one case I am now aware of where the domain extends into
some combinations of prosodically independent words – Literary Macedonian.5
4 Enlarged stress domains in Macedonian
4.1 Overview
We have seen that while stem level affixes virtually always are taken into account in
the first pass of foot building, many languages have stress neutral word-level affixes like
English #ly, #ness, that are not part of the stress domain. Indeed, stress neutrality seems
to be a recurrent, if not definitional characteristic of non-cohesiveness. This suggests that
rhythm is not easily maintained or phonologized over large domains. Continuing along
this cline, we saw that there are clitics of various sorts, some of which are considered
in the first pass of stress assignment, some in a second pass that takes into account
previously assigned stresses, and some of which are never counted for stress. Finally, we
saw that compound words are only rarely the domain of initial footing.
I am very grateful to Ryan Bennett for bringing case to my attention.
2 The domain of stress assignment: Word-boundedness and frequent collocation
A survey of stress assignment, beginning with the compendious Hayes (1995) confirms the general impression that stress-assigning processes are lexical, not postlexical.
But there is an exception, well-known among Slavicists, that in a sense proves the rule.
Macedonian as described by Lunt (1952) and Koneski (1976) (analyzed in generative terms
by Franks 1987; 1989 can build initial feet over certain units larger than the word. The
example treated by Lunt and Koneski comes from what is termed ‘Literary Macedonian.’
At first I was concerned that this might be an artificial language and that the cases of
large domain stress could be the creation of scholars. However, Lunt (1952: 5–6) and
Franks (1987) explain that the literary language is simply a pan-dialectal creation-bycommission that takes features from several western and central dialects but does not
invent them. The large domain stress effects are found in several dialects, though some
of the details differ from spoken dialect to spoken dialect.
Literary Macedonian has regular antepenultimate stress:
(17) Literary Macedonian (Lunt 1952; Franks 1987; 1989)
a. ˈbeseda
b. beˈseda-ta
c. voˈdeniʧar
d. vodeniˈʧar-i -te
As is usual in such systems, monosyllables are stressed and disyllables are stressed on
the penult. (See Halle & Vergnaud 1987: 53 for a full analysis.) But some syntactically
complex strings can be stressed as single units, termed ‘enlarged stress domains’. These
domains are not, for the most part, unfamiliar to phonologists, as they involve potentially prosodically deficient function words such as negative or interrogative particles
and pronouns plus the following word. However, there are also some strings of modifiers plus nouns. There are also polysyllabic propositions, which are cross-linguistically
unlikely to be prosodically deficient (i.e., to be clitics), yet are stressed as a unit with
their objects. I summarize Lunt’s list in (18):
(18) Literary Macedonian enlarged domains (Lunt 1952: 23–25)
a. monosyllabic words which have no accent of their own, both proclitic and
enclitic. The proclitics include personal pronouns, particles and prepositions.
The enclitics include definite articles and certain indirect personal pronouns.
b. the negative particle plus the verb, and any elements that fall between them,
such as the present tense of the verb ‘to be,’ even though these normally have
their own accent.
c. the interrogatives meaning ‘what,’ ‘how,’ ‘where,’ and ‘how many,’ plus the
following verb, and any stressless elements between them.
Ellen Kaisse
d. prepositions and their pronominal objects.
e. prepositions “when used with concrete, spatial meanings” “and in a number
of set phrases” when their object is non-definite.
f. a numeral and the noun it modifies.
g. some combinations of adjectives and the nouns they modify.
Let us begin with (18g), which is the most typologically unusual. We will return to the
also-surprising prepositions 18d) in §4.3.
4.2 Prenominal adjectives
It is worth quoting Lunt’s remarks about the combination of adjectives and nouns in
their entirety.
The combination of adjective+substantive under a single accent is common to many,
but not all, of the central dialects on which the literary language is based, and in
any case it is not productive. Such a shift of accent is impossible if either the noun
or adjective come from outside the narrow sphere of daily life. Therefore this usage
is not recommended. Conversational practice is extremely varied. Place names
tend to keep the old [i.e. single domain, antepenultimate-EMK] accent… Often used
combinations tend to keep the single accent: ‘soured milk’ [yoghurt], ‘dry grapes=
raisins’, ‘the left foot,’ ‘the lower door,’ ‘(he didn’t see a) living soul’. Still one usually
hears [the words stressed as separate domains]. Only with the numbers and perhaps
a few fixed phrases [dry grapes] is the single stress widespread in the speech of
Macedonian intellectuals. (Lunt 1952: 24–25)
What we should note then, is that open class content words are only grouped into a
single stress domain when the words are frequent (“narrow sphere of everyday life”),
especially when such words are also in frequent collocation with one another (“oftenused combinations”). And even though this system arose in some of the dialects on
which the new literary language was based, it apparently is not easily learned. Lunt
here reports that the single domain stress on attributive adjective plus noun has not
been taken up by the educated speakers who adopted it.
Franks (1987; 1989) combines Lunt’s examples with those of Koneski (1976) and others
he and colleagues elicited. Here are several representative ones, including those mentioned in the above quotation.
(19) Literary Macedonian (Lunt 1952; Franks 1987: 989)
a. dolnaˈta porta
lower gate
b. kiseˈlo mleko
sour milk
2 The domain of stress assignment: Word-boundedness and frequent collocation
c. suˈvo grozje
dry grapes
d. Crˈvena voda
red water
‘name of a village’
e. ˈstar ʧovek
old man
f. novaˈta kukʲa
new house
The examples in (19) are not argued to be compound words by the various sources, probably because they have the regular syntax of noun phrases and many such as (a), (e)
and (f) have compositional semantics; however, they are strings that that are in frequent
collocation and may have come to take on a specialized, less predictable meaning. Given
the silence of the sources on the question of compound versus phrase, perhaps the most
noncommittal and appropriate term for them is Erman and Warren’s (2000) “prefabs.”
Prefabs are not idioms or lexicalized compounds with unpredictable meaning and peculiar syntax but simply common and conventional collocations that can be stored in the
lexicon while having compositional meaning and normal syntax.
The matter of whether a string of words that occur together frequently is a compound
or a phrase is a vexed one, even for well-studied languages like English. (See Plag et al.
2008 for an overview of the controversy.)
4.3 Another compound or adjective-noun case
Ryan Bennett informs me that some dialects of Modern Irish show a similar phenomenon
in their adjective plus noun pairs. For instance, Mhac An Fhailigh & Éammon (1968)
states for the dialect of Erris that while such a phrase typically has word stress on
each element, frequent collocations can show a single stress. Thus the infrequent [ˈdrɑx
ˈxlɑdəx] ‘bad shore’ has stress on each member, but [ˈʃɑn vʲɑn] ‘old woman,’ has only a
single stress. Mhac An Fhailigh calls these ‘loose’ versus ‘close’ compounds, rather than
phrases vs. compounds, but does not give clear criteria for what defines a compound
versus a phrase. I have not yet found an extended description of this phenomenon in
Irish so I mention it only in passing here.
The adjective plus noun enlarged domains of Macedonian (and perhaps of Erris Irish)
are the exception that proves the rule – phrases are not normally the domain of stress
assignment. If what leads to phonologization of rhythmic tendencies within words is
frequent collocation, it is gratifying that the only extensively described example of an
enlarged domain that I could find is one involving frequent collocation of common words,
and it is consistent with my hypothesis that the Erris Irish seems to accord with the
same generalization. But Lunt tells us that the Macedonian example is hard to learn
Ellen Kaisse
and is being eliminated over time. Stress rules really don’t like to apply outside a single
content word, its affixes (or some of them) and, sometimes, its clitics.
4.4 Macedonian prepositions and their objects
The stress behavior of prepositions in Macedonian is also worth looking at in a bit of
detail. Since prepositions are the heads of phrases and especially since they can be polysyllabic in Macedonian, one might expect they would be independent prosodic words
on their own. Looking at a more familiar case, English monosyllabic prepositions like
to and for are optionally proclitic on their complements, but polysyllabic prepositions
like behind and above are not. But Macedonian prepositions, regardless of their apparent
prosodic heft in terms of syllable count, never receive stress as a domain on their own.
They are, apparently, as prosodically deficient as clitics. Lunt devotes several pages (⁇–
⁇) to the individual behavior of some two dozen prepositions because their stress behaviors are somewhat idiosyncratic. The basic generalization is that prepositions have
or receive no accent of their own. They either form a single stress domain with their
nominal or pronominal complement, acting like internal clitics (20a); or, in some cases,
they behave like free clitics, never receiving a stress of their own nor receiving the single
stress of the enlarged domain (20b).
(20) Literary Macedonian (Franks 1989)
a. okoˈlu selo
near village
‘near the village’
b. otkaj ˈgradot
from direction
‘from the direction (of)’
Prosodic deficiency in prepositions makes sense from the point of view of the frequency
of collocations. Prepositions are closed class items and they require a nominal or pronominal complement. Thus, when they are heard, they are always in collocation with a noun
4.5 Summary
As we mentioned earlier, it is helpful to think of the Macedonian collocations of full
words that are stressed as a single word as the inverse of the Prosodic Word clitics of
Bilua in (7), re-illustrated below in (21). PWd clitics are underlyingly prosodically deficient but can receive their own stress under duress. The Macedonian adjectives are
elements that are not inherently prosodically deficient, even in Macedonian, yet in certain common collocations, they are being treated as a piece of a single prosodic word, like
internal clitics. This is illustrated in (22). From a cross-linguistic perspective, the Macedonian polysyllabic prepositions are a bit unexpected in being inherently prosodically
2 The domain of stress assignment: Word-boundedness and frequent collocation
deficient, parallel either to internal (20a) or free (20b) proclitics. These are illustrated in
the second line of (22) and in (23).
Prosodic independence of underlyingly prosodically deficient clitic
‘he called’
Prosodic dependence of underlyingly prosodically independent words and of
‘internal’ prepositions
‘sour milk’
‘near the village’
(23) Macedonian unstressable preposition parallel to free proclitic
Preposition N
‘from the direction’
5 Conclusion
Why is stress assignment almost always word-bounded? Why is it restricted only to cohering affixes in some languages, and how does it manage to extend outward to clitics in
others? The hypothesis offered in this paper and in Kaisse (forthcoming) is that rhythm,
like long-distance harmony, is grammaticized when certain syllables are heard together
Ellen Kaisse
over and over.6 This usually only happens within words, including their affixes, and,
occasionally, with words plus slightly more independent closed class items such as clitics. Because word-level affixes have fewer selectional restrictions than stem-level ones,
stem-level affixes are the most likely to be included in the initial domain of footing. And
since clitics have even fewer selectional restrictions than word-level affixes, they are
even less likely to be included. Since there are thousands of independent content words,
the collocation of any two is much less likely to have many exemplars. However, both
compound words (as in Modern Greek) and ‘prefabs’ – collocations of frequent words in
frequently heard phrases like ‘old man’ – are more prone to becoming a single domain
than just any sequence of content words. Similarly, collocations of closed class items
like prepositions may cross the border and form a single initial stress domain.
There may also be a functional motivation for the prosodic independence of content
words. Extending the domain for the initial building of feet would eliminate the demarcative function linguists often ascribe to stress. Because stress is culminative, with exactly
one primary stressed syllable per content word, it allows us to identify the independent
lexical words of a phrase. And because stress is often predictably located on an initial,
final, or penultimate syllable, it also allows us to identify word boundaries, helping the
listener locate the end of one word and the beginning of the next.
Let us return to a question I raised in §⁇. Does the regular, alternating stress assignment found within words stem from the grammaticization of rhythmic tendencies in
sentences? Or are rhythmic tendencies in sentences the result of “leakage” from regular
phonology within the word? The counterparts of word-bounded phonological processes
are present and, in the current century, detectable, as statistical, gradient tendencies in
larger domains such as compounds and phrases. Shih (2016) shows that syntactic choices
are gradiently influenced by the rhythmic principles that we see operating obligatorily
inside words as alternating stress assignment, rhythm rules, stress clash avoidance, lapse
avoidance, and so on. She argues that word order and construction choice can be recruited in response to these phonological pressures from inside the lexical phonology,
unearthing small effects in the choice of optional syntactic variants that favor more eurhythmic outputs. Thus she finds that the genitive constrictions X’s Y and Y of X are
skewed toward avoidance of stress clashes and lapses. Similarly, the dative alternation
verb IO DO vs. verb DO to IO trends in a small but discoverable way towards phrases
that are more eurhythmic (Anttila, Adams & Speriosu 2010; Shih 2016). Along the same
lines, Hammond (2013) finds that the Brown and the Buckeye corpora contain statistically fewer instances of underlyingly clashing prenominal adjective-noun pairs than
would be expected, so that adjectives like aloof and unknown, with final primary stress,
are underrepresented when they are prenominal, while adjectives like happy and finite,
with initial primary stress are not. Here we run into a chicken and egg problem. The
more traditional view in historical linguistics is the Neogrammarian one: these pressures
Although tone rules are more powerful in their ability to escape the word, for reasons that, as far as I know,
are not well understood (see footnote 2), word-boundedness is nonetheless certainly robustly attested for
tone rules as well as other iterative processes, and I would propose that the same factors of frequent
collocation should underlie that domain restriction.
2 The domain of stress assignment: Word-boundedness and frequent collocation
are there all along as variation, with the syntax reflecting them before they are phonologized. This is how I envisioned the progression in Kaisse (forthcoming). Along the
same lines, Myers & Padgett (2015) note the tendency of learners to extend phenomena
like utterance-final devoicing to the word-level. Participants in their artificial languagelearning experiments generalized beyond the training data, applying word-final devoicing inside of utterances in novel strings, not just the kind of utterance-final devoicing
they were trained on. Myers & Padgett (2015: 399) conclude that “learners are biased
towards word-based distributional patterns.” They speculate that this is because we hear
and store many more exemplars of words than we do of utterances. Baumann & Ritt
(2012) take a similar view of the direction of development of lexical stress patterns, investigating through game theoretic simulations the lexicalization of stress in English bisyllabic words, which can have either initial (ˈlentil, ˈresearchN ) or final (hoˈtel, reˈsearchV )
prominence. They argue that “words adopt, on average, those stress patterns that produce, on average, the best possible phrasal level rhythm.” But Shih and Martin’s views
are equally plausible: the obligatory, regular stress system of the language spreads from
words into gradient choices in the syntax. I suspect that both directions are operating at
the same time, since language change results from a constant push and pull of variation
and optimization.
Whatever the direction of change, we have seen that it is rare for rhythmic tendencies beyond the word to become phonologized and to be reflected as foot construction
processes that take material outside of words as their domain. I have argued that there
are simply not enough exemplars of adjacent words heard repeatedly together to result
in the reification of such a large domain.
I am grateful to Ryan Bennett, Larry Hyman, Andrew Livingston, David Odden, the
referees for this paper, and the audience at the Annual Meeting on Phonology 2015 (University of British Columbia).
Anderson, Stephen R. 1992. A-morphous morphology. Cambridge: Cambridge University
Anderson, Stephen R. 2011. Clitics. In Marc van Oostendorp, Colin J. Ewen, Elizabeth
Hume & Keren Rice (eds.), The Blackwell companion to phonology, 2002–2018. Chichester: Wiley-Blackwell.
Anttila, Arto, Matthew Adams & Michael Speriosu. 2010. The role of prosody in the English dative alternation. Language and Cognitive Processes 25(7–9). 946–981.
Athanasopoulou, Angeliki & Irene Vogel. 2014. Compounds in Modern Greek. Journal
of the Acoustical Society of America 136(4). 2176.
Ellen Kaisse
Barnes, Jonathan. 2006. Strength and weakness at the interface: Positional neutralization
in phonetics and phonology. Berlin: Mouton de Gruyter.
Baumann, Andreas & Nikolaus Ritt. 2012. Transferring mathematics to English Studies. In Manfred Markus & Herbert Schendl (eds.), Transfer in English Studies, 219–239.
Wien: Braumüller.
Bermudez-Otero, Ricardo. 2015. Amphichronic explanation and the life cycle of phonological processes. In Patrick Honeybone & Joseph Salmons (eds.), The Oxford handbook
of historical phonology, 374–399. Oxford: Oxford University Press.
Bybee, Joan. 2006. From usage to grammar: The mind’s response to repetition. Language
82(4). 711–733.
Chomsky, Noam & Morris Halle. 1968. The sound pattern of English. New York: Harper &
Erman, Britt & Beatrice Warren. 2000. The idiom principle and the open choice principle.
Text 20(1). 29–62.
Fabb, Nigel. 1988. English suffixation is constrained only by selectional restrictions. Natural Language and Linguistic Theory 6(4). 527–539.
Fhailigh, Mhac An & Éammon. 1968. The Irish of Erris, Co. Mayo: A phonemic study.
Dublin: Dublin Institute for Advanced Studies.
Flemming, Edward. 1997. Phonetic detail in phonology: Towards a unified account of
assimilation and coarticulation. In K. Suzuki & D. Elzinga (eds.), Proceedings of the 1995
Southwestern Workshop in Optimality Theory (SWOT). Tucson: University of Arizona.
Franks, Steven L. 1987. Regular and irregular stress in macedonian. International Journal
of Slavic Linguistics and Poetics 35–36. 93–142.
Franks, Steven L. 1989. The monosyllabic head effect. Natural Language and Linguistic
Theory 7(4). 551–563.
Halle, Morris & Jean-Roger Vergnaud. 1987. An essay on stress. Cambridge, MA: MIT
Hammond, Michael. 2013. Data input optimization in English. Journal of the Phonetic
Society of Japan 17. 26–38.
Hansson, Gunnar Ó. 2010. Consonant harmony: Long-distance interaction in phonology.
University of California Publications in Linguistics, 145. Berkeley, CA: University of
California Press.
Hayes, Bruce. 1984. The phonology of rhythm in English. Linguistic Inquiry 15(1). 33–74.
Hayes, Bruce. 1995. Metrical stress theory: Principles and case studies. Chicago: University
of Chicago Press.
Hyman, Larry. 2011. Tone is different. In John Goldsmith, Jason Riggle & Alan C. L. Yu
(eds.), The handbook of phonological theory, 197–239. Somerset, NJ: Wiley.
Inkelas, Sharon, Jonathan Barnes, Jeffrey Good, Darya Kavitskaya, Orhan Orgun, Ronald
Sprouse & Alan Yu. 2001. Stress and vowel-to-vowel coarticulation in Turkish. Presented
at the annual meeting of the Linguistic Society of America.
Kaisse, Ellen M. 1993. Rule reordering and rule generalization in lexical phonology: A
reconsideration. In Sharon Hargus & Ellen M. Kaisse (eds.), Studies in lexical phonology,
343–363. San Diego: Academic Press.
2 The domain of stress assignment: Word-boundedness and frequent collocation
Kaisse, Ellen M. 2005. Word formation and phonology. In Pavel Štekauer & Rochelle
Lieber (eds.), Handbook of word-formation, 25–47. Studies in natural language and linguistic theory 64. New York: Springer Verlag.
Kaisse, Ellen M. Forthcoming. What kinds of processes are postlexical? And how powerful are they? In Hongming Zhang (ed.), Prosodic studies: Challenges and prospects.
London: Routledge.
Kiparsky, Paul. 1982. Lexical morphology and phonology. In Linguistic Society of Korea
(ed.), Linguistics in the morning calm: Selected papers from SICOL–1981, 3–91. Seoul:
Koneski, Bozidar. 1976. Gramatika na makedonskiot literaturen jazik. Skopje: Kultura.
Kubozono, Haruo. 2008. Japanese accent. In Shigeru Miyagawa & Mamoru Saito (eds.),
The Oxford handbook of Japanese linguistics, 165–191. Oxford: Oxford University Press.
Liberman, Mark. 2015. Man: reduced or not? Language Log. Accessed January 18, 2017.
Lunt, Horace G. 1952. A grammar of the Macedonian literary language. Skopje: Drzavno
Knigoizdatelstvo na NR Makedonija.
Martin, Andrew T. 2007. The evolving lexicon. Los Angeles, CA: University of California
Myers, Scott & Jaye Padgett. 2015. Domain generalization in artificial language learning.
Phonology 31(3). 399–433.
Newman, Stanley. 1944. Yokuts language of California. Vol. 2 (Viking Fund publications
in anthropology). New York.
Obata, Kazuko. 2003. A grammar of Bilua: A Papuan language of the Solomon Islands.
Vol. 540. Canberra: Pacific linguistics.
Ohala, John J. 1992. What’s cognitive, what’s not, in sound change. In Günter Kellermann & Michael Morrissey (eds.), Diachrony within synchrony: Language history and
cognition, 309–355. Frankfurt: Peter Lang.
Öhman, Sven E. G. 1966. Coarticulation in CVC utterances: Spectrographic measurements. Journal of the Acoustical Society of America 39(1). 151–168.
Peperkamp, Sharon. 1997. Prosodic words (HIL dissertations 34). The Hague: Holland Academic Graphics.
Plag, Ingo, Gero Kunter, Sabine Lappe & Maria Braun. 2008. The role of semantics, argument structure, and lexicalization in compound stress assignment in English. Language 84(4). 760–794.
Poser, William. 1990. Evidence for foot structure in Japanese. Language 66(1). 78–105.
Ralli, Angela. 2013. Compounding in Modern Greek. New York: Springer Verlag.
Selkirk, Elisabeth O. 1995. The prosodic structure of function words. In Jill N. Beckman,
Laura Walsh Dickey & Suzanne Urbanczyk (eds.), University of Massachusetts occasional papers 18: papers in Optimality Theory, 439–469. Amherst: GLSA.
Shih, Stephanie S. 2016. Phonological influences in syntactic alternations. In Vera Gribanova (ed.), The morphosyntax-phonology connection. (pages not yet available). Oxford:
Oxford University Press.
Ellen Kaisse
Torres-Tamarit, Francesc & Clàudia Pons-Moll. 2015. Enclitic-triggered stress shift in Catalan. Paper delivered at Going Romance, 29. Nijmegen: Radboud University.
Zwicky, Arnold M. & Geoffrey K. Pullum. 1983. Cliticization vs. inflection: English N’T.
Language 59(3). 502–513.
Chapter 3
Compensatory lengthening and
structure preservation revisited yet
Darya Kavitskaya
University of California, Berkeley
In their seminal paper, de Chene & Anderson (1979) make a strong claim that pre-existing
vowel length contrast is a necessary condition for the phonologization of vowel length
through compensatory lengthening. Compensatory lengthening is thus predicted to be always a structure-preserving change. Since that time, the claim has been challenged in numerous works (Gess 1998; Hock 1986; Morin 1992, among others). A closer examination
of the cited counterexamples to de Chene and Anderson’s claim reveals certain generalizations. Some apparent counterexamples, such as Samothraki Greek (Kiparsky 2011), involve
the full vocalization stage of the consonant with the subsequent coalescence of that consonant with the preceding vowel. In other cases, such as Old French (Gess 1998) and Komi
Ižma (Hausenberg 1998), heterosyllabic or heteromorphemic identical vowel sequences are
attested elsewhere in the language. The former cases involve the reanalysis of vowel length
before weakened consonants that is indeed strengthened by the independent existence of
the vowel length contrast in the languages in question, in support of de Chene and Anderson’s claim. The former cases are not truly compensatory, and phonemic vowel length is
introduced into the language through coalescence.
1 Introduction
In their seminal paper on compensatory lengthening (CL), de Chene & Anderson (1979)
make a strong claim that the independent existence of a vowel length contrast is a necessary condition for the phonologization of vowel length through compensatory lengthening. CL is thus predicted to be always a structure-preserving change that cannot introduce contrastive vowel length into a language. Since that time, the generalization
in its stronger version (certain sound changes are always structure preserving) or in its
weaker version (structure preservation is a tendency in sound change) has been accepted
and developed by linguists otherwise advocating very diverse and sometimes incompatible approaches to sound change, in particular, in research programs by Paul Kiparsky
Darya Kavitskaya. 2017. Compensatory lengthening and structure preservation revisited yet again. In Claire Bowern, Laurence Horn & Raffaella Zanuttini (eds.),
On looking into words (and beyond), 41–58. Berlin: Language Science Press.
Darya Kavitskaya
(Kiparsky 1995; 2003) and Juliette Blevins (Blevins 2004a; 2009). However, the generalization has also been challenged in several works. For instance, Gess (1998) takes issue with
de Chene and Anderson’s claim, suggesting that in general, “structure preservation is irrelevant as a theoretical construct” and proceeds to argue that de Chene and Anderson’s
interpretation of the Old French data, which is their main example, is incorrect, and that
in Old French compensatory lengthening happened before the introduction of the other
sources of length distinction into the language, contrary to de Chene and Anderson’s
analysis. CL through onset loss, such as in Samothraki Greek (Topintzi 2006; Kiparsky
2011; Katsika & Kavitskaya 2015), is also a potential counterexample to the claim that CL
is a structure-preserving change, along with the case of Occitan (Morin 1992). In other
languages without pre-existing vowel length contrast, such as Andalusian Spanish (Hock
1986), Ilokano (Hayes 1989) and the Ngajan dialect of Dyirbal (Dixon 1990), vowel length
that is the result of CL remains allophonic and predictable. In yet another type of cases,
such as Komi Ižma (Harms 1967; 1968; Hausenberg 1998), vowel length from CL appears
to be quasi-phonemic and on its way to phonologization.
CL is a common sound change that has occurred independently in many languages
across the world, and only a few potential counterexamples to de Chene and Anderson’s
claim have been reported. In principle, we could have been done simply restating this observation that supports a weaker but less controversial claim that there is a tendency for
CL to occur in languages with pre-existing vowel length, in the spirit of proposals about
structure-preserving sound change by either Kiparsky (2003) or Blevins (2009), but we
will proceed to examining the most widely discussed counterexamples to de Chene and
Anderson’s claim. A closer examination of these counterexamples reveals certain generalizations. The working analyses of some cases proposed in the literature involve the full
vocalization stage of the consonant with the subsequent coalescence with the preceding
vowel, such as in Samothraki Greek (Sumner 1999; Kiparsky 2011). In other cases, such as
Old French (Gess 1998) and Komi Ižma (Hausenberg 1998), heterosyllabic or heteromorphemic long vowels (or rather vowel sequences) are attested elsewhere in the language.
We shall argue that the cases of CL that do not involve full vocalization (Hayes 1989; Kavitskaya 2002) are in a sense truly compensatory, as opposed to instances of consonant
vocalization and subsequent vowel coalescence. The former cases involve the reanalysis
of vowel length before weakened consonants that is indeed strengthened by the independent existence of the vowel length contrast in the languages in question. In the latter
cases, phonemic vowel length is introduced into the language through coalescence.
2 The problem
CL through consonant loss is defined as a process whereby a vowel lengthens in compensation for the loss of a tautosyllabic consonant. CL through coda loss is the most
typologically wide spread process, as either a diachronic change or a synchronic alternation.1 An example of this kind of CL in the Ižma dialect of Komi (a Uralic language
I do not address CL through vowel loss in this paper.
3 Compensatory lengthening and structure preservation revisited yet again
of the Permian subgroup) is shown in Table 1 (Harms 1967; 1968; de Chene & Anderson
Table 1: CL through coda loss in Komi Ižma (after Harms 1968: 104).
Past 1sg
In Komi Ižma, the lateral /l/ deletes in the coda position with the lengthening of the
preceding vowel, as illustrated in (b) of Table 1.2 De Chene & Anderson (1979) propose
that CL through consonant loss should be analyzed as an instance of the conversion of
coda consonants, /l/ in the case of Komi Ižma, to glides (either semivocalic or laryngeal),
/w/ in the case of Komi Ižma, with the subsequent monophthongization of the resulting vowel-glide sequence in the syllable nucleus, as in, for example, *kɨl.nɨ > *kɨw.nɨ >
kɨː.nɨ ‘to hear’, with the intermediate stage unattested in Komi Ižma but present in other
dialects of Komi, such as Vychegda Komi (Lytkin 1966; Lytkin & Teplyashina 1976).
De Chene and Anderson (1979: 508) emphasize that their account is phonetic in nature
and accounts for CL as a historical sound change, and not as a synchronic alternation:
We will argue that these processes can be understood as the transition of the consonant, through loss or reduction of its occlusion, to an eventual glide G. It is the
monophthongization of the resulting sequence (X)VG(Y) which gives rise to a syllable nucleus that is interpreted as distinctively long. In consequence, cases of apparent compensatory lengthening can be analysed (as far as their phonetic bases
are concerned) as a combination of consonantal weakening in certain positions
followed by monophthongization; and compensatory lengthening per se can be
eliminated as an independent member of any inventory of phonetic process-types.
This insight into the phonetics of CL serves as the basis for the analysis developed in
Kavitskaya (2002), who maintains that CL is the result of the reanalysis of the longer
phonetic duration of vowels as phonological length with the loss of tautosyllabic consonants. Kavitskaya (2002) maintains that vowels are more likely to be reanalyzed as
phonologically long in the environment of more sonorous consonants after the loss of
the said consonants, which makes the differences in vowel length unpredictable. De Chene and Anderson’s (1979) analysis of CL as a process whereby consonants weaken to
glides supports Kavitskaya’s phonetic analysis, which is shown in Table 2.
The schematic representation in Table 2 considers two possible situations where the
consonants X and Y are lost. Prior to the deletion of the consonants, both vowels are
The lengthened /a/ surfaces as [oː].
Darya Kavitskaya
Table 2: CL through coda loss (Kavitskaya 2002: 9).
Stage 1
(before consonant loss)
Stage 2
(consonant loss)
correctly analyzed as phonologically short. In the case when the listener mishears the
more sonorous consonant X as absent, the longer transitions are reinterpreted as a part
of the vowel, which is subsequently reanalyzed as long. The vocalic transitions to the
less sonorant consonant Y are shorter, and with the loss of this consonant, there is no
reinterpretation of vowel length based on its duration. The divide between X and Y is
arbitrary, and the more sonorous the deleting consonant is, the more likely its deletion
to be compensated by the lengthening of the vowel.
Several later accounts of CL are mostly phonological. The most well-known of those
is an account by Hayes (1989), who analyzes CL through consonant loss as the deletion
of a weight-bearing coda while preserving its weight and reassigning it to the preceding
vowel, as illustrated in (1) for Komi Ižma. The account holds that when the underlying
coda /l/ is deleted, its mora is left behind (in an intermediate stage) and spreads to the
preceding vowel, making it bimoraic and thus long:
CL through coda loss in Komi Ižma (after Hayes 1989)
μ μ
μ →
μ μ
μ →
μ μ
k ɨ l n ɨ
k ɨ
n ɨ
k ɨː
n ɨ
The reason for the necessity of the phonetic explanation in de Chene and Anderson’s
analysis and its conspicuous absence from Hayes analysis lies in the difference between
the general approaches to CL taken by these two accounts. De Chene and Anderson’s account carefully distinguishes between a sequence of phonetic processes that comprises
the weakening of occlusion followed by a monophthongization of the resulting vowelglide sequences and the phonological reinterpretation of some of the outputs of this
monophthongization as long vowels. While Hayes uses historical examples to illustrate
his points (one of the examples being Attic Greek, where CL is arguably only a historical
process with no synchronic alternations), the account he proposes is synchronic in nature and does not consider either phonetic or phonological stages of the sound change
analyzed by de Chene and Anderson.
3 Compensatory lengthening and structure preservation revisited yet again
One of the important predictions of de Chene and Anderson’s account concerns the
systemic constraints on the phonologization of vowel length. They propose that the
phonologization of vowel lengthening as a result of CL can happen if and only if the
language in question has a pre-existing vowel length contrast. This prediction does not
follow directly from de Chene and Anderson’s analysis, nor it is necessary for the accounts in the spirit of Hayes. It a sense, it is not a prediction per se, but rather a generalization about the nature of CL as a sound change. In the following sections, I will
consider several counterexamples to this claim, discuss the similarities among these examples, and offer some speculation on why de Chene and Anderson’s generalization is
at least a tendency in the languages of the world.
3 CL with no pre-existing vowel length: Apparent
As can be inferred from an (admittedly small) survey of languages with CL in Kavitskaya
(2002), CL is more often a structure-preserving sound change. In the majority of the
cases of CL, this tendency indeed holds: out of 80 languages with historical CL sound
changes listed in Kavitskaya’s (2002) survey, 72 or 90% occur in languages with preexisting long/short vowel contrasts, while only 8 or 10% are found in languages without
a pre-existing vowel length contrast.3 These 8 languages constitute counterexamples
to the stronger version of the claim, which holds that CL as a sound change is always
structure-preserving. However, first, the presence of counterexamples does not make
the tendency false (it is just not a universal). Second, there seems to be an important
difference between the cases that are structure-preserving and the cases in which vowel
length (mostly allophonic) is potentially introduced into a language through CL.
3.1 Old French
Old French is the central example used by de Chene and Anderson to illustrate that CL
as a sound change does not happen unless contrastive vowel length is independently
present in the language. According to de Chene & Anderson (1979: 527–528), stated
after Pope (1934: 79, 191), the diphthong [aw], inherited both from Indo-European and
from Vulgar Latin, monophthongized to a short [o] in French by the middle of the 9th
century, as in (2a). The loss of other consonants, such as velars before l and n and p/b
and t/d before p/b, t/d, and s, that took place at approximately the same time, was not
accompanied by CL either, as exemplified in (2b):
This information is not explicitly present in Kavitskaya (2002) and was compiled by Blevins (2009).
Darya Kavitskaya
a. Monophthongization of [aw] to [o] in Old French (circa 850 AD)
or < aurum ‘gold’
oser < ausare ‘to dare’
forge ‘forge’ < *faurga < fabrica ‘workshop’
parole < paraula < parab(o)la ‘word’
b. Loss of consonants g, k, p, and d in Old French (before 850 AD)
agneau [aɲo] < agnellum ‘lamb’
maille [may] < mac(u)lam ‘stain’
route ‘road’ < (via) rupta ‘broken road’
après ‘after’ < adpressum ‘near’
Another wave of monophthongization, presumably through the weakening of the
coda [l] to a labiovelar glide, happened in Old French by the 16th century, this time
resulting in a long [oː], as illustrated in (3a). The loss of other pre-consonantal consonants, such as nasals (complete by the middle of the 16th century) and fricatives s and z
(earlier), was accompanied by CL. Examples of the loss of fricatives are in (3b), and CL
through the loss of nasals is illustrated in (3c).
The examples in (3b) show the orthographic s that was preserved in such words until
1740 (de Chene & Anderson 1979: 520). However, Pope (1934: 151) mentions that 12th
century poetry suggests that the fricative had begun to drop before voiced consonants
by this period. French loanwords in English, such as blame, male, and isle, do not have a
pronounced [s], which adds more support to this conclusion:
a. Monophthongization of [aw] < [al] to [oː] in Old French (16th century)
autre [o:tr] < alterum ‘other’
aube [o:b] ‘dawn’ < alba ‘white’ (
b. Loss of fricatives with CL in Old French (12th century)
mȇler (ModFrench) < mesler (Old French) < misculāre ‘to mix’
île (ModFrench) < isle (Old French) < insula ‘island’
c. Loss of nasals with CL in Old French (16th century)
fendre [fãːdr] < findere ‘to split’
rompre [rɔ̃ːpr] < rumpere ‘to break’
De Chene and Anderson claim that the difference between the outcomes of the two
Old French monophthongizations, as well as between the loss of consonants without
and with vowel lengthening, lies in the fact that in the 9th century, the Old French
vowel system did not exhibit contrastive vowel length, and so vowels did not lengthen
as a response to the loss of consonants, while by some time in the 12th to 16th century
a vowel length distinction was introduced into the system independently, and this preexisting vowel length contrast made it possible for the reanalysis of the vowels that
preceded the lost consonants as long.
It is argued in de Chene (1985) that languages typically acquire vowel length through
vowel coalescence (dubbed as vowel hiatus or geminate vowel clusters, de Chene &
Anderson 1979: 520). De Chene and Anderson (1979) state that French obeys this rule
3 Compensatory lengthening and structure preservation revisited yet again
and acquires long vowels through the deletion of intervocalic consonants and subsequent vowel coalescence in the period between the changes exemplified in (2) and (3).
The examples of consonant loss and vowel coalescence are presented in Table 3.
Table 3: Intervocalic consonant loss between identical vowels (de Chene &
Anderson 1979 after Pope 1934).
Modern French
Old French
‘to yawn’
‘suitable’ (
Gess (1998) takes issue with de Chene and Anderson’s claim that CL is only possible
in languages with a preexisting vowel length contrast. He argues with the claim on the
basis of the evidence from Old French. The objection is that the putative long vowels are
treated as disyllabic in 12th and 13th century poetry in Old French, as shown in (4) for one
of the examples in Table 3. From the scansion of the octosyllabic line in (4), it is evident
that graal ‘dish’ consists of two syllables for the purposes of poetic syllabification:
(4) Le Roman de Perceval, late 12th century (Roach 1959: 3, 11, 76–77; Gess 1998: 357)
Ce est le contes de graal,
1 2 3 4 5 6 7 8
‘This is the story of the Grail,’
Dont li quens li bailla le livre.
‘About which the count gave him the book.’
The evidence from the scansion provided by Gess (1998) is questionable since syllabification in poetry is often conservative and reflects an earlier stage of the language.
The scansion is also consistent with the possibility that vowel coalescence has already
happened and long vowels scan as two syllables, with a poetic line becoming moracounting rather than syllable-counting (see discussion in de Chene (1985: 76, 84ff) about
such developments in Japanese and Tongan).4 If this is the case, then Old French does
not constitute a counterexample to de Chene and Anderson’s generalization.
The metrical evidence presented by Gess (1998) is thus inconclusive. However, even
if Gess’ interpretation is correct and indeed his examples illustrate that at the time of
CL in Old French there were heterosyllabic sequences of identical vowels, it could have
been sufficient to strengthen the possibility of CL, as we shall further discuss for another
example in the next section.
I am grateful to a reviewer for the discussion of this point.
Darya Kavitskaya
3.2 Komi Ižma
It would be informative now to return to Komi Ižma, which does not have contrastive
long vowels in the inventory, or any other allophonically long vowels, except for those
that are derived by CL (Lytkin 1966; Lytkin & Teplyashina 1976; Hausenberg 1998). Thus,
in principle, Komi Ižma constitutes a counterexample to de Chene and Anderson’s claim
interpreted broadly, as noticed by Gess (1998). The forms in Table 4 illustrate CL alternations in Komi Ižma:
Table 4: CL through coda loss in Komi Ižma (after Harms 1968: 104–105).
Past 1sg
‘to hear’
‘to stand’
The deletion of l in Komi Ižma went through the stage of the loss of the occlusion
of the liquid to the labiovelar glide w, followed by the monophthongization of the Vw
sequence.5 The diphthongal stage is synchronically attested in related dialects of Komi,
spoken in Vychegda and Syktyvkar, and there is also a dialect group in Komi that preserves the lateral (cf. vɘː /vɘl/ ‘horse’ in Komi Ižma vs. vɘv /vɘl/ ‘horse’ in Vychegda
Komi vs. vɘl /vɘl/ ‘horse’ in Komi Yazva) (Lytkin 1966: 44–49, Lytkin & Teplyashina 1976:
De Chene and Anderson (1979) maintain that Komi Ižma data do not counterexemplify their generalization since the language has heteromorphemic long vowels (or vowel
sequences), so sufficient contrast in vowel duration is present for CL to go through.
Hausenberg (1998: 309) states that in dialects long vowels may develop through assimilation in forms like una-an < una-ën ‘many’, baba-as < baba-ïs ‘his wife’.
In a narrower sense, Komi Ižma is not a counterexample since CL does not introduce
vowel length contrast into the language: vowel length is allophonic and predictable, and
even though ‘son’ and ‘cloud’ look like a minimal pair, they are underlyingly /pi/ ‘son’
vs. /pil/ ‘cloud’. Abondolo (1998: 13) calls vowel quantity in Komi Ižma “nascent”, thus
interpreting vowel length distinction in the language as quasi-phonemic and possibly
on its way to phonemicization. However, another view on the facts of Komi Ižma CL is
possible that provides additional evidence in support of de Chene and Anderson’s claim.7
Syllable-final /l/ frequently undergoes vocalization; cf., for instance, l-vocalization in BCS (South Slavic):
beo /bel/ ‘white-masc’ (vs. bela ‘white-fem’), video /videl/ ‘see-past.masc’ (vs. videla ‘see-past-fem’).
Yet another dialect of Komi, Komi Inva, vocalizes /l/ into [w] in all positions (Lytkin 1966: 44–49).
I am much indebted to a reviewer for the following discussion of vowel length and morphology.
3 Compensatory lengthening and structure preservation revisited yet again
These long vowels can arise through either inflection, as in Table 4, or derivation, as in
(5) CL in Komi Ižma (Collinder 1957: 309 via de Chene & Anderson 1979: 525)
‘he hangs (trans), as a cross on one’s breast’
In an important paper that defines the place of morphology in grammar, Anderson
(1982) proposes that the traditional category of inflection is the subset of morphology
that is relevant to the syntax. As a consequence, inflection depends on the results of
syntactic operations and is post-syntactic, while derivation happens before syntax. Thus,
according to Anderson’s model, the units of lexical storage are stems that “include all
internal structure of a derivational sort” (Anderson 1982: 592). Endorsing this approach
amounts to saying that, since long vowels resulting from the addition of derivational
material are robustly attested in Komi Ižma, the language has lexical long vowels even
if none of them are morpheme-internal.
3.3 Samothraki Greek
One of the languages in which CL introduces phonemic vowel length into a system
without pre-existing vowel length contrast is a dialect of Greek spoken on the island of
Samothraki (Newton 1972a,b; Hayes 1989; Katsanis 1996; Sumner 1999; Kavitskaya 2002;
Topintzi 2006; Kiparsky 2011; Katsika & Kavitskaya 2015). Samothraki Greek is not a
usual case of CL in yet another respect since it is the loss of the onset, not the coda, that
triggers tautosyllabic vowel lengthening, as illustrated in Table 5.
In Samothraki Greek, the prevocalic r deletes with the lengthening of the following
vowel in a) the word-initial onset of either stressed or unstressed syllable, as in Table 5a,
and b) after a consonant in a complex onset, both in biconsonantal clusters, as in Table 5b,
and triconsonantal clusters, as in Table 5c, both in stressed and unstressed word-initial
and word-medial/final syllables:
Table 5: CL through onset loss in Samothraki Greek (after Katsika & Kavitskaya
Standard Greek
Samothraki Greek
Darya Kavitskaya
The examples in Table 6 show the synchronic status of r-deletion in Samothraki Greek:
the rhotic surfaces in the coda and as a first consonant in a complex onset, but deletes
intervocalically. On the basis of such alternations, Kiparsky (2011) argues that the presence of r-zero alternations constitutes evidence for the synchronic status of CL in the
Table 6: Alternations in Samothraki Greek (Katsika & Kavitskaya 2015: 7).
‘little hands’
‘little feet’
However, as Katsika & Kavitskaya (2015) point out, there are no synchronic alternations where the deletion of /r/ is accompanied by vowel lengthening. In other words,
there are no attested examples in which one member of a semantically related pair has a
surface [ɾ], while the other exhibits a long vowel as a consequence of the r-deletion. On
the basis of this, Katsika & Kavitskaya (2015) conclude that it would be more accurate
to analyze r-zero alternation as a synchronic process, and CL through the loss of r as a
sound change in Samothraki Greek.
CL through onset loss presents a problem for the theories that treat CL as weight conservation (such as Hayes 1989), which predict that only the deletion of coda consonants
can result in vowel lengthening. It is generally assumed that, unlike codas, onsets cannot
bear weight and do not count as moraic.9 Several such problematic cases, including CL
through onset loss in Samothraki Greek, are reanalyzed in Hayes (1989). Hayes extends
Newton’s (1972) idea that rC clusters underwent vowel epenthesis of the form VrC →
VriC → ViC and proposes that identical vowel epenthesis happened in Cr clusters as
well, yielding CrVi → CVi rVi → CVi ː. The deletion of the intervocalic r could then be
followed by vowel coalescence, just like in other VrV → Vː cases in Samothraki Greek.
However, as shown by Topintzi (2006), the Samothraki Greek CL resists such a reanalysis since the deletion of the word-initial r cannot be accounted for by metathesis.
In addition, Kiparsky (2011) claims that Hayes’ analysis is problematic because it incorrectly predicts the merger of the outputs of the r-deletion from CrV and VrV. While
after the loss of r, the original *rV sequence where the vowel is accented becomes a long
vowel accented on the first mora, as in θrími → θíim ‘shard’, the original *VrV sequence
where the second vowel is accented becomes a long vowel accented on the second mora,
as in xará → xaá ‘joy’. However, Heisenberg (1934: 91) notes that if r-deletion results
in a sequence of identical vowels with the stress on the second vowel, the stress shifts
from the second vowel to the first one, as in /karávi/ → [káav] ‘ship’. Newton (1972a: 79)
In Samothraki Greek, unstressed high vowels /i/ and /u/ delete unless the deletion creates phonotactically
unacceptable structures. Unstressed mid vowels /ɛ/ and /ɔ/ raise to [i] and [u] (Newton 1972a: 79).
Ryan (2014) presents statistical evidence from stress and meter showing that onsets are factors in syllable weight, though they are subordinate to the rhyme with respect to weight. For the discussion of the
possibility of moraic onsets, see Curtis (2003), Davis (1999), among others.
3 Compensatory lengthening and structure preservation revisited yet again
interprets the stress shift as evidence for vowel contraction (coalescence), while Heisenberg (1934: 90) and Margariti-Rogka & Tsolaki (2011) ascertain that the vowels remain
separate and belong to different syllables in such cases.
While the moraic weight approach does not seem to account for the Samothraki Greek
CL, Kavitskaya (2002) proposes a phonetic/historical account. According to Kavitskaya
(2002: 99), r is vocalic enough to be reinterpreted as additional vowel length. Kiparsky
(2011) argues that neither purely phonetic models nor purely phonological (weight conservation) models are sufficient to account for CL in Samothraki Greek. He develops
an account that relies on the observation that r is excluded from the onset position
cross-linguistically (Zec 2007). Typologically, high sonority segments are dispreferred
in the onset, which is evident from the fact that many languages, such as Korean, various Turkic languages, Basque, Piro, Telefol, etc., do not allow rhotics in word-initial or
syllable-initial positions (de Lacy 2001; Smith 2003) even though they have some type of
r in their consonant inventories. Languages employ different strategies to avoid onset
rhotics, such as prothesis, deletion, fortition, anti-gemination, and incorporation into
the nucleus (Kiparsky 2011: 26). Specifically, in Samothraki Greek the prohibition on
the rhotic in the onset is resolved through the latter strategy: the rhotic is syllabified
as a part of the nucleus so that the r and the following vowel form a rising diphthong,
and then deletes with CL. Katsika & Kavitskaya (2015) develop an articulatory phonetic
account of Samothraki Greek CL that builds both on Kavitskaya (2002) and Kiparsky
(2011). To resolve the dispreference for the onset rhotic, the tongue tip constriction of
the r is deleted, but the tongue body constriction is kept, preserving some of the segmental and temporal information of the r. The resulting segment is highly vocalic and
is subsequently incorporated into the nucleus. Thus, Katsika and Kavitskaya’s (2015)
account provides articulatory motivation to Kiparsky’s idea that in Samothraki Greek,
the onset r goes through a vocalic stage followed by the coalescence with the following
We can thus conclude that the best analysis of Samothraki Greek CL treats it as a
two-stage process, under which the vocalization of the onset r happens first, followed
by the coalescence of the two vocalic elements.10
3.4 Towards the explanation of CL as a sound change
From the point of view of contrast maintenance and loss, CL can be described as the
loss of contrast in a certain position. In the case of CL through the loss of consonants,
it is usually the coda consonant that deletes with the lengthening of the tautosyllabic
A reviewer points out that there are cases when the deletion of a coda consonant happens simultaneously
with intervocalic deletion of the same consonant, as, for example, in Turkish (de Chene & Anderson 1979,
Kavitskaya 2002: 23). The reviewer suggests that this renders such examples consistent with de Chene
and Anderson’s generalization. If the same was the case in Samothraki Greek, and the coalescence was
phonetically complete at the time of r-deletion, this, by itself, would be enough to exclude Samothraki
Greek from potential counterexamples to de Chene and Anderson’s generalization. I believe, however,
that CL through onset loss in Samothraki Greek is best re-analyzed as vowel coalescence, in the spirit of
Kiparsky (2011).
Darya Kavitskaya
vowel. In a system with no phonologically long vowels, the result of this process could
in principle be the introduction of a new vowel length contrast (the phonologization of
vowel length in a narrow sense). However, in the case of the pre-existing vowel length
distinction, the result is the introduction of the merger of the new long vowels with
existing long vowels (the phonologization of vowel length in a certain position, in a
broader sense of phonologization).
On the basis of the examples discussed above as well as other instances of CL, it an
be argued that CL as a sound change should indeed be defined as the lengthening of
the vowel after the loss of the tautosyllabic consonant as a result of the phonological
reanalysis of the additional vowel length, either in the spirit of de Chene & Anderson
(1979) or of Kavitskaya (2002). De Chene & Anderson (1979) dub this process “monophthongization”, that is, roughly, a vowel shift under which a monosyllabic vowel of two
vowel qualities becomes a monosyllabic vowel of one vowel quality (as exemplified by
Old French and Komi Ižma, among many others). From our admittedly incomplete survey of CL, some kind of a pre-existing length contrast is a necessary condition for CL
in the cases where such reanalysis is involved, and no clear cases that counterexemplify
this prediction have yet been found if this contrast is interpreted as including heterosyllabic and heteromorphemic sequences of identical vowels. Thus, CL is best described as
phonologization in a broader sense, that is, a merger of existing long vowels with new
long vowels that are the result of CL. It is possible that Samothraki Greek could also be
reanalyzed along these lines (as discussed in footnote 10), but, according to Kiparsky’s
(2011) account and the phonetic evidence amounted in Katsika & Kavitskaya (2015), it
stands out since the sound change goes through an intermediate stage, whereby the consonant becomes a full vocalic entity. CL in this case is a misnomer, and Samothraki Greek
is really an instance of vowel coalescence, which is a well-known and uncontroversial
source of vowel length in the languages of the world.
We can thus conclude that Samothraki Greek is not a counterexample to the generalization because the lost consonant vocalizes completely and then vowel coalescence
happens, with the result that is reminiscent of CL as it has the initial stage of consonant
plus vowel and a final stage of a long vowel, but is not, in fact, CL, but rather an instance
of VV > Vː coalescence. In turn, Old French is not a counterexample because, even if
Gess’s analysis of the Old French on the basis of the metrical scansion is correct, the
presence of a sequence of identical heterosyllabic vowels, which are likely to be phonetically identical to long vowels, provides sufficient contrast. Finally, Komi Ižma is not a
counterexample to the most restricted version of the generalization because the presence
of a sequence of identical heteromorphemic vowels is sufficient contrast.
4 Sound change, mergers, splits, and contrast
A broad question that remains to be discussed is the reason for why the CL sound change
that proceeds by gliding followed by monophthongization (de Chene & Anderson 1979)
tends to be structure-preserving, that is, is more likely to acquire vowel length in certain
positions with the loss of the consonant if vowel length is already contrastive elsewhere
in the system?
3 Compensatory lengthening and structure preservation revisited yet again
Two distinct proposals in the literature address the question of the relevance of structure preservation to sound change. One view on the structure-dependence of sound
change is expressed in Kiparsky (1995; 2003) and is to various extent present in other
work by Paul Kiparsky. Another view on structure-preserving sound change is presented in Blevins (2004a) and developed in Blevins (2009). As Anderson (2016) notes,
Blevins and Kiparsky advocate quite different views on the explanation of sound change.
While Blevins (2004a; 2006) puts the main burden of explanation of the sound change
on the phonetic factors, Kiparsky (2006) in a critique of Blevins’ program views individual grammars as a result of both “what change can produce and of what the theory
of grammar allows” (Anderson 2016: 17). Interestingly, both Blevins and Kiparsky see a
place for structure preservation in the theory of sound change, either as belonging to the
grammar (Kiparsky 1995; 2003) or emerging through acquisition (Blevins 2004a; 2009).
Kiparsky (2003: 328) comments on “the textbook story” of phonologization, where redundant features become phonemic with the loss of conditioning environment (e.g., in
the CL sound change, vowel length phonologizes with the loss of the tautosyllabic consonant). However, as Kiparsky (2003) points out, in many similar cases the redundant
features fail to phonologize and disappear with the loss of the conditioning environment. Kiparsky goes on to posit a priming effect, which is a diachronic manifestation of
structure preservation, formulated as in (6):
(6) Priming effect in phonologization (Kiparsky 2003: 328)
Redundant features are likely to be phonologized if the language’s phonological
representations have a class node to host them.
Kiparsky (2003) distinguishes between the two types of sound change, perceptionbased and articulation-based, and claims that while perception-based changes, such as
CL, metathesis, tonogenesis, and assimilation, are more likely to be structure-preserving
(phonologization in a broad sense, as defined in §3.4), articulation-based changes, such
as lenition, umlaut, etc., are usually structure-changing (phonologization in a narrow
sense, as defined in §3.4). Among the structure-changing processes that create long
vowels are vowel coalescence and also vowel lengthening in specific prosodic conditions
(for instance, under stress). Kiparsky (2003: 329) notes that Korhonen (1969: 333–335)
suggested that only certain allophones have a functional load that allows for the phonemicization with the loss of conditioning environment. Korhonen (1969) dubs these allophones quasi-phonemes. Having claimed that the classical phoneme is “something of a
straightjacket” when it comes to understanding of the introduction and loss of phonological contrast, Kiparsky (2013) proposes a system, where he distinguishes between
contrastiveness, as a structural notion, and distinctiveness, as a perceptual notion, as
shown in Table 7.
By the system in Table 7, quasi-phonemes are not contrastive, but distinctive, and
thus they represent a necessary stage to the secondary split. Since distinctiveness is a
perceptually-defined notion, only those sound changes that are perceptually-based are
predicted to follow this pattern. As was discussed in §3.3, vowel length in Komi Ižma is
quasi-phonemic and thus is a likely candidate for the phonologization of vowel length
in the language.
Darya Kavitskaya
Table 7: Contrastiveness vs. distinctiveness (Kiparsky 2013).
Near contrasts
Blevins (2004a; 2009) pursues a research agenda that is very different from Kiparsky’s
theory of sound change. However, she also notes that certain sound changes tend to
be structure-preserving, and that these changes tend to be perceptually-based. Blevins
(2004a) posits a principle of structural analogy, stated in (7):
Structural Analogy (Blevins 2004a: 154)
In the course of language acquisition, the existence of a (non-ambiguous)
phonological contrast between A and B will result in more instances of sound
change involving shifts of ambiguous elements to A or B than if no contrast
between A and B existed.
The consequence of such principle for sound change is a tendency towards structure
preservation. Blevins (2009) presents an overview of two known cases of sound changes
that have this tendency, such as CL (de Chene & Anderson 1979; Kavitskaya 2002) and
metathesis (Blevins & Garrett 1998; Blevins 2004a,b; Hume 2004) and then proceeds to
a case study of the Principle of Structural Analogy, unstressed vowel syncope in Austronesian.
According to Blevins (2009), unstressed vowel syncope in the Austronesian languages
discussed is a perceptually-based sound change that is the result of the ambiguous vocalic
status of hypoarticulated short unstressed vowel. The loss of the second vowel in a
CV.CV.CV sequence creates a structure CVC.CV where the first syllable is closed. The
Principle of Structural Analogy predicts that languages that contrast open and closed
syllables will have a stronger tendency towards this kind of syncope. Indeed, as Blevins
(2009) shows, the prediction is borne out.
Blevins’ (2009) example is different from the case of CL in an interesting and fundamental way. While CL as a sound change amounts to the introduction of a new allophone
and potentially a new phoneme with a positional loss of a segment, unstressed vowel
syncope is the introduction of a new prosodic structure with a positional loss of a segment.11 This example provides additional support to the generalization that the presence
of contrast in the system affects sound change that potentially creates similar structures.
As a reviewer notes, length is prosodic structure as well, and in this sense, there is little difference between
the case discussed by Blevins and the cases of CL. However, while CL (potentially) introduces a new element to the inventory of phonemes, Blevins discusses an example that introduces a new structure to the
inventory of syllables.
3 Compensatory lengthening and structure preservation revisited yet again
5 Conclusions
De Chene and Anderson (1979) had an important insight about the structure-preserving
nature of CL that holds in the majority of the languages with this sound change and
thus cannot be ignored. I have presented examples in which systemic considerations
play an important role in the phonologization of newly introduced phonetic detail in
perception-based sound changes, such as vowel duration in CL. I have shown a way
to address potential counterexamples to the generalization, reanalyzing CL in Samothraki Greek as vowel coalescence and arguing that in the cases of Old French and Komi
Ižma the presence of identical tautosyllabic vowels elsewhere in the system might have
constituted a sufficient contrast for the phonologization of vowel length through CL.
Abondolo, Daniel (ed.). 1998. The Uralic languages. London/New York: Routledge.
Anderson, Stephen R. 1982. Where’s morphology? Linguistic Inquiry 13(4). 571–612.
Anderson, Stephen R. 2016. Synchronic versus diachronic explanation and the nature of
the language faculty. Annual Review of Linguistics 2. 11–31.
Blevins, Juliette. 2004a. Evolutionary phonology: The emergence of sound patterns. Cambridge: Cambridge University Press.
Blevins, Juliette. 2004b. Phonetic explanations for recurrent sound patterns: D iachronic
or synchronic? In Eric Raimy & Charles Cairns (eds.), Phonological theory: Representations and architecture. Paper presented at CUNY Symposium on Phonology Theory.
Cambridge, MA: MIT Press.
Blevins, Juliette. 2006. New perspectives on English sound patterns: ‘Natural’ and ‘unnatural’ in Evolutionary Phonology. Journal of English Linguistics 34. 6–25.
Blevins, Juliette. 2009. Structure-preserving sound change: A look at unstressed vowel
syncope in Austronesian. In Andrew Pawley & Alexander Adelaar (eds.), Austronesian
historical linguistics and culture history: A festschrift for Bob Blust, 33–49. Canberra:
Pacific Linguistics.
Blevins, Juliette & Andrew Garrett. 1998. The origins of consonant-vowel metathesis.
Language 74(3). 508–556.
Collinder, Björn. 1957. Survey of the Uralic languages. Stockholm: Almqvist & Wiksell.
Curtis, Emily. 2003. Geminate weight: Case studies and formal models. University of Washington Ph.D. Dissertation.
Davis, Stuart. 1999. On the representation of initial geminates. Phonology 16. 93–104.
de Chene, Brent E. 1985. The historical phonology of vowel length. New York: Garland.
de Chene, Brent E. & Stephen R. Anderson. 1979. Compensatory lengthening. Language
55(3). 505–535.
de Lacy, Paul. 2001. Markedness in prominent positions. In Ora Matushansky, A. Costa,
Javier Martin-Gonzalez, Adam Szczegielniak & Lance Nathan (eds.), Proceedings of
HUMIT 2000, MIT Working Papers in Linguistics 40, 53–66. Cambridge, MA: MITWPL.
Darya Kavitskaya
Dixon, Robert M. W. 1990. Compensating phonological change: An example from the
northern dialects of Dyirbal. Lingua 80. 1–34.
Gess, Randall. 1998. Compensatory lengthening and structure preservation revisited.
Phonology 15. 353–366.
Harms, Robert T. 1967. Split, shift, and merger in the Permic vowels. Ural-Altaische
Jahrbücher 39. 163–198.
Harms, Robert T. 1968. Introduction to phonological theory. Englewood Cliffs, NJ: PrenticeHall.
Hausenberg, Anu-Reet. 1998. Komi. In Daniel Abondolo (ed.), The Uralic languages, 305–
326. London/New York: Routledge.
Hayes, Bruce. 1989. Compensatory lengthening in moraic phonology. Linguistic Inquiry
20(2). 253–306.
Heisenberg, August. 1934. Neugriechische Dialekttexte. Leipzig, Germany: Harrassowitz.
Hock, Hans Henrich. 1986. Compensatory lengthening: In defense of the concept ‘mora’.
Folia Linguistica 20. 431–460.
Hume, Elizabeth. 2004. The indeterminacy/attestation model of metathesis. Language
80(2). 203–237.
Katsanis, Nicolaos. 1996. Το γλωσσικό ιδίωμα της σαμοθράκης. Thessaloniki, Greece: Aristotle University of Thessaloniki Press.
Katsika, Argyro & Darya Kavitskaya. 2015. The phonetics of /r/ deletion in Samothraki
Greek. Journal of Greek Linguistics 15. 34–65.
Kavitskaya, Darya. 2002. Compensatory lengthening: Phonetics, phonology, diachrony.
New York: Routledge.
Kiparsky, Paul. 1995. The phonological basis of sound change. In John Goldsmith (ed.),
The handbook of phonological theory, 640–670. Oxford: Blackwell Publishing.
Kiparsky, Paul. 2003. The phonological basis of sound change. In Brian D. Joseph &
Richard D. Janda (eds.), The handbook of historical linguistics, 313–342. Oxford: Blackwell Publishing.
Kiparsky, Paul. 2006. The Amphichronic Program vs. Evolutionary Phonology. Theoretical Linguistics 32(2). 217–236.
Kiparsky, Paul. 2011. Compensatory lengthening. In Eric Raimy & Charles Cairns (eds.),
Handbook of the syllable, 33–69. Leiden: Brill.
Kiparsky, Paul. 2013. Phonologization. In Patrick Honeybone & Joseph Salmons (eds.),
Handbook of historical phonology. Oxford: Oxford University Press.
Korhonen, Mikko. 1969. Die Entwicklung der morphologischen Methode im Lappischen.
Finnisch-Urgische Forschungen 37. 203–262.
Lytkin, V. I. (ed.). 1966. The contemporary Komi language. Syktyvkar.
Lytkin, V. I. & T. I. Teplyashina. 1976. Permskie iazyki. Osnovy finno-ugorskogo iazykoznaniia. Moskva: Nauka.
Margariti-Rogka, Marianna & Maria Tsolaki. 2011. Τα μακρά φωνήεντα στο ιδίωμα της
σαμοθράκης. In C. Mpasea-Mpezantakou, I. Manolessou, A. Afroudakis, S. Mpeis &
G. Katsouda (eds.), Νεοελληνική διαλεκτολογία 6, 273–294. Athens, Greece: Academy
of Athens.
3 Compensatory lengthening and structure preservation revisited yet again
Morin, Yves-Charles. 1992. Phonological interpretations of historical lengthening. Paper
presented at the 7th International Phonology Meeting. Krems.
Newton, Brian. 1972a. Generative interpretation of the dialect. Cambridge: Cambridge University Press.
Newton, Brian. 1972b. Loss of /r/ in a Modern Greek dialect. Language 48. 566–572.
Pope, Mildred K. 1934. From Latin to Modern French with especial consideration of AngloNorman: Phonology and morphology. Manchester: Manchester University Press.
Roach, William (ed.). 1959. Chrétien de Troyes – Le Roman de Perceval ou Le Conte du
Graal. Geneva: Droz.
Ryan, Kevin. 2014. Onsets contribute to syllable weight: Statistical evidence from stress
and meter. Language 90. 309–341.
Smith, Jennifer. 2003. Onset sonority constraints and subsyllabic structure. Rutgers Optimality Archive 608.
Sumner, Megan. 1999. Compensatory lengthening as coalescence: Analysis and implications. In Sonya Bird, Andrew Carnie, Jason D. Haugen & Peter Norquest (eds.), WCCFL
18 proceedings, 532–544. Somerville, MA: Cascadilla Press.
Topintzi, Nina. 2006. A (not so) paradoxical instance of compensatory lengthening: Samothraki Greek and theoretical implications. Journal of Greek Linguistics 7. 71–119.
Zec, Draga. 2007. The syllable. In Paul de Lacy (ed.), The Cambridge handbook of phonology, 161–194. Cambridge: Cambridge University Press.
Chapter 4
Phonological exceptionality is localized
to phonological elements: the argument
from learnability and Yidiny word-final
Erich R. Round
University of Queensland
Anderson (2008) emphasizes that the space of possible grammars must be constrained by
limits not only on what is cognitively representable, but on what is learnable. Focusing on
word final deletion in Yidiny (Dixon 1977a), I show that the learning of exceptional phonological patterns is improved if we assume that Prince & Tesar’s (2004) Biased Constraint
Demotion (BCD) with Constraint Cloning (Pater 2009) is subject to a Morphological Coherence Principle (MCP), which operationalizes morphological analytic bias (Moreton 2008)
during phonological learning. The existence of the MCP allows the initial state of Con to be
simplified, and thus shifts explanatory weight away from the representation of the grammar
per se, and towards the learning device.
I then argue that the theory of exceptionality must be phonological and diacritic. Specifically,
I show that co-indexation between lexical forms and lexically indexed constraints must be
via indices not on morphs but on individual phonological elements. Relative to indices on
phonological elements, indices on morphs add computational cost for no benefit during
constraint evaluation and learning; and a theory without indices on phonological elements
is empirically insufficient. On the other hand, approaches which represent exceptionality
by purely phonological means (e.g. Zoll 1996) are ill-suited to efficient learning. Concerns
that a phonologically-indexed analysis would overgenerate (Gouskova 2012) are unfounded
under realistic assumptions about the learner.
1 Exceptionality
What is the nature of representations which are passed from the morphology to the
phonology? Anderson (1992) demonstrates that the processes that create those representations can be elaborate and complex. Operations that act upon morphological forms, to
realize units of morphologically-relevant meaning, involve not only the concatenation
Erich R. Round. 2017. Phonological exceptionality is localized to phonological elements: the argument from learnability and Yidiny word-final deletion. In Claire Bowern, Laurence Horn & Raffaella Zanuttini (eds.), On looking into words (and beyond),
59–98. Berlin: Language Science Press. DOI:10.5281/zenodo.495439
Erich R. Round
of formatives, but also selection among alternatives and non-concatenative modifications to intermediate representations (see also Anderson 2015; 2016; 2017). However,
what of the final result, which comprises some number of morphs that must then be
interpreted phonologically? A constant concern of generative phonology since its inception has been to account adequately for patterned phonological exceptionality, the
phenomenon in which segments in a restricted class of morphs exhibit phonologically
distinctive behavior as triggers, targets or blockers of alternations, or as participants in
exceptional featural, phonotactic or prosodic surface structures. For example, in Yidiny
(Dixon 1977a,b) vowels delete word-finally, if that deletion would prevent the word from
surfacing with an unfooted syllable. This is seen in the root gaɟara- ‘possum’ in (1a) and
the suffix -ɲa accusative in (1b), where feet are marked by parentheses. However, in a
restricted set of morphs the final vowel behaves exceptionally, resisting deletion, as in
the root guɟara- ‘broom’ (2a) and the suffix -na purposive (2b).
(ga ɟa:r)
(bim bi:ɲ)
(gu ɟa:) ra
(ga li:) na
In order for the phonology to treat morph-specific, exceptional segments appropriately,
it must receive from the morphology some kind of discriminating information which it
can act upon. For much of the generative period it has been argued that this information is associated with morphs as a whole, and not with their individual phonological
elements. Here I present an argument for the contrary view. The contribution, then, is to
clarify the nature of one important aspect of the interaction between the morphological
and phonological components of grammar. The principle line of evidence is learnability,
namely the learnability of an optimality-theoretic grammar for phonological exceptionality. Anderson (2008) has emphasized that the space of possible human grammars must
be constrained not only by limits on what is cognitively representable, but also on what
is learnable. The crux of the argument here relies not on specifics, but ultimately on general properties of learnable grammars, and thus I would hope should remain valid even
as specific theories undergo refinement as they move closer to answering Anderson’s
(2008) challenge.1
The chapter falls into two broad parts. In §2–§5 I discuss the processes and principles
required to learn exceptionality. This leads to the positing of a Morphological Coherence
A reviewer asks whether the machinery presented here is necessary if one assumes an exemplar-based
model of phonology. I assume that learners do store rich, exemplar-like representations of linguistic experiences. However, natural language morphology in general has enough combinatorial complexity that
reliance upon retrieved episodes will not be sufficient to reproduce the full range of creative behavior that
humans display. Consequently some generative machinery is necessary, which performs not merely simple analogies and concatenations, but which can reproduce with precision the complex patterns generated
by a realizational morphology such as Anderson’s (1992), and by a formal phonological grammar such as
entertained here.
4 Phonological exceptionality is localized to phonological elements
Principle in §6, which operationalizes a morphological bias that ensures successful learning for certain cases. In §7–§9 I am concerned with the underlying theory of these processes and principles. I evaluate two broad approaches to phonological exceptionality:
phonological approaches, which represent exceptionality as a property of individual
segments (Bloomfield 1939; Kiparsky 1982a; Inkelas 1994; Zoll 1996), and morphological approaches which represent it as a property of morphs (Chomsky 1964; Chomsky &
Halle 1968; Zonneveld 1978; Pater 2000). The result is an argument in favor of a diacritic
phonological approach. On this account, exceptionality is represented at the level of
individual phonological elements, not morphs; however the means of marking it is by
diacritics which are visible to the phonology but not manipulable by it, in contradistinction to the concrete phonological approach, where the crucial representations are
themselves phonological elements. As I show, the function of these “Φ-indices” is essentially identical to “M-indices” which would mark morphs, only there is no assumption
that all exponents of a morph m be indexed identically. As we shall see, freedom from
that assumption is both coherent theoretically and desirable, computationally and empirically. The discussion is illustrated throughout by the facts of word final deletion in
Yidiny, to which we turn now in §2.
2 Word-final deletion in Yidiny
2.1 The phenomenon
Yidiny (Dixon 1977a) belongs to the Yidinyic subgroup of the Pama-Nyungan language
family. Traditionally it was spoken in the rainforest region southwest of Cairns, in Northeastern Australia. Most examples below are from Dixon’s (1977) detailed descriptive
grammar; examples marked † are from Dixon’s (1991) dictionary and texts. An inventory
of underlying segments is in Table 1.
Table 1: Yidiny underlying segments, after Dixon (1977a: 32).
Lateral, trill
l, r
i, a, u, i:, a:, u:
Syllable shapes are tightly constrained. Onsets are obligatory and simple. Codas
permit only sonorants other than /w/. Codas in word-final position are simple; wordinternal codas also permit disegmental, continuant–nasal sequences. Morphologically,
the language is almost entirely suffixing and largely agglutinative. Roots are minimally
Erich R. Round
disyllabic and suffixes are maximally disyllabic (Dixon 1977a: 35,90). An online appendix2 discusses the morphological constituency of verbal inflection.
Of Yidiny’s phonological alternations, those to receive the greatest attention have
been stress placement, vowel length and to a lesser extent, word-final deletion (Dixon
1977a,b; Hayes 1982; 1985; Kager 1993; Crowhurst & Hewitt 1995; Halle & Idsardi 1995;
Hall 2001; Pruitt 2010; Hyde 2012; Bowern, Round & Alpher in revision, inter alia).
Yidiny’s stress and length alternations in particular have featured in significant theoretical works on meter and prosody over the past four decades, and both are nontrivial
topics in themselves. Word-final deletion, however, can be studied largely independently
of them for reasons that follow.
Although stress placement in Yidiny has proven contentious (Pruitt 2010; Bowern,
Round & Alpher in revision), word-final deletion is not sensitive to stress per se, but
rather only to the position of foot boundaries. These have been uncontroversial since
their analysis by Hayes (1982): feet in Yidiny are disyllabic and left-aligned within the
phonological word.
Many words with word-final deletion also exhibit vowel lengthening; however the
phenomena show little to no mutual interaction. In a rule-based theory permitting simultaneous application (Anderson 1974) lengthening and deletion would apply simultaneously; neither rule feeds or bleeds the other.3 See Round (in progress) for an analysis
of Yidiny lengthening.
Word-final deletion is sensitive to foot placement, and foot placement is sensitive
to phonological word boundaries. In Yidiny, phonological words commence at the left
edge of each root and each disyllabic suffix (Dixon 1977a: 88–98).4 Phonological words
therefore begin with either a polysyllabic root or a disyllabic suffix and are followed by
zero or more monosyllabic or entirely consonantal suffixes. Word-final deletion targets
unfooted syllables and therefore only affects prosodic words which, modulo deletion,
would be at least trisyllabic. As a consequence, we are interested here in three kinds
of phonological word: those comprised of bare roots of three or more syllables; those
comprised of roots plus one or more monosyllabic suffixes; and those comprised of a
disyllabic suffix plus one or more additional, monosyllabic suffixes. The third kind is
rare,5 and so discussion will focus on the first two.
Word-final deletion applies only if the word thereby avoids surfacing with an unfooted
syllable. For example, the roots gindanu- ‘moon’ and gubuma- ‘black pine’ both contain
three vowels, each of which is a potential syllabic nucleus at the surface. In (3a,4a) they
have undergone deletion of their final vowel to prevent it from surfacing in an unfooted
syllable; compare (3b,4b) where the roots are non-final in the word, and the final vowels
Available from 10.6084/m9.figshare.4579696
Deletion counter-bleeds lengthening, thus in a strictly serial analysis lengthening would precede wordfinal deletion (Dixon 1977a,b; Hayes 1985; Crowhurst & Hewitt 1995).
Yidiny’s only prefix, [ɟa:-] ‘in a direction’ occupies its own phonological word (Dixon 1977a: 98,162).
For an illustration, see example (25).
4 Phonological exceptionality is localized to phonological elements
(gin da:n)
*(gin da:) nu
→ (gin da) (nuŋ gu)
(4) a.
‘black pine[abs]’
‘black pine-purp’
(gu bu:m)
*(gu bu:) ma
→ (gu bu) (ma gu)
Final vowel deletion may also affect suffixes. In (5a,c,6a), the vowels of the nominal
comitative suffix -yi and verbal comitative suffix -ŋa have undergone deletion, thereby
preventing the surfacing of an unfooted syllable. In (5b,6b) the suffixes are non-final in
the word, and the vowel surfaces.
(5) a.
(6) a.
‘black bream-com’
(bu ɲa:y)
*(bu ɲa:) yi
→ (bu ɲa) (yiŋ gu)
→ (gu lu) (gu lu:y)
*(gu lu) (gu lu:) yi
(ga da:ŋ)
* (ga da:) ŋa
→ (ga da:) (ŋal ɲu)
Word-final deletion interacts with restrictions on word-final consonants, and the interaction plays out differently in roots versus suffixes. In roots, deletion will fail to apply
if the result would be an illicit word-final coda, containing either a stop or /w/ (7) or a
cluster (8). One conceivable alternative, to also delete the consonant, is not attested in
roots (7–8).6
‘sugar ant[abs]’
‘place name[abs]’
(wa gu:) ɟa
* (wa guɟ)
* (wa gu:)
→ (gu da:) ga
* (gu da:g)
* (gu da:)
→ (ba la:) wa
* (ba la:w)
* (ba la:)
→ (ŋa lu:m) ba
* (ŋa lu:mb)
* (ŋa lu:m)
Neither Dixon’s grammar (1977) nor dictionary (Dixon 1991, which cites underlying forms) records a surface
form for the roots in (7c) and (7d), or for roots illustrating the same pre-final consonant or comparable
consonant clusters. However, Dixon (1977a: 57–58) specifically reports that the roots balawa- and gindalbado not undergo deletion; the surface forms provided here are what we would expect if this is so.
Erich R. Round
(bi na:r) ŋa
* (bi na:rŋ)
* (bi na:r)
In contrast, deletion in suffixes applies not only to the final vowel, but also to a single consonant that precedes it, if that consonant would be illicit word-finally, as in (9).
This form of CV deletion respects phonotactic constraints while also avoiding unfooted
‘grey possum-erg’
(mar gu:ŋ)
(wa wa:l)
(bi nar) (ŋal ɲu:n)
However, word-final deletion never deletes the initial segment of a suffix (and consequently, it will never delete an entire suffix), as illustrated in (10).
(bu ɲa:) ba
* (buɲ ba)
→ (guy ga:l) ni
* (guy ga:ln)
* (guy ga:l)
Deletions do not occur word internally (11a,b), nor do word-final, licit codas delete (11b).
All Yidiny roots and suffixes that are consonant-final end underlyingly with licit coda
consonants, so no morph undergoes spontaneous deletion of an underlyingly-final consonant (11c).
(bu ɲa:b)
* (bu ɲa:)
→ (ba gi:) ram
* (ba gi:rm)
* (ba gi:r)
→ * (ba gi:r)
To summarize, word-final deletion applies only so as to avoid the surfacing of unfooted
syllables. It may delete the final vowel from a root and the final (C)V sequence from
a suffix, but will not delete a suffix-initial segment. Deletion is blocked (in roots) or
expanded (in suffixes, from V deletion to CV deletion) in order to obey phonotactic restrictions on word-final codas. These are the regular conditions under which word-final
deletion occurs.
In addition to its regular application, Yidiny contains roots and suffixes which are
exceptional non-undergoers of word-final deletion. In (13), the non-undergoer roots
The “dative subordinate” is marked by what Round (2013: 26) has called “compound suffixation”, comprising
two monosyllabic suffixal morphs, /-lɲu; -nda/. That these are not a single, disyllabic suffix is evident in
the fact that they fail to be parsed into a their own phonological word, separate from the root.
4 Phonological exceptionality is localized to phonological elements
mulari-, guɟara-, ɟudulu-, baŋgamu- all resist word-final deletion despite their pre-final
consonant being permissible as a coda, and despite the fact that the consequence is an
unfooted, word-final syllable.
‘initiated man[abs]’
‘brown pigeon[abs]’
(mu la:) ri
*(mu la:r)
→ (gu ɟa:) ra
*(gu ɟa:r)
→ (ɟu du:) lu
*(ɟu du:l)
→ (baŋ ga:) mu
*(baŋ ga:m)
Dixon (1977a: 59) reports 115 trisyllabic roots whose phonotactic shape would, under
regular conditions, expose them to word-final deletion. Of these, 34, or around 30%,
are exceptional non-undergoers. The distinction is idiosyncratic; neither Dixon (1977a:
58) nor subsequent researchers have found any phonological, semantic or grammatical
factor that categorically determines whether a root will be a non-undergoer.8
Suffixes also may be exceptional non-undergoers. In (17) the non-undergoer suffixes
-nda, -lɟi and -na resist word-final deletion and allow an unfooted syllable to surface.
Avoidance of regular, word-final CV deletion is seen in (13a,b) and V deletion in (13c).
(13) a.
‘grey possum-dat’
(mar gu:n) da
*(mar gu:n)
→ (wa wa:l) ɟi
*(wa wa:l)
→ (ga li:) na
*(ga li:n)
Tables 2 and 3 list all suffixal allomorphs in Yidiny which, on phonotactic grounds, could
plausibly delete.9 Regular undergoers are in Table 2 and non-undergoers in Table 3.
Historically speaking, borrowed forms may account for many of these items (Barry Alpher p.c.); synchronically, however, their motivation is opaque.
9 Such suffixes must be vowel-final and monosyllabic. If just the final vowel is to delete, then it must leave
behind a single, licit-coda consonant in word final position. This will require the suffix to be -CV, and be
preceded by a vowel, not a consonant. Alternatively, if the final CV is to delete, then the suffix must be
-CCV, since suffix-initial segments do not delete, and it too must attach to a vowel-final stem. Data here is
from a comprehensive search of Dixon (1977a), in which relevant information can be found on pp.50–54,
151. “Emphatic” -ɲa (Dixon 1977a: 151) is excluded. It behaves as a phonological clitic that occupies a distinct
phonological word, and does not undergo final deletion.
Erich R. Round
Table 2: Monosyllabic suffixes which undergo word-final deletion.
-ni, -nu
past tense inflection
comitative derivation
dative subordinate inflection
-lɲu, -ɻɲu
Table 3: Monosyllabic suffixes which escape word-final deletion.
purposive inflection
lest nominalizing derivation
-lna, -ɻna
-nɟi, -lɟi, -ɻɟi
Exceptional non-undergoers, both roots and suffixes, only block the deletion of their
own segments; the exceptionality does not spread to neighboring morphs. Accordingly
in (14), the exceptional non-undergoer lest does not block deletion in the following,
regular undergoer, ergative suffix.
(wi wi:) (ɟin ɟi:ŋ)
*(wi wi:) (ɟin ɟi:ŋ) gu
Likewise, the presence of a regular undergoer will not undo the blocking effect of an exceptional non-undergoer. In (15) the regular undergoer comitative does not undermine
the blocking of deletion in the exceptional non-undergoer purposive, which follows it.
‘walk up-com-purp’
(ma ɟin) (da ŋa:l) na
*(ma ɟin) (da ŋa:l)
It will be recalled that roots in Yidiny can undergo word-final deletion of vowels, but
not of the consonants that precede them. More specifically, roots that end in CCV do
not delete final CV, whereas some suffixes do, and nor does final CʹV delete from roots
that end in VCʹV, where Cʹ is an impermissible coda. Two conceivable accounts for this
The dative subordinate is marked by a string of two monosyllabic suffixes -lɲu-nda, cf. fn.7.
4 Phonological exceptionality is localized to phonological elements
may be distinguished. On one account, the grammar of Yidiny expressly prohibits rootfinal CV deletion. On the other, it happens just by chance that all CCV-final and VCʹVfinal roots are exceptional non-undergoers. On the latter account, the grammar would
enforce CV deletion from roots, if only the lexicon provided the right inputs; on the
former account it would not. The level of empirical support for these hypotheses can be
assessed statistically. Table 4 compares counts of CCV- and VCʹV-final roots and CCVfinal suffixes which either do or do not delete. The distribution is strongly unbalanced,
and we can reject with confidence the null hypothesis that it is due to chance (χ 2 df=1
= 47.9 p < 10-10 ). Table 5 compares counts of roots that are CCV- and VCʹV-final with
those that are VCV-final, i.e., where C is a permissible coda. Again, the counts are highly
unbalanced and we reject the hypothesis that the absence of deletion in CCV- and VCʹVfinals is by chance (χ 2 df=1 = 125.8. p < 10-10 ). The only empirically-supported conclusion
is that the lack of consonant deletion in Yidiny roots is systematic, not due to chance. A
satisfactory formal analysis should reflect this.11
Table 4: Deletion of coda-ilicit pre-final C in roots versus suffixes.
No deletion
CCV- and VCʹV-final roots
CCV-final suffixes
Table 5: Deletion in roots with pre-final coda-illicit C versus prefixal coda-licit
No deletion
CCV- and VCʹV-final roots
VCV-final suffixes
2.2 Constraint rankings
A briefly sketch now follows of how the facts above would be analysed in OT. Foot placement in Yidiny is due to FootBinarity ≫ ParseSyllable ≫
Align(Ft,L,PrWd,L) (Prince & Smolensky 2004[1993], McCarthy & Prince 1993a, McCarthy & Prince 1995). Of these, only ParseSyllable (Prs) will be of interest for our
purposes; I assume that other prosodic constraints are satisfied optimally. Absolute re11
As a reviewer observes, there is an interesting historical background to be clarified here, and an account
of it is planned. Naturally, the object of a synchronic analysis differs ontologically from that of a historical
one. The two are complementary, but neither account would substitute for or serve as a counter-analysis
to the other.
Erich R. Round
strictions against obstruents and /w/ in codas are due to Sonorant/Coda (e.g. Lombardi
2002) and *w/Coda; I assume these are unviolated.
Regular word-final deletion in Yidiny can be analysed straightforwardly by ranking
Prs ≫ Maximality (Max, McCarthy & Prince 1995). This causes deletion of final vowels
in preference to the surfacing of unfooted syllables, but not if an illicit coda results.
Segments may delete from the right edge of the word only, not the left or wordinternally. High-ranking Anchor-Left(morph) penalizes deletion from the left edge
of any morph and Contig-IO(PrWd) penalizes deletion internally (McCarthy & Prince
Yidiny permits complex codas word-internally, but not word-finally. Ranking Cntg
≫ *ComplexCoda (Bernhardt & Stemberger 1998) accounts for this; ranking both above
Prs accounts for the absence of deletion after pre-final clusters in roots and the defeat
of candidates which delete only a final vowel from word-final CCV suffixes.
Word-final deletion applies differently to roots and suffixes. Roots will not undergo
consonant deletion, even if the consequence is an unfooted syllable. The ranking of undominated Max-C/root (McCarthy & Prince 1995) above Prs accounts for this. Suffixes
do not violate Max-C/rt and consequently are free to undergo consonant deletion, however highly-ranked Anc penalizes the deletion of morph-initial segments. This accounts
for the fact that a consonant may delete from a -CCV suffix but not from -CV.
At this point, regular word-final deletion occurs whenever satisfaction of the markedness constraint Prs requires the violation of the lower-ranked faithfulness constraint
Max. Deletion is blocked unexceptionally whenever Prs itself is violated in order to satisfy higher-ranking constraints, which are of two kinds: those which penalize marked
codas, son/Coda, *w/Coda, *Cplx; and those which penalize deletion in specific morphological contexts, namely at left edges of morphs, Anc, and consonants in roots, MaxC/rt. We see that the driver of word-final deletion in Yidiny is the ranking of Prs ≫
Max. Deletion occurs when Prs is satisfied but Max is not. Regular blocking results
when Prs must be violated, in which case Max can be satisfied.
Exceptional non-undergoers avoid deletion. For them, Max is always satisfied, even
at the expense of Prs. Consequently, while regular undergoers are subject to a ranking
of Prs ≫ Max, exceptional non-undergoers must be subject to Max ≫ Prs. In §4 I
consider two approaches that will ensure this is the case, one morphological and one
phonological. First though, a remark about constraint violations.
3 Relativized constraint violation
I introduce here a simple expression for relating the violations of certain pairs of constraints, which will aid discussion in later sections.
For any constraint C and candidate cand, there will be zero or more violations of C.
Given the definition of C, those violations will be due to certain parts, or loci, in cand, either in the output of cand or in the correspondences between input and output elements
(McCarthy & Prince 1995). We can define the set of loci of violation, V(C, cand), as
the loci in cand which cause violations of C (McCarthy 2003, Łubowicz 2005). Now,
4 Phonological exceptionality is localized to phonological elements
some pairs of constraints C1 , C2 are related such that for any cand, the loci of violation
of C2 are a subset of the loci of violation of C1 . In many cases, the latter are precisely
those members of the former which also contain some particular kind of phonological
element. For example V(Max-C, cand) are those members of V(Max, cand) which also
contain input consonants. In that case, we can express V(C2 , cand) terms of the intersection of the set V(C1 , cand) and some appropriately defined second set, that picks out
loci containing the criterial elements. Let us define the set of “ϕ-loci”, Lϕ (D(ϕ), cand),
as the set of loci in cand that contain a phonological element ϕ of the kind denoted by
predicate D(ϕ). For example, V(Max-C, cand) can be defined in relative terms, as in (16),
where the predicate input_consonant(ϕ) denotes input consonants. (For brevity I omit
the “cand” from the expression for each set.)
(16) V(Max-C) =def V(Max) ∩ Lϕ (input_consonant(ϕ))
This relativized method will be used below to define new constraints CN in terms of a
reference constraint, CR , and a set of phonological elements which restrict the violations
of CN relative to those of CR .
4 Preliminary analysis of word-final deletion
4.1 A morphological approach
We now consider an OT implementation of the morphological approach to Yidiny exceptionality, using lexically indexed constraints (Pater 2000; 2006; 2009). A lexically
indexed constraint CM behaves precisely like its unindexed counterpart, C, except that it
can be violated only by structures which contain exponents of a specific set M of morphs,
each of which has been assigned a diacritic mark which I will term a lexical M-index,
that co-indexes it to CM . The definition can be expressed relatively as in (17), following
a similar formulation by Finley (2010).
(17) V(CM ) =def V(C) ∩ Lϕ (m∈M & Exp(ϕ, m)), where:
M is the set of morphs co-indexed to CM .
Exp(ϕ, m) states that element ϕ is an exponent of morph m
If we now define two sets of Yidiny morphs, U the set of regular undergoers of wordfinal deletion, and N the set of exceptional non-undergoers, then either of the rankings
in (18) will ensure that the correct sets of morphs is subject to the desired partial ranking
of Prs and Max.
a. PrsU ≫ Max ≫ Prs
b. MaxN ≫ Prs ≫ Max
In (18a), all phonological exponents of undergoer morphs will be subject to PrsU ≫
Max, and non-undergoers to Max ≫ Prs. In (18b), all phonological exponents of nonundergoer morphs will be subject to MaxN ≫ Prs, and undergoers to Prs ≫ Max. For
Erich R. Round
now I will use ranking (18a); the reason for this will become clear in §5.12,13 Examples in
(19a–19b) illustrate word-final deletion of regular undergoers which are indexed U, the
root malanu-U and suffix ergative -ŋguU , while (19c–19d) show the absence of deletion
for exceptional non-undergoers mulari- ‘initiated man’ and dative nda.
/malanuU / ‘right hand[abs]’
(ma la:n) ≻ (ma la:) nu
/margu-ŋguU / ‘grey possum-erg’
(mar gu:ŋ) ≻ (mar gu:ŋ) gu
/mulari/ ‘initiated man[abs]’
(mu la:) ri ≻ (mu la:r)
/margu-nda/ ‘grey possum-dat’
(mar gu:n) da ≻ (mar gu:n)
/maɟinda-ŋaU -lna/ ‘walk up-com-purp’
(ma ɟin) (da ŋa:l) na ≻ (ma ɟin) (da ŋa:l)
Example (19e) illustrates the fact that violations of PrsU require not merely the presence of a U -indexed morph in the word, but a locus of violation which contains a phonological exponent of a U -indexed morph (17). Namely, the final syllable of (19e), na, is unfooted. However since that syllable contains no phonological exponent of a U -indexed
morph, no violation of PrsU results. This is true despite the presence of a U -indexed
morph elsewhere in the word.
4.2 A phonological approach
The phonological approach correlates the (un)exceptionality of a segment with representational properties of the segment itself. Implementations differ as to which property is
used. Zoll (1996) analyses segments which resist deletion as having root nodes in their
input, whereas segments that delete more readily lack root nodes, and are termed subsegments. Under these assumptions, a ranking Max(Seg) ≫ Prs ≫ Max(Subseg) ensures that segments with input root nodes are subjected to Max(Seg) ≫ Prs, while those
without are subjected to Prs ≫ Max(Subseg).14 Examples are in (20), where segments
without root nodes are underlined.
Briefly, procedures for learning OT grammars improve in performance if they opt to rank markedness
higher than faithfulness when given a choice. Consequently the ranking in (18a) will be learned in preference to (18b); see §5.
An early proposal that only faithfulness constraints be indexable (Benua 1997; Itô & Mester 1999; Fukazawa
1999) has proven untenable (Pater 2000; 2006; Flack 2007b; Flack 2007a; Inkelas & Zoll 2007; Gouskova
2007; Mahanta 2008; Jurgec 2010).
Assuming undominated *Float (Myers 1997), which prohibits surface subsegments, and low-ranked
Dep(Root) (Zoll 2001).
4 Phonological exceptionality is localized to phonological elements
(ma la:n) ≻ (mu la:) nu
(mar gu:ŋ) ≻ (mar guŋ) gu
(mu la:) ri ≻ (mu la:r)
(mar gu:n) da ≻ (mar gu:n)
I wish to draw a distinction now between two conceivable kinds of phonological analysis. A concrete phonological analysis represents exceptionality using regular phonological material, such as features, root nodes and prosodic units, or perhaps their absence. An abstract phonological analysis uses diacritic lexical indices, which I will
term lexical Φ-indices, on segments, much like the morphological analysis uses lexical
M-indices on morphs. Some objections which have been raised to phonological analyses
are specific to the concrete approach. These include doubts over whether sufficiently
many concrete phonological contrasts would be available in languages with very many
exceptional patterns (Gouskova 2012), and concerns over whether learners can choose
between multiple, alternative concrete representations (Kiparsky 1973, Pater 2009). I
will set these concrete-specific concerns aside for now, and instead assume an abstract
phonological approach. I return to the concrete approach in §9, where I argue on independent grounds that it is poorly adapted to efficient learning.
Accordingly, I will use lexical Φ-indices u and n to index undergoer and non-undergoer
segments respectively, and define Φ-indexed constraints, CΦ , in relative terms as in (21).
V(CΦ ) =def V(C) ∩ Lϕ (ϕ∈Φ), where:
Φ is the set of phonological elements co-indexed to CΦ .
Returning to the phonological account of Yidiny exceptionality, a constraint ranking
Max-n ≫ Prs ≫ Max, or Prs-u ≫ Max ≫ Prs, will be sufficient for our purposes.
Tableau (22) shows examples using the latter ranking; u-indexed segments are underlined.
(ma la:n) ≻ (mu la:) nu
(mar gu:ŋ) ≻ (mar guŋ) gu
(mu la:) ri ≻ (mu la:r)
(mar gu:n) da ≻ (mar gu:n)
(ma ɟin) (da ŋa:l) na ≻ (ma ɟin) (da ŋa:l)
Erich R. Round
A recent criticism of the phonological approach to exceptionality in OT is that it overgenerates (Gouskova 2012). Adapting Gouskova’s arguments to the facts of Yidiny: if
we adopt the ranking Prs-u ≫ Max ≫ Prs, then it is no longer necessary to assign a
high ranking to the morphologically-sensitive constraints Anc and Max-C/Rt, which
penalize the deletion of morph-initial segments and root consonants. Rather, so long as
all morph-initial segments and all root consonants lack a lexical u-index, then by virtue
of the partial ranking Max ≫ Prs, they will resist deletion irrespective of the ranking of
Anc and Max-C/Rt. By the same token however, if Anc and Max-C/Rt do receive a low
ranking, then the analysis will fare poorly in the context of Richness of the Base (Prince
& Smolensky 2004-1993]), since without high-ranked Anc and Max-C/Rt ensuring that
morph-initial and root-consonant deletion is impossible, there is nothing to prevent segments from deleting in those positions if they are u-indexed in the lexicon. For example,
a root such as *binarŋa could undergo CV deletion; a suffix *-ni could delete entirely;
and *mulari could delete from the left, thereby failing to capture the generalization that
the absence of such forms is not an accident of the lexicon, but a systematic property
of the grammar. This is perhaps the most significant apparent flaw of the phonological
approach: it fails to rule out unattested patterns. This is in contrast to the morphological
approach, which does rule them out. Or at least, so it would seem. In §5 I show that the
true situation can be otherwise, once learning is taken into account.
4.3 Alternatives
Before proceeding to learning, I mention two OT alternatives to the analysis of exceptionality in Yidiny word-final deletion.
Co-phonological approaches handle exceptionality as a type of cyclicity effect (Orgun
1996, Kiparsky 2000, Inkelas & Zoll 2007, Bermúdez-Otero 2016). On each morphological
cycle the result of a morphological operation is submitted to an appropriate phonological
subgrammar, of which the language may possess many. Problematic for any cyclicitybased approach to exceptionality in Yidiny word-final deletion is that the Yidiny case is
non-cyclic. Instead, undergoers are subject to deletion only if word-final. For example,
in building both words in (23a,b) the first step would be to introduce the undergoer
root bigunu- ‘shield’. However at that point, the “deleting” subgrammar should only be
applied if the root will end up word final, as in (23a) but not in (23b).
(23) a.
(bi gu:ŋ)
(bi gu) (nu yi:ŋ)
*(bi gun) (yiŋ gu)
Selecting the correct subgrammar in (23) thus requires information about the next
step in the derivation. Crucially though, it requires forewarning, not only of whether or
not there is more morphology to come, but also of what the phonological ramifications
will be. This is because the relevant domain for word final-deletion in Yidiny is not the
morphological word but the prosodic word. For example, in (24) the roots gaɟula- ‘dirty’
and gumaɻi- ‘red’ are followed by suffixes. Since the suffixes are monosyllabic, just one
4 Phonological exceptionality is localized to phonological elements
prosodic word results and the roots are non-final in their prosodic word. In (25) however,
the roots are followed by the disyllabic inchoative suffix daga, which commences a
second prosodic word. As a consequence, the roots are final in their prosodic word and
deletion is possible: the undergoer gaɟula- deletes while the non-undergoer gumaɻi- does
[(ga ɟu) (la ŋa:l)PWd ]
[(gu ma) (ɻi ŋa:l)PWd ]
[(ga ɟu:l)PWd ] [(da ga:ɲ)PWd ]
[(gu ma:) ɻi PWd ] [(da ga:ɲ)PWd ]
Any cyclic, look-ahead mechanism in Yidiny would therefore need to know how the
word would be prosodically parsed on the next cycle, before it can decide whether or
not to apply the “deleting” subgrammar on the current cycle. The look-ahead mechanism would therefore require the power of a subgrammar itself, yet if the theory were
augmented in this manner, then other core mechanisms such as scope, or “bracket erasure”, effects (Inkelas & Zoll 2007) would be undermined. I conclude that co-phonology
theory as it stands cannot analyse exceptionality in Yidiny word-final deletion.
Another approach would be to lexically list two allomorphs for all undergoer morphs
in the language, and have the grammar select them either optimally (Mester 1994, Kager
1996, Mascaró 1996, and Tranel 1996a,b) or with some degree of stipulation (Bonet, Lloret
& Mascaró 2007; Round 2013; Wolf 2015). On this approach, “deletion” is apparent only,
due in reality to the selection between two input allomorphs, one of which contains only
a subset of the segments in the other (for a proposal not unlike this for Yidiny, see Hayes
1997). An example is shown in (26), where the grammar optimally selects between two
input allomorphs of the undergoer root bigunu- ‘shield’.
{/bigunu/, /bigun/} ‘shield[abs]’
/bigun/ :: (bi gu:n)
/bigun/ :: (bi gu:) nu
/bigunu/ :: (bi gu:n)
/bigunu/ :: (bi gu:) nu
Two objections can be raised. First, because the approach simply lists alternant pairs, it
misrepresents their resemblances as accidents, rather than relating them systematically.
Relatedly, in the context of Richness of the Base, the analysis would allow the apparent
deletion of morph-initial and -medial segments as well as root consonants, by leaving
them out of an underlying allomorph, in a pair such as {/bigunu/, /gunu/}. Ranking Anc
and Max-C/root highly would not ameliorate the problem, as shown in (27).
Erich R. Round
{/bigunu/, /gunu/}
/gunu/ :: (gu nu)
/bigunu/ :: (bi gu:) nu
/buɟala-ŋa-lna/ :: (bu ɟa) (la ŋa:l) na
/buɟal-ŋa-lna/ :: (bu ɟal) (ŋal na)
†{/buɟala, buɟal/}-ŋa-lna/
‘finely ground-cause-purp’
Second, it is unclear how the analysis would prevent apparent deletion in word medial
positions in the event that it is optimsing, as in (28), where the true output buɟala-ŋa:-lna
violates Prs while the more optimal false winner *buɟal-ŋa-lna does not. The constraint
Cntg will not prevent this occurring.
I conclude that neither the co-phonological approach nor the allomorph-selection approach offers a viable alternative for Yidiny word-final deletion.
5 Learning exceptionality
5.1 Biased Constraint Demotion
I turn now to consider how exceptionality is, or isn’t, learned. After introducing Prince
and Tesar’s (2004) Biased Constraint Demotion (BCD) algorithm and adaptations of it
for the learning of indexed constraints, I show that the learning of Yidiny word-final
deletion does not proceed as one might expect from the discussion in §4. A solution is
then offered in §6.
Prince and Tesar’s BCD is a computationally efficient algorithm for the learning of
OT grammars. It builds upon Tesar’s earlier Recursive Constraint Demotion (RCD) algorithm (Tesar 1995, Tesar & Smolensky 2000), deterministically learning a grammar,
conditional on the data, by ranking constraints in a series of steps, or recursions. At the
first step, one or more constraints is assigned to the highest-ranked constraint stratum in the grammar. A stratum is a set of constraints whose relative ranking against
one another is indeterminate given the data, but whose ranking relative to constraints
in other strata is significant. The act of assigning constraints to a stratum is termed installation. At each subsequent step, one or more additional constraints are installed in
the next-highest stratum, and so on, until all constraints are ranked. The determination
of which constraint(s) are installed next is based on evidence from winner–loser pairs
(WLPs). For each WLP, any constraint yet to be installed will favor the winner in the
pair, the loser, or neither. The full table of WLPs and constraints yet to be installed is
termed the support. A fragment of a support is shown in (29). The relative order of
constraints and WLPs in a support is inconsequential, though for ease of inspection I set
4 Phonological exceptionality is localized to phonological elements
(mar gu:n) ≻ (mar gu:) ni
(guy ga:l) ni ≻ (guy ga:l)
(guy ga:l) ni ≻ (guy ga:ln)
(guy ga:l) ni ≻ (guy ga:l ni)
(bulm ba) ≻ (bul ba)
out markedness constraints to the left of a vertical double line, and faithfulness to the
In the original RCD algorithm, the sole criterion for installing a constraint was that it
favor no losers. This is true of the constraints FtBin, Cntg and Anc in (29). When a constraint, C, is installed, all of the WLPs for which C favors the winner are removed from
the support, since the constraint ranking has now accounted for them. In the RCD, all
constraints meeting this criterion at any recursion are installed, and the result at the end
of all recursions is a correct grammar for the data. Nevertheless, the grammars inferred
by the RCD are not optimal (Prince & Tesar 2004). The suboptimality relates to the subset
problem (Baker 1979; Angluin 1980), a general problem in algorithmic learning from positive evidence, namely that the system which results from learning will correctly assess as
grammatical all attested items, but will fail to rule out certain systematically unattested
items. This in turn relates to the notion of restrictiveness: a learning algorithm ought
ideally to learn the most restrictive grammar consistent with the data. The RCD does not
do this. In practice, meeting this desideratum is challenging for an efficient algorithm.
However Prince & Tesar (2004) demonstrate that good headway can be made by enhancing the RCD with a small set of biases, hence the name Biased Constraint Demotion,
or BCD. The BCD differs from the RCD in two main respects. The first is the principle
of faithfulness delay. According to this, at every recursion faithfulness constraints
are not installed, even when they favor no losers, unless there are no other installable
constraints. In (29) for example, the BCD would install the markedness constraint FtBin
but not the faithfulness constraints Cntg and Anc. If we do this, and remove from (29)
all the WLPs for which FtBin favors the winner, namely (29d), and remove FtBin, we
have (30), in which only faithfulness constraints, Cntg and Anc favor no losers; under
these conditions, faithfulness delay would permit the installation of Cntg and Anc.
Erich R. Round
(mar gu:n) ≻ (mar gu:) ni
(guy ga:l) ni ≻ (guy ga:l)
(guy ga:l) ni ≻ (guy ga:ln)
(bulm ba) ≻ (bul ba)
However, there is a second principle to consider also. A principle of “freeing up
markedness” states that when there is a choice between installing several faithfulness
constraints, the algorithm should install the smallest subset possible, whose installment
would cause a markedness constraint to become installable in the next recursion. For example, in (30), installing Cntg would remove WLP (30e), thereby freeing up the markedness constraint *Cplx at the next recursion; no comparable gain would flow from installing Anc. On those grounds, from (30) the BCD would install Cntg.
5.2 A support for learning Yidiny exceptionality
I now consider several learning scenarios for Yidiny exceptionality. Each begins directly
after the installation of undominated constraints. Table 6 contains a set of WLPs that is
representive of all combinations of roots and suffixes which are relevant to the grammar
of word-final deletion: it is not the complete support, but it represents the complete
support well. Segments which can delete are underlined. To economize on space below,
WLPs will be referred to by the letters in the first column of Table 6.
5.3 Learning the phonological account (preliminary version)
We begin with the learning of the phonological account of Yidiny exceptionality described previously in §4.2. For the moment, I assume that input segments are already
lexically Φ-indexed as u or n. We begin after undominated constraints have been installed, with a support as in (31).
4 Phonological exceptionality is localized to phonological elements
Table 6: Support for learning Yidiny exceptionality.
a, h, i.
d, g.
j, k, l, o.
m, n.
(mar gu:n)
≻(mar gu:) ni
(guy ga:l) ni
≻(guy ga:l)
(guy ga:l) ni
≻(guy ga:ln)
(mar gu:ŋ)
≻(mar gu:ŋ) gu
(mar gu:ŋ)
≻(mar gu:ŋg)
(bi gu) (nu yi:ŋ)
≻(bi gun) (yiŋ gu)
(wa wa:l)
≻(wa wa:l) ɲu
(ga li:ŋ)
≻(ga li:) ŋa
(ga ɟa:r)
≻(ga ɟa:) ra
(mar gu:n) da
≻(mar gu:n)
(wa wa:l) na
≻(wa wa:l)
(ga li:) na
≻(ga li:n)
(gu ɟa:) ra
≻(gu ɟa:r)
(ma ɟin) (da ŋa:l) na ≻(ma ɟin) (da ŋa:l)
(bulm ba)
≻(bul ba)
Support (31) does not contain any markedness constraints that favor no losers. Two
faithfulness constraints favor no losers: Cntg, which would free up *Cplx if installed,
and Anc, which would not free up any markedness constraints. Consequently, Cntg
is installed next, removing WLPs (f) and (p) from the support. After that, the newly
freed-up *Cplx is installed, removing WLPs (c) and (e), and leaving (32).
Erich R. Round
a, h, i.
d, g.
j, k, l, o.
m, n.
In 32 only Anc favors no losers, and so is installed. This removes (b), freeing up
Prs-u, which is installed next, removing (a,h,i) and (d.,g), leaving (33). From (33), Max
will be installed since it frees up Prs. This leaves Prs and Max-C, which according to
faithfulness delay, will be ranked last as Prs ≫ Max-C, as in (34).
j, k, l, o.
m, n.
Cntg ≫ *Cplx ≫ Anc ≫ Prs-u ≫ Max ≫ Prs ≫ Max-C
Some comments are in order. First, the BCD algorithm has learned the key constraint
ranking Prs-u ≫ Max ≫ Prs responsible for the core of Yidiny exceptionality. Secondly
however, it has also ranked Anc ≫ Prs-u, in which case the learned grammar expressly
prohibits morph-initial deletion. Indeed, had Max-C/rt been included in (31), it would
also have been ranked highly since it only ever favors winners, meaning the grammar
would also expressly prohibit CV deletion in roots (the reasons for my excluding MaxC/rt are clarified in §6). This means that the algorithm is learning precisely the rankings
required to prevent the phonological solution from overgenerating, thereby voiding the
major criticism of the phonological approach which was introduced in §4.2. This is perhaps surprising, so why is the ranking learned? It is learned because the BCD algorithm
attempts to construct a restrictive grammar. The typical assumption, that grammars
implementing a phonological approach would not assign redundant, high rankings to
constraints like Anc, is predicated on an implicit assumption that the learner would be
seeking a permissive grammar; doing so leads to overgeneration. However no successful learner would adopt that assumption, because successful learning in general requires
a restrictive approach. For the theory of exceptionality, this is significant. It means
the result obtained here, in which a phonological approach to exceptionality has been
learned without overgeneration, is not dependent on some minor detail of the BCD, or
the constraints used, or even OT. Rather, it follows from a general principle of learning.
Consequently, the adoption of realistic assumptions about learning narrows the performance gap between the phonological and morphological approaches. I will examine the
phonological approach further in §7.3.
4 Phonological exceptionality is localized to phonological elements
5.4 Learning indexed constraints and the morphological analysis
We consider next the learning of the morphological approach. The support begins, after
installation of undominated constraints, as (35). These are the same constraints and
WLPs as in the previous section, but without Prs-u. The support begins with no lexically
indexed constraints; how they are learned is considered shortly. I also do not include
Max-C/rt in the support. Max-C/rt is essentially a variant of Max-C, indexed to all root
morphs. This is the kind of constraint we might reasonably expect the morphological
approach to learn.
a, h, i.
d, g.
j, k, l, o.
m, n.
Turning now to the BCD algorithm, neither of the markedness constraints in support
(35) favors no losers. Cntg does, and would free up *Cplx. Anc also does, but would
not free up any markedness constraints. Accordingly, Cntg is installed next, removing
WLPs (f) and (p) are from the support, and *Cplx after that, removing (c) and (e), leaving
a, h, i.
d, g.
j, k, l, o.
m, n.
Anc is installed next, removing WLP (b), which leaves (37), a support in which there
is no constraint which favors no losers.
Erich R. Round
a, h, i.
d, g.
j, k, l, o.
m, n.
Supports in this state are said to have reached inconsistency. An inconsistency, however, is not a failure.
Inconsistencies indicate that the combination of data and assumptions currently under
consideration have not led to a working grammar. Accordingly (assuming the data is
correct), a revision of the assumptions is warranted. Suppose, in this case, that a revision
could be made which leaves intact all previously installed constraints and their rankings,
and the validity of all previously accounted-for WLPs, that is, a revision that would
change only what is in the support. Suppose also that as a result of this revision the
support came to contain a constraint that favors no losers. Such a revision would resolve
the inconsistency. The BCD could restart and, one hopes, lead to a working grammar.
Revisions that meet these criteria can be considered a type of learning. One such revision
is to add a new, lexically M-indexed constraint to Con.
Pater (2009) describes a method for learning M-indexed constraints and assigning coindices to morphs, which takes a BCD inconsistency as its starting point. Coetzee (2009)
extends this to Output-Output constraints, which I will not consider here. Becker’s modifications (Becker 2009; Becker, Ketrez & Nevins 2011) are addressed in §8.
Central to Pater’s method is the operation of constraint cloning, a process I describe informally here and return to in detail in §8. Within the stalled support, a constraint C is sought which, if it were indexed to some set M of morphs, would (i) favor at
least one winner15 and (ii) favor no losers. Assuming such a constraint C can be identified, it is then cloned, which is to say, a lexically M-indexed version of it, CM , is added
to the support. Because CM favors no losers, it is installed next. For example, support
(38) is the same as (37) but now displays information about which morphs are involved.
I have annotated relevant undergoers as U and non-undergoers as N.
According to the criteria for cloning, all three of Prs, Max and MaxC are candidates
for cloning (indexed to sets U, N and N respectively). I assume that owing to faithfulness delay, markedness constraints are cloned in preference to faithfulness when both
are available, in which case Prs will be cloned. In (39) the cloned, lexically M-indexed
constraint PrsU is added to the support. Installing it removes WLPs (a,d,g,h,i) which
frees up Max, whose installation is followed by Prs and Max-C. The resulting ranking
is (40), which requires comment.
The new constraint needs to favor at least one winner to have any chance of freeing up another constraint
once it is installed.
4 Phonological exceptionality is localized to phonological elements
/margu-niU /
/margu-niU /
/margu-ŋguU /
/gali-ŋaU /
/gaɟaraU /
/margu-ndaN /
/wawa-lnaN /
/guɟaraN /
/maɟinda-ŋa-lnaN /
a, h, i.
d, g.
j, k, l, o.
m, n.
(40) Cntg ≫ *Cplx ≫ Anc ≫ PrsU ≫ Max ≫ Prs ≫ Max-C
The algorithm has successfully learned the key constraint ranking PrsU ≫ Max ≫
Prs. However, it did not create an indexed version of Max-C for roots, and thus has
not learned to expressly prohibit CV deletion in roots. To be sure, no individual roots
ending in CCV or VCʹV (where Cʹ would be an illicit coda) will have been co-indexed to
PrsU during the cloning operation (see §8 for details) and so none of those roots will be
subject to CV deletion, however the ranking in (40) predicts that if the lexicon did contain
a root such as *binarŋaU , then that root and any like it would undergo CV deletion. This
is overgeneration of the same kind which was believed to beset phonological accounts.
Thus, while §5.3 showed that grammars learned for the phonological account may suffer
less than expected from overgeneration once learning is taken into consideration, §5.4
shows that grammars for the morphological account may suffer from overgeneration
more than expected. In the next section, I propose a solution.
6 Morphological analytic bias: the Morphological
Coherence Principle
In §5.4 the grammar which was learned for a morphological analysis of Yidiny exceptionality suffers from a manifestation of the subset problem. Although the algorithm
Erich R. Round
correctly handled all attested data, it did not learn the more restrictive generalization
which applies also to unattested data, that roots in Yidiny do not undergo consonant
deletion. The problem arises because the cloning procedure assesses morphology on a
morph-by-morph basis only, whereas the true generalization in Yidiny applies to a class
of morphs, in this instance, to roots. The remedy to be pursued here has two parts. It
adds a new kind of constraint cloning, which indexes a constraint not to an idiosyncratic
lexical list of morphs, but to a general class. It then biases constraint cloning so that classindexed (or K-indexed) cloning is preferred over lexically indexed cloning. Effectively,
this introduces an analytic bias (Moreton 2008) from morphology to phonological learning at BCD inconsistencies.
Now, supposing that the algorithm is seeking a constraint that it will clone and Kindex to some non-idiosyncratic class of morphs, which classes should be available
for the learner to consider? Important here is the fact that human phonological learning will need to proceed in parallel with, and interleaved with, morphological learning
(Tesar 2007, Merchant 2008: 6). Accordingly, I assume the learner has access both to
universally-defined classes such as “root”, and those classes which have been morphologically learned, such as ergative case. The biasing principle, which I term the Morphological Coherence Principle is stated in (41), where criterion 2 provides an additional
bias towards maximal restrictiveness.
The Morphological Coherence Principle:
1. At a BCD inconsistency, attempt to create a K-indexed constraint,
co-indexed to some universal or learned morphological class K, before
attempting to create a lexically-indexed constraint.
2. If multiple constraints are eligible for K-indexation, select the one whose
co-indexed class is most general.
The MCP has some desirable theoretical properties. If the universal state of Con at
the commencement of learning is Coninit , then the MCP obviates the need for Coninit to
contain any constraints that are relativized to universal or learned morphological classes,
since such constraints will be learned on demand, if and only if needed. In effect, this
reduces the size of Coninit without any change in the explanatory capacity of the theory.
And, since it allows the grammar to build constraints for language-specific morphological classes it makes those constraints available to the learner without problematically assuming them universal (Russell 1995, Hammond 2000, see also Smith 2004, Flack 2007b).
The MCP operationalizes, in a specific manner, the kind of insight into linguistic theory
that Anderson (2008) argues ought to follow from an improved understanding of the
learning device.
Let us now return to Yidiny exceptionality, equipped with the MCP. Learning begins
and proceeds as in §5.4 until the inconsistency in (38), at which point a constraint is
sought for cloning. The MCP states that if possible, a constraint should be cloned and
K-indexed. In (38) Max-C would favor no losers if it were K-indexed to the entire class
of roots, so it is cloned and accordingly K-indexed. This is the functional equivalent of
adding Max-C/rt to Con, and the reason why in §5 I did not include Max-C/rt in the
4 Phonological exceptionality is localized to phonological elements
support at the outset. Adding Max-C/rt to the support results in (42). From (42), MaxC/rt is installed and WLP (j) is removed, whereupon we return to inconsistency, in (51).
As in §5.4, the process from that point results in the cloning of Prs and the installation
of PrsU , then Max, Prs and Max-C, yielding the desired constraint ranking (43).
a, h, i.
d, g.
k, l, o.
m, n.
a, h, i.
d, g.
k, l, o.
m, n.
(44) Cntg ≫ *Cplx ≫ Anc ≫ Max-C/rt ≫ PrsU ≫ Max ≫ Prs ≫ Max-C
To summarize, results from §5.3 suggested that, provided a learner is seeking a restrictive grammar, the phonological approach to exceptionality may not suffer from overgeneration. This contradicts recent arguments, which on examination appear to adopt the
implausible assumption that a learner would be seeking a permissive grammar. That being said, I have not yet clarified how the learner would arrive at the requisite Φ-indices
required by the phonological approach. That will be discussed in §7.3. Meanwhile, §5.4
revealed that without further refinement, the BCD is prone to learning grammars that
overgenerate even in a morphological approach to exceptionality, due to an overly atomistic method of morphological generalization. This was remedied in §6 by the Morphological Coherence Principle (41), which solves the learning problem and simplifies Coninit .
7 The theoretical status of lexical indices
In §7 I set Yidiny to one side and consider some matters of theory.
7.1 Lexical M-indices
Lexical M-indices are representations which are visible to the phonology, but they are
not phonological elements per se. In OT, Gen cannot alter M-indices. It cannot add
or remove them, or displace them from one morph to another. There is therefore no
need for mechanisms such as M-index “faithfulness”, rather it is simply assumed that
Erich R. Round
the lexical affiliation of a morph m with an M-index M is identical in the input and
output. This set of properties is shared with other kinds of lexical affiliation, such as
the affiliation of a phonological element with its morph, and is termed Consistency of
Exponence (McCarthy & Prince 1993b, Van Oostendorp 2007).
Taking a historical view, M-indices closely resemble the rule features and alphabet
features of early generative phonology (GP) (Chomsky & Halle 1968, Lakoff 1970, Coats
1970, Zonneveld 1978, inter alia). Both sets of formalisms fulfill the function of determining for cases of exceptionality whether a morph m participates in certain phonological
patterns or not, by ensuring that m is visible or not visible as required, to OT’s constraints or GP’s phonological rules. Diacritic features were investigated extensively in
GP. It was argued that the theory should not allow the phonology to manipulate diacritic
features (Kiparsky 1973, Zonneveld 1978). The same applies to M-indices in OT. It was argued that not all idiosyncrasies in the phonology can be analysed satisfactorily in terms
of rule exception features, and that there is an additional role for cyclicity (Chomsky
& Halle 1968, Kiparsky 1982b) and the same has been recognized for M-indices (Pater
2009). In GP, it was also assumed that the diacritic features of morph m were distributed
across, and directly characterized, each of the phonological elements (namely, segments)
in m. We might ask whether this is also true of M-indices in OT. Suppose that it is, so
that the M-indices of a morph m directly characterize each phonological element ϕ that
is lexically affiliated with m (that is all ϕ which are exponents of m). In that case, the
relative definition of an M-indexed constraint (25), repeated here as (45), can be revised
and simplified as (46).
(45) V(CM ) =def V(C) ∩ Lϕ (m∈M & Exp(ϕ, m)), where:
M is the set of morphs co-indexed to CM .
Exp(ϕ, m) states that element ϕ is an exponent of morph m
(46) V(CM ) =def V(C) ∩ Lϕ (ϕ∈ΦM ), where:
ΦM is the set of phonological elements co-indexed to CM .
It will be recalled that the relative definition of a constraint CM is expressed as the set
intersection between the loci of variation of the unindexed constraint C, written V(C),
and the set of loci, Lϕ (D(ϕ)) which contain some criterial type of phonological element
ϕ, described by predicate D(ϕ). Importantly, this means that M-indexed constraints are
defined directly in terms of phonological elements, ϕ, and only indirectly in terms of
morphs m. The indirectness shows up in the complexity of D(ϕ) in (45), which links
morphs to their exponent ϕ elements via the function Exp(ϕ, m). This is in contrast with
(46), where the assumption is that all ϕ elements are directly characterized by the Mindex borne by their affiliated morph. The constraint definition no longer refers to the
morph itself, and so the predicate D(ϕ) is simpler.
At risk of laboring the point, the phonology itself assesses violations of M-indexed
constraints directly in terms of ϕ elements, not morphs. While it is possible to refer to the
morphs in the definitions of M-indexed constraints as in (17/45), it is not necessary. Nor is
it possible to refer only to the morphs and not to the ϕ elements, since the loci of violation
of these constraints are defined inherently at a sub-morphological, phonological level.
4 Phonological exceptionality is localized to phonological elements
7.2 Lexical Φ-indices
Let us now consider the nature of lexical Φ-indices of the type I invoked in §4.2 and §5.3.
My proposal is that these are exactly like M-indices: they are non-phonological indices of
lexical affiliation, visible to, but not manipulable by, the phonology and used for making
particular phonological elements visible or not, as required, to OT’s constraints in order
to provide a coherent account of exceptionality. The only distinction between Φ-indices
and M-indices lies in the supplementary assumption attached to M-indices, in (47).
The M-index assumption:
A lexical index which characterizes phonological element ϕ i will also
characterize all other phonological elements ϕ j affiliated with the same morph m.
Φ-indices are not subject to this redundancy; they are affiliated with only those ϕ elements for which the affiliation makes any difference to the analysis of language. As I
will show in §8, that makes Φ-indices somewhat simpler to learn, since they correspond
more directly to the evidence in the data.
The reader may also have noticed that the definition of a Φ-indexed constraint in
(21) is almost exactly like the simplified definition of an M-indexed constraint in (46).
This reflects the fact that for the operation of the phonology, it is ϕ elements, and the
indexation of specific ϕ elements, that matter. Whether or not one chooses to adopt
supplementary assumption (47) in fact has no material consequence for the evaluation of
an individual indexed constraint. The question of whether there are other consequences,
and whether they are desirable, is taken up in §9.
7.3 Learning lexical Φ-indices
Given the proposal above, the learning of Φ-indices is quite parallel to the learning of
M-indices. I assume that the MCP still applies, so that class-based exceptionality and
K-indexed constraints continue to be learned with priority over idiosyncratic exceptionality, even though the latter will now be accounted for by Φ-indexed constraints, not
M-indexed. This is a coherent assumption to make. The MCP is concerned with the
learning of class-based generalizations, whereas Φ- and M-indexed constraints are alternative devices for learning idiosyncrasies. Accordingly, in a stalled support once there
are no K-indexed constraints available for cloning, the algorithm seeks a constraint C
which, were it indexed to some set Φ of phonological elements, would (i) favor at least
one winner and (ii) favor no losers. All else proceeds as for M-indexed constraints. In
the learning of Yidiny word-final deletion, the process begins as in §6, leading to a first
inconsistency resolved by the addition of Max-C/rt to Con, and proceeding from there
to the second inconsistency (43), repeated here in part and in more detail as (48).
Erich R. Round
(mar gu:n) ≻ (mar gu:) ni
(mar gu:ŋ) ≻ (mar gu:ŋ) gu
(ga li:ŋ) ≻ (ga li:) ŋa
(ga ɟa:r) ≻ (ga ɟa:) ra
(mar gu:n) da ≻ (mar gu:n)
(ga li:) na ≻ (ga li:n)
(gu ɟa:) ra ≻ (gu ɟa:r)
(ma ɟin) (da ŋa:l) na ≻ (ma ɟin) (da ŋa:l)
In (48), no K-indexed constraint is available for cloning.16 Turning to potential Φindexed constraints, we see that the constraint Prs would, if it were co-indexed to all
underlined phonological elements, favor at least one winner and favor no losers, and so
it is cloned and co-indexed resulting in (49).
(50) Cntg ≫ *Cplx ≫ Anc ≫ Max-C/rt ≫ Prs-u ≫ Max ≫ Prs
Actually this is not strictly true. All past suffixes for example are undergoers, in which case the MCP would
generate and rank PrsPST . Notwithstanding this, the essential argument remains, since other morphological classes exist, such as ergative and “root”, that are not uniformly (non)undergoers, and still need to
be handled by lexically-indexed, not K-indexed, constraints. This minor correction applies equally to the
learning process in §6.
4 Phonological exceptionality is localized to phonological elements
From there the algorithm proceeds in the now-familiar fashion, resulting in grammar
(50). With its high-ranking Max-C/rt and Anc, (50) does not overgenerate. Moreover,
given the argument in §7.2, that for Eval there is no detectable difference between Mindexed and Φ-indexed constraints, we can see that grammar (50) is in all material aspects identical to grammar (44) learned in §6.
8 Constraint cloning
8.1 Assessing eligibility for cloning
It is necessary now to examine more precisely the processes by which constraints are
deemed eligible for cloning (§8.1), by which a viable set of co-indexed elements is identified (§8.2), and by which a selection is made between multiple eligible constraints (§8.3).
Earlier, I introduced criteria by virtue of which a constraint becomes eligible for cloning. These are restated in (51) in a generalized from, so that the set S is: a coherent
class of morphs for K-indexing; an idiosyncratic set of morphs for M-indexing; or an
idiosyncratic set of lexical phonological elements for Φ-indexing.
A constraint should be sought for cloning which, if it were indexed to set S,
would (i) favor at least one winner, and (ii) favor no losers.
Criterion (⁇ii) ensures that once the cloned constraint is added to the support, it can
be installed; (51i) ensures that its installation will remove at least one WLP from the
support, and thereby have some hope of freeing up other constraints. The formulation
in (51) improves upon Pater’s (2009: 144) criterion, which is to seek a constraint that
favors no losers “for all instances” of some morph.17 To see why Pater’s criterion fails,
consider WLPs (h,l,o) from the stalled support (38), reproduced in part and in detail in
(52). For the purposes of discussion, I assume we are attempting to learn an M-indexed
constraint, though the argument generalizes to other kinds.
/gali-ŋa/ ‘go-com[imp]’
(ga li:ŋ) ≻ (ga li:) ŋa
/wawa-lna/ ‘see-purp’
(wa wa:l) na ≻ (wa wa:l)
/maɟinda-ŋa-lna/ ‘walk up-com-purp’
(ma ɟin) (da ŋa:l) na ≻ (ma ɟin) (da ŋa:l)
In (52), WLPs (h) and (o) both contain the suffix -ŋa, a regular undergoer which our
procedure ought to co-index to the M-indexed constraint PrsU . In WLP (h) word-final
ŋa is subject to deletion, and Prs favors the winner. In WLP (o) non-final ŋa is parsed
Pater’s phrase “favors only winners” is equivalent to my “favors no losers”.
Erich R. Round
into a foot and escapes deletion. Nevertheless, for WLP (o) Prs favors the loser. This
has nothing to do with ŋa, but is due to the non-deletion of the unparsed, word-final
non-undergoer -lna. Pater’s co-indexing criterion asks whether Prs favors no losers “for
all instances” of -ŋa in the support. The answer is “no”, because (o) contains an instance
of -ŋa and Prs favors the loser for (o). This is the wrong result; the suffix -ŋa ought to
get co-indexed to PrsU . It comes about because Pater’s criterion does not discriminate
between morphs that contribute to violations and those which are present in the word,
but do not contribute. The criteria in (51) avoid this problem because they refer directly
to how the co-indexed constraint would perform, were it created. The next two sections
detail how to operationalize them.
8.2 Specifying co-indexed sets
The question considered here is which set S ought to be co-indexed to a given constraint
C if we wish to clone C? The answer varies depending on which kind of indexed constraint we are constructing. One possible answer is that no such set exists, and C cannot
be cloned. Seen from that angle, the question here is also: is C eligible for cloning?
K-indexed constraints can be co-indexed only to the morphological classes K 1 , K 2 …K n
in the language (§6). In (41) I suggested that the preferred class for co-indexation is the
most general one. Thus, to efficiently assess if constraint C is eligible for cloning and Kindexing, the learner should proceed stepwise through the available classes, ordered by
decreasing generality. The process is one of trial and error. At each step, the constraint
CK is built and applied to all WLPs in the support. If CK meets criteria (51) then it is
successful; the process halts and CK is used, otherwise the trial and error continues. If
by the end, no successful constraint CK1 …CKn is found, then C is ineligible for cloning.
For M-indexed and Φ-indexed constraints, the desired set S can be identified by focusing attention on loci of violation. Suppose we are considering constraint C for cloning.
For any WLP, p, its loci of violation of constraint C fall into three classes: the class w(p),
responsible for violations of C that favor the winner (i.e., the locus occurs in the loser
only), class l(p) which favor the loser (locus occurs in the winner only) and class n(p)
which favor neither (occurs in both). Next define Φw(p) as the set of phonological elements ϕ contained in any of the loci in w(p), and Φl(p) as the set of ϕ elements contained
in any of the loci in l(p). Finally, define ΦW as the union of Φw(p) for all WLPs, p1 , p2 …
pn , in the support, and ΦL as the union of all Φl(p) in the support. Now, consider the set
(ΦW – ΦL ), the set difference between ΦW and ΦL . This is the set of all ϕ elements which
both (i) appear in at least one locus that in at least one WLP causes C to favor a winner,
and (ii) never appear in a locus that causes C to favor a loser. For a Φ-indexed constraint
this is an optimal set S. If for a given constraint C, (ΦW – ΦL ) is the null set, then we may
conclude that C is ineligible for cloning.18
To be precise, if (ΦW – ΦL ) is the null set then it is possible that there still exists some additional, viable
set S which contains fortuitous elements ϕ i which are elements of both ΦW and ΦL such that in every
WLP p in which ϕ i is contained in some number n of the loci w(p) there are at least n offsetting loci in
l(p) which contain other elements ϕ j which are also in S. Identifying these fortuitous elements ϕ i , or even
determining if any exist, would very likely be prohibitively expensive computationally.
4 Phonological exceptionality is localized to phonological elements
To find the equivalent for an M-indexed constraint, it is necessary to extrapolate from
ΦW and ΦL to morphs: set S will be the set (M W – M L ) where M W is the set of all morphs
mw , such that any of mw ’s phonological exponents is an element of ΦW ; and M L is the set
of all morphs ml , such that any of ml ’s phonological exponents is an element of ΦL . Note
that M W and M L can be calculated only after the calculation of ΦW and ΦL is performed.
In §7.1 I considered what is involved computationally in assessing violations of Φand M-indexed constraints, and argued that the calculations for both are essentially concerned with ϕ elements, not morphs. Here we see that the same is true when learning
the co-indexed set. As in §7.1, one can bring morphs into the picture, to be sure, but in
both cases doing so requires additional computational effort, for no effective difference
in how the grammar will work. In §9 I will argue the theory to be preferred is one which
admits lexically Φ-indexed constraints, but not M-indexed.
8.3 Selecting among eligible constraints
Suppose there are multiple lexically-indexed constraints which are eligible for cloning;
which do we choose? The principles of faithfulness delay and freeing-up of markedness
constraints will eliminate some options (§5.1). Beyond that, I suggest the learner chooses
the constraint which favors the most winners, and whose installment would therefore
remove the greatest number of WLPs. A desirable consequence will be a bias toward
restrictiveness. For example, suppose Max is eligible. If so, then so too is Max-C, MaxV, Max-p, etc. This “maximize-winners” criterion would select Max, and increase the
restrictiveness of the grammar, relative to the other options.
Interestingly, Becker (2009) proposes a minimize-winners criterion, whose effect is
to generate many, very specific cloned constraints, each indexed to highly specific subclasses in the lexicon. The aim is to account for a particular phenomenon, which I describe here. I argue that other accounts are possible, and that Becker’s solution has
undesirable consequences.
When language learners assign novel words to existing grammatical categories, they
do so on the basis of statistical correlations that exist in the lexicon, for example between category membership and aspects of the members’ phonological forms (Poplack,
Pousada & Sankoff 1982; Albright 2002). One such task is to assign a word as exceptional or non-exceptional, given evidence which underdetermines that choice. The key
question here is, what existing statistical knowledge do speakers use, and what do they
ignore? In Turkish, speakers appear to ignore correlations between the (non)alternation
of a stop’s laryngeal features and the quality of its neighboring vowel. It is proposed
(Becker 2009; Becker, Ketrez & Nevins 2011) that this is because speakers do not access
lexical statistics per se, rather they attend to the statistics of constraint indexation. Importantly, Con lacks constraints such as *[+high]tV which refer to a stop and the quality of its vocalic neighbor. Consequently, no such constraint can be indexed, making
such correlations invisible and hence irrelevant to a speaker when she assigns a novel
word to a (non)exceptional lexical category. Assuming this is the case, then in order for
fine-grained knowledge to be available to speakers, an atomizing, “minimize-winners”
Erich R. Round
criterion for cloning is needed. However, this solution would seem neither necessary
nor warranted.
Notwithstanding the facts of Turkish, speakers in other languages and performing
other novel-word tasks do use lexical correlations which lack a corresponding constraint
in Con (Moreton & Amano 1999, Albright 2002, Albright & Hayes 2002, Ernestus &
Baayen 2003), indicating that speakers are capable of such computation. In that case,
atomized indexed constraints alone are not enough to produce the Turkish results. An
additional stipulation is required, that this ability is suppressed when assigning novel
words to exceptionality classes; yet this leads to a curious view of phonology. Whereas
the grammar is usually the store of generalizations, just in the case of exceptionality, it is
a store of highly detailed idiosyncrasy, and just in that case speakers ignore their usual,
lexical store of idiosyncrasy and turn to the grammar. More satisfying would be to find
some other explanation of the Turkish data. While that would take us well beyond this
scope of this paper, it can be noted that what is required is a mechanism that can filter
the lexical information in some way. That mechanism needn’t be part of the OT grammar. Indeed, if it is true that learners build certain constraints during learning (Flack
2007b, Hayes & Wilson 2008, Hayes 2014), then there must exist extra-grammatical
generalization devices, which may provide the lexicon-filtering power needed. For now
I conclude that that Becker’s proposal follows from just one possible solution to an interesting puzzle, however both the puzzle and solution are outliers relative to what else
we know. In contrast, a “maximize-winners” criterion leads to the learning of restrictive
grammars, and on those general grounds would appear correct.
9 Discussion
9.1 The case against concrete accounts
Throughout this paper, I have considered only the abstract phonological approach to
analyzing exceptionality, gradually building the argument that its superiority to the morphological approach lies in the fact that it localizes exceptionality to specific ϕ elements,
which are the elements in terms of which the relevant computation must be carried out.
Concrete phonological approaches also localize exceptionality at a sub-morphological
level, but compared to the abstract approach they are ill-suited to learning, and to seriality, as follows.
Lexical indexation is an ideal response to BCD inconsistency, because it annotates
the lexicon with indices which are invisible to all previously installed constraints. This
guarantees, without needing to check, that all previously accounted-for WLPs remain
accounted for. Even if some of them contain lexical ϕ elements which acquire a new
index, their violations of all previously ranked constraints remain unchanged, since no
previously-ranked constraint is sensitive to the new index. In contrast, the alteration of
phonological form — for example, removing a root node from certain segments — may
very well alter the evaluation of WLPs by already-ranked constraints, thus it requires a
re-evaluation of the entire ranking. It is not possible to simply repair an inconsistency
4 Phonological exceptionality is localized to phonological elements
and resume the BCD process. An abstract phonological account is therefore easier to
In serial theories, concrete phonological approaches face the problem that in noninitial strata, it is possible that a preceding stratum will have removed, altered, moved or
introduced, those aspects of phonological form which should function as pseudo-indices,
which lack Consistency of Exponence. This opens up the possibility of all manner of
phonological manipulations of exceptionality, for which I am unaware of any evidence.
Taking a more historical view, Chomsky (1964) criticized concrete phonological accounts espoused by structuralists (e.g. Bloomfield 1939) for the proliferation of underlying segments that they entailed. To the extent that such concerns matter to modern
phonological theories, Φ-indexation avoids such proliferation by augmenting representations with non-phonological indices (cf §7), rather than additional underlying phonological distinctions.
9.2 The case against M-indexing
In §7 and §8 I showed that for both constraint evaluation and constraint learning, exceptionality is calculated in terms of phonological elements, not morphs. Morphs can
be brought into the picture, but at additional computational cost and to no effect. Perhaps, however, it is nevertheless empirically true that exceptionality is inherently morphbound. If that were so, then phonological exceptionality in any morph m would always
be either (i) uniform throughout all phonological exponents of m or (ii) entirely predictably located within m. Yet this is not the case. If we accept something along the lines
of Anderson’s (1982) analysis of French schwa as an exceptionally-deleting /ø/ vowel,
then that exceptional property is neither uniform throughout morphs nor does it have
a predictable location. Similarly, in Turkish, non-high round vowels are phonotactically
exceptional outside the first syllable (Clements & Sezer 1982; Hulst & Weyer 1991), yet the
location of the exception is not predictable, as seen in a comparison of otoban ‘highway’,
monoton ‘monotone’, fenomen ‘phenomenon’ and paradoks ‘paradox’. There is no doubt
that in most known cases, exceptionality does happen to be either uniform or predictable
within a morph, but this follows uninterestingly from the fact that most exceptional
morphs are short, or that most phonological alternations are either local, in which case
their location inside a morph is predictably restricted to an edge, or domain-spanning,
in which case the morph acts uniformly. However, when such uninformative cases are
set aside, the small, informative residue of evidence does not support the morph-based
A second argument in defense of M-indices might be that morphs, and not ϕ elements,
belong to lexical strata, and that a single morphological diacritic can therefore coherently
index a whole set of phonological exceptionality patterns, patterns which impact different parts of the morph and which therefore would be only incoherently represented by
individual diacritics on ϕ elements. Yet the empirical falsity of this claim has long been
recognized. SPE (Chomsky & Halle 1968) permitted both stratal diacritics, later labeled
morphological features (Postal 1968) and more specific rule features (Lakoff 1970),
Erich R. Round
in view of the fact that distinct phonological patterns associated with strata are not uniformly attested in all morphs. For more recent work, see for example Labrune (2012:
71,72,85ff) on Japanese.
A third argument in defense of M-indices might be that since some kinds of phonological exceptionality are cyclic (§7.1), and since cycles are inherently tied to morphology, not
ϕ elements, then something like M-indices are required anyhow, in which case Φ-indices
are redundant. I would suggest that this is a category mistake. While it is true that cycles
are inherently tied to morphology, they are tied not to morphs, but to morphological
operations. Some operations are non-concatenative and hence morph-free (Anderson
1992). Cyclicity effects, therefore, are about how phonological subgrammars correlate
with operations; in contrast, Φ-indices are about correlation with forms. M-indices fall
uncomfortably in between. Since they are inherently attached to morphs, they will be
unavailable for the triggering of cyclicity effects associated with non-concatenative operations. And, as we have seen above, they are inefficient, and in all likelihood insufficient,
devices for exceptionality of forms.
10 Conclusion
For most of the generative period, an implicit assumption has been that we must choose
between a concrete phonological and a diacritic morphological approach to phonological
exceptionality.19 But the argument from learning is that the correct theory is phonological and diacritic, based on lexical phonological indices which are visible to the phonology
but not manipulable by it. The concrete phonological approach, whose pseudo-indices
are manipulable by the phonology, is ill-suited to efficient learning (§9.1). Diacritic approaches are well suited to learning; however the computation of exceptionality is simply not carried out in terms of morphs, rather its currency is lexical phonological elements. This is true for both constraint evaluation (§7.1) and the learning of co-indexation
(§8.2). Concurrently, plausible assumptions about learning ensure that a diacritic phonological account does not suffer from overgeneration (§5.3), and reveal the need for a morphological analytic bias, operationalized here as the Morphological Coherence Principle
(§6). Finally, a morph-based diacritic theory appears empirically insufficient in the inevitably small number of cases that are informative (§9.2). No doubt there is much more
to be said on the topic of exceptionality, but I hope to have established that the nature
of exceptionality is, in essence, phonological and diacritic.
Abbreviations conform with the Leipzig glossing rules; in addition are: lest ‘lest’ and
set ‘inclusion/one of a group’ (Dixon 1977a).
Except, trivially, in purely abstract theories (e.g. Lamb 1966, Fudge 1967).
4 Phonological exceptionality is localized to phonological elements
Albright, Adam. 2002. Islands of reliability for regular morphology: evidence from Italian.
Language 78(4). 684–709.
Albright, Adam & Bruce Hayes. 2002. Modeling English past tense intuitions with minimal generalization. In Proceedings of the ACL-02 workshop on Morphological and phonological learning-Volume 6, 58–69. Association for Computational Linguistics.
Anderson, Stephen R. 1974. The organization of phonology. New York: Academic Press.
Anderson, Stephen R. 1982. Where’s morphology? Linguistic Inquiry 13(4). 571–612.
Anderson, Stephen R. 1992. A-morphous morphology. Cambridge: Cambridge University
Anderson, Stephen R. 2008. Phonologically conditioned allomorphy in the morphology
of Surmiran (Rumantsch). Word Structure 1(2). 109–134.
Anderson, Stephen R. 2015. Morphological change. In Claire Bowern & Bethwyn Evans
(eds.), The Routledge handbook of historical linguistics, 264–285. New York: Routledge.
Anderson, Stephen R. 2016. The role of morphology in transformational grammar. In
Andrew Hippisley & Gregory T. Stump (eds.), The Cambridge handbook of morphology,
587–608. Cambridge: Cambridge University Press.
Anderson, Stephen R. 2017. Words and paradigms: Peter H. Matthews and the development of morphological theory. Transactions of the Philological Society 115. 1–13.
Angluin, Dana. 1980. Inductive inference of formal languages from positive data. Information and control 45(2). 117–135.
Baker, Carl L. 1979. Syntactic theory and the projection problem. Linguistic Inquiry 10(4).
Becker, Michael. 2009. Phonological trends in the lexicon: The role of constraints. Amherst,
MA: University of Massachusetts Ph.D. dissertation.
Becker, Michael, Nihan Ketrez & Andrew Nevins. 2011. The surfeit of the stimulus: Analytic biases filter lexical statistics in Turkish laryngeal alternations. Language 87(1).
Benua, Laura. 1997. Transderivational identity: phonological relations between words. University of Massachusetts, Amherst PhD thesis.
Bermúdez-Otero, Ricardo. 2016. Stratal Phonology. In S.J. Hannahs & Anna R. K. Bosch
(eds.), The Routledge handbook of phonological theory. Abingdon: Routledge.
Bernhardt, Barbara H. & Joseph P. Stemberger. 1998. Handbook of phonological development from the perspective of constraint-based nonlinear phonology. Academic press.
Bloomfield, Leonard. 1939. Menomini morphophonemics. Travaux du Cercle Linguistique
de Prague 8. 105–115.
Bonet, Eulàlia, Maria-Rosa Lloret & Joan Mascaró. 2007. Allomorph selection and lexical
preferences: Two case studies. Lingua 117(6). 903–927.
Bowern, Claire, Erich R. Round & Barry J. Alpher. in revision. The Phonetics and Phonology of Yidiny Stress.
Chomsky, Noam. 1964. Current issues in linguistic theory. Mouton: The Hague.
Erich R. Round
Chomsky, Noam & Morris Halle. 1968. The sound pattern of English. New York: Harper &
Clements, George N. & Engin Sezer. 1982. Vowel and consonant disharmony in Turkish.
In Harry Van der Hulst & Norval Smith (eds.), The structure of phonological representations, vol. 2, 213–255.
Coats, Herbert S. 1970. Rule environment features in phonology. Papers in Linguistics 2(?).
Coetzee, Andries W. 2009. Learning lexical indexation. Phonology 26(1). 109–145.
Crowhurst, Megan & Mark Hewitt. 1995. Prosodic overlay and headless feet in Yidiny.
Phonology 12(1). 39–84.
Dixon, Robert M. W. 1977a. A grammar of Yidiny. Cambridge: Cambridge University
Dixon, Robert M. W. 1977b. Some Phonological Rules in Yidiny. Linguistic Inquiry 8(1).
Dixon, Robert M. W. 1991. Words of our country: stories, place names and vocabulary in
Yidiny, the Aboriginal language of the Cairns-Yarrabah region. Brisbane: University of
Queensland Press.
Ernestus, Mirjam & R. Harald Baayen. 2003. Predicting the unpredictable: interpreting
neutralized segments in dutch. Language 79(1). 5–38.
Finley, Sara. 2010. Exceptions in vowel harmony are local. Lingua 120(6). 1549–1566.
Flack, Kathryn. 2007a. Templatic morphology and indexed markedness constraints. Linguistic Inquiry 38(4). 749–758.
Flack, Kathryn G. 2007b. The sources of phonological markedness. Amherst, MA: University of Massachusetts Ph.D. dissertation.
Fudge, Eric C. 1967. The nature of phonological primes. Journal of Linguistics 3(1). 1–36.
Fukazawa, Haruka. 1999. Theoretical implications of OCP effects on features in Optimality Theory. University of Massachusetts, Amherst Ph.D. Dissertation. https://rucore.
Gouskova, Maria. 2007. The reduplicative template in Tonkawa. Phonology 24(3). 367–
Gouskova, Maria. 2012. Unexceptional segments. Natural Language & Linguistic Theory
30(1). 79–133.
Hall, Nancy. 2001. Max-Position drives iterative footing. In Karine Megerdoomian &
Leora A. Bar-el (eds.), Proceedings of the 20th West Coast Conference on Formal Linguistics. Cascadilla Press.
Halle, Morris & William J. Idsardi. 1995. General properties of stress and metrical structure. In John Goldsmith (ed.), The handbook of phonological theory, 403–443. Oxford:
Blackwell Publishing.
Hammond, Michael. 2000. There is no lexicon. Coyote Papers 10. 55–77.
Hayes, Bruce. 1982. Metrical structure as the organizing principle of Yidiny phonology.
In Harry Van der Hulst & Norval Smith (eds.), The Structure of Phonological Representations, Part I, 97–110. Dordrecht: Foris.
Hayes, Bruce. 1985. A metrical theory of stress rules. New York: Garland.
4 Phonological exceptionality is localized to phonological elements
Hayes, Bruce. 1997. Anticorrespondence in Yidiɲ. Ms., University of California, Los Angeles.
Hayes, Bruce. 2014. Comparative phonotactics. In. Second International Workshop on
Phonotactics. Pisa: Dept. of Linguistics Scuola Normale Superiore.
Hayes, Bruce & Colin Wilson. 2008. A maximum entropy model of phonotactics and
phonotactic learning. Linguistic Inquiry 39(3). 379–440.
Hulst, Harry van der & Jeroen van de Weyer. 1991. Topics in Turkish Phonology. In R.
Boeschoten & L. Verhoeven (eds.), Turkish Linguistics Today, 11–159. Leiden: Brill.
Hyde, Brett. 2012. The odd-parity input problem in metrical stress theory. Phonology
29(3). 383–431.
Inkelas, Sharon. 1994. The consequences of optimization for underspecification. In Jill
Beckman (ed.), Proceedings of NELS 25, 287–302. Amherst: GLSA.
Inkelas, Sharon & Cheryl Zoll. 2007. Is grammar dependence real? A comparison between cophonological and indexed constraint approaches to morphologically conditioned phonology. Linguistics 45(1). 133–171.
Itô, Junko & Armin Mester. 1999. The phonological lexicon. In Natsuko Tsujimura (ed.),
The handbook of Japanese linguistics, 62–100. Oxford: Blackwell Publishing.
Jurgec, Peter. 2010. Disjunctive lexical stratification. Linguistic Inquiry 41(1). 149–161.
Kager, René. 1993. Alternatives to the iambic-trochaic law. Natural Language & Linguistic
Theory 11(3). 381–432.
Kager, René. 1996. On affix allomorphy and syllable counting. In Ursula Kleinhenz (ed.),
Interfaces in phonology (Studia Grammatica 41), 155–171. Berlin: Akademie-Verlag.
Kiparsky, Paul. 1973. Abstractness, opacity and global rules. Bloomington: Indiana University Linguistics Club.
Kiparsky, Paul. 1982a. From cyclic phonology to lexical phonology. In Harry van der
Hulst & Norval Smith (eds.), The structure of phonological representations, 131–175. Dordrecht: Foris.
Kiparsky, Paul. 1982b. From cyclic phonology to lexical phonology. In Harry Van der
Hulst & Norval Smith (eds.), The Structure of Phonological Representations, Part I, 131–
175. Dordrecht: Foris.
Kiparsky, Paul. 2000. Opacity and cyclicity. The Linguistic Review 17(2–4). 351–366.
Labrune, Laurence. 2012. The Phonology of Japanese. Oxford: Oxford University Press.
Lakoff, George. 1970. Irregularity in syntax. New York: Holt, Rinehart & Winston.
Lamb, Sydney M. 1966. Prolegomena to a theory of phonology. Language 42(2). 536–573.
Lombardi, Linda. 2002. Coronal epenthesis and markedness. Phonology 19(02). 219–251.
Łubowicz, Anna. 2005. Locality of conjunction. In John Alderete, Chung-hye Han &
Alexei Kochetov (eds.), Proceedings of the 24th West Coast Conference on Formal Linguistics, 254–262.
Mahanta, Shakuntala. 2008. Directionality and locality in vowel harmony: With special
reference to vowel harmony in Assamese. Netherlands Graduate School of Linguistics
Ph.D. Dissertation.
Mascaró, Joan. 1996. External allomorphy and contractions in Romance. Probus 8(2). 181–
Erich R. Round
McCarthy, John & Alan Prince. 1993a. Generalized alignment. In Geert Booij & Jaap van
Marle (eds.), The yearbook of morphology 1993, 79–153. Dordrecht: Kluwer Academic
McCarthy, John & Alan Prince. 1993b. Prosodic morphology i: constraint interaction and
satisfaction. Ms., University of Massachusetts, Amherst.
McCarthy, John & Alan Prince. 1995. Faithfulness and reduplicative identity. In University
of Massachusetts occasional papers 18: Papers in Optimality Theory, 249–384. Amherst:
Graduate Linguistic Student Association, UMass.
McCarthy, John J. 2003. Comparative markedness. Theoretical linguistics 29(1-2). 1–51.
Merchant, Nazarré Nathaniel. 2008. Discovering underlying forms: Contrast pairs and
ranking. New Brunswick, NJ: Rutgers University PhD thesis.
Mester, Armin. 1994. The quantitative trochee in Latin. Natural Language and Linguistic
Theory 12(1). 1–61.
Moreton, Elliott. 2008. Analytic bias and phonological typology. Phonology 25(1). 83–127.
Moreton, Elliott & Shigeaki Amano. 1999. Phonotactics in the perception of Japanese
vowel length: evidence for long-distance dependencies. In Proceedings of the 6th
European Conference on Speech Communication and Technology, 82. https : / / pdfs .
Myers, Scott. 1997. OCP effects in Optimality Theory. Natural Language and Linguistic
Theory 15(4). 847–892.
Orgun, Cemil. 1996. Sign-based morphology and phonology with special attention to Optimality Theory. Berkeley, CA: University of California, Berkeley PhD thesis.
Pater, Joe. 2000. Non-uniformity in English secondary stress: the role of ranked and lexically specific constraints. Phonology 17(2). 237–274.
Pater, Joe. 2006. The locus of exceptionality: morpheme-specific phonology as constraint
indexation. In Leah Bateman & Adam Werle (eds.), University of Massachusetts Occasional Papers 32: Papers in Optimality Theory, 1–36. Amherst, MA: GLSA.
Pater, Joe. 2009. Morpheme-specific phonology: Constraint indexation and inconsistency
resolution. In Steve Parker (ed.), Phonological Argumentation: Essays on Evidence and
Motivation, 123–154.
Poplack, Shana, Alicia Pousada & David Sankoff. 1982. Competing influences on gender assignment: Variable process, stable outcome. Lingua 57(1). 1–28. http : / / www ., accessed 2016-10-02.
Postal, Paul. 1968. Aspects of phonological theory. New York: Harper & Row.
Prince, Alan & Paul Smolensky. 2004. Optimality Theory: Constraint interaction in generative grammar. Malden, MA: Wiley-Blackwell.
Prince, Alan & Bruce Tesar. 2004. Learning phonotactic distributions. In René Kager,
Joe Pater & Wim Zonneveld (eds.), Constraints in phonological acquisition, 245–291.
Cambridge: Cambridge University Press.
Pruitt, Kathryn. 2010. Serialism and locality in constraint-based metrical parsing. Phonology 27(3). 481–526.
Round, Erich R. 2013. Kayardild Morphology and Syntax. English. Oxford: Oxford University Press.
4 Phonological exceptionality is localized to phonological elements
Round, Erich R. in progress. Unsyllabified moras and length in Yidiny.
Russell, Kevin. 1995. Morphemes and candidates in Optimality Theory.
Smith, Jennifer L. 2004. Making constraints positional: toward a compositional model of
CON. Lingua 114(12). 1433–1464.
Tesar, Bruce & Paul Smolensky. 2000. Learnability in Optimality Theory. Cambridge: MIT
Tesar, Bruce B. 1995. Computational optimality theory. Boulder: University of Colorado
Ph.D. dissertation.
Tesar, Bruce B. 2007. Learnability. In Paul de Lacy (ed.), The Cambridge handbook of phonology, 555–574. Cambridge: Cambridge University Press.
Tranel, Bernard. 1996a. Exceptionality in Optimality Theory and final consonants in
French. In Karen Zagona (ed.), Grammatical Theory and Romance Languages, 275–293.
Amsterdam: John Benjamins.
Tranel, Bernard. 1996b. French liaison and elision revisited: A unified account within
Optimality Theory. In Claudia Parodi, Carlos Quicoli & Mario Saltarelli (eds.), Aspects
of Romance linguistics, 433–455. Washington, DC: Georgetown University Press.
Van Oostendorp, Marc. 2007. Derived environment effects and consistency of exponence.
In Sylvia Blaho, Patrick Bye & Martin Krämer (eds.), Freedom of analysis?, 123–148.
Berlin/New York: Mouton de Gruyter.
Wolf, Matthew. 2015. Lexical insertion occurs in the phonological component. In Eulàlia
Bonet, Maria-Rosa Lloret & Joan Mascaró (eds.), Understanding Allomorphy: Perspectives from Optimality Theory, 361–407. London: Equinox Publishing.
Zoll, Cheryl. 2001. Constraints and representation in subsegmental phonology. In Linda
Lombardi (ed.), Segmental phonology in Optimality Theory: Constraints and representations, 46–78.
Zoll, Cheryl C. 1996. Parsing below the segment in a constraint based framework. Berkeley,
CA: University of California, Berkeley Ph.D. dissertation.
Zonneveld, Wim. 1978. A formal theory of exceptions in generative phonology. Lisse: Peter
de Ridder.
Chapter 5
U-umlaut in Icelandic and Faroese:
Survival and death
Höskuldur Thráinsson
University of Iceland
Although Icelandic and Faroese are closely related and very similar in many respects, their
vowel systems are quite different (see e.g. Anderson 1969b; Árnason 2011). This paper compares u-umlaut alternations in Icelandic and Faroese and shows that the Faroese umlaut has
a number of properties that are to be expected if the relevant alternations are morphological
(or analogical) rather than being due to a synchronic phonological process. In Icelandic, on
the other hand, u-umlaut has none of these properties and arguably behaves like a living
phonological process. This is theoretically interesting because the quality of the vowels involved (both the umlaut trigger and the target) has changed from Old to Modern Icelandic.
In addition, u-umlaut in Modern Icelandic is more opaque (in the sense of Kiparsky 1973)
than its Old Icelandic counterpart, i.e. it has more surface exceptions. An epenthesis rule inserting a (non-umlauting) /u/ into certain inflectional endings is the cause of many of these
surface exceptions. Yet it seems that u-umlaut in Icelandic is still transparent enough to
be acquired by children as a phonological process. In Faroese, on the other hand, u-umlaut
became too opaque and died out as a phonological rule. It is argued that this has partly
to do with certain changes in the Faroese vowel system and partly with the fact that the
u-epenthesis rule was lost in Faroese.
1 Introduction
Anderson put the process of u-umlaut in Icelandic on the modern linguistic map with
the analysis he proposed in his dissertation (Anderson 1969b) and several subsequent
publications (Anderson 1969a; 1972; 1973; 1974; 1976). Because of changes in the vowel
system from Old to Modern Icelandic, the nature of the umlaut process changed somewhat through the ages (see e.g. Benediktsson 1959). The most important part of u-umlaut,
and the only part that is alive in the modern language, involves /a/ ~ /ǫ/ alternations in
the old language (phonetically [a] ~ [ɔ], as shown in 2), which show up as /a/ ~ /ö/ alternations in the modern language (phonetically [a] ~ [œ], cf. 2). This is illustrated in (1)
with the relevant vowel symbols highlighted:
Höskuldur Thráinsson. 2017. U-umlaut in Icelandic and Faroese: Survival and death.
In Claire Bowern, Laurence Horn & Raffaella Zanuttini (eds.), On looking into words
(and beyond), 99–113. Berlin: Language Science Press. DOI:10.5281/zenodo.495440
Höskuldur Thráinsson
Old Icelandic:
saga ʽsagaʼ, obl sǫgu, pl sǫgur
hvass ‘sharp’, dat hvǫssum
tala ‘speak’, tǫlum
Modern Icelandic:
saga, obl sögu, pl sögur
hvass, dat hvössum
tala, 1pl tölum
As these examples suggest, the quality of the root vowel /a/ changes when a /u/ follows
in the next syllable. The relevant proecesses can be illustrated schematically as in (2). For
the sake of simplicity I use conventional orthographic symbols to represent the vowels
and only give IPA-symbols for the vowels that are important for the understanding of
the umlaut processes. The umlaut-triggering vowels are encircled:1
a. u-umlaut in Old Icelandic and the system of short vowels:
[-round] [+round]
[-round] [+round]
u [u]
ǫ [ɔ]
a [a]
b. u-umlaut in Modern Icelandic and the system of monophthongs:2
[-round] [+round]
[-round] [+round]
ú [u]
u [ʏ]
ö [œ]
a [a]
o [ɔ]
The gist of Anderson’s analysis of u-umlaut can then be illustrated semi-formally as
in the traditional generative phonological notation in (3), with the assimilating features
highlighted (see also Rögnvaldsson 1981: 31, Thráinsson 2011: 89–90):3
a. u-umlaut
[ in Old] Icelandic:
/a/ → +round / C0 V [+round]
b. u-umlaut
[ in Modern
/ C0 V [+round]
/a/ →
Note that in the representation of the Modern Icelandic vowel system, the accents over vowel symbols
have nothing to do with quantity but simply denote separate vowel qualities. Thus /í/ is [i], /i/ is [ɪ], /ú/ is
[u] and /u/ is [ʏ], as the schematic representation in (2) suggests.
I am assuming here, like Thráinsson (1994) and Gíslason & Thráinsson (2000: 34), for instance, that Modern Icelandic only distinguishes between three three vowel heights and that /e/ [ɛ] and /ö/ [œ] are both
phonologically [+low], like /a/ [a] and /o/ [ɔ]. For different assumptions see e.g. Árnason (2011: 60).
Here, and elsewhere in this paper, I will use the kinds of formulations of rules and conditions familiar
from classical generative phonology since much of the work on u-umlaut has been done in that kind of
framework. For analyses employing more recent frameworks see Gibson & Ringen 2000, Hansson 2013
and Ingason 2016. Most of the argumentation in this paper should be relatively framework-independent,
5 U-umlaut in Icelandic and Faroese: Survival and death
As the illustration in (3) shows, the modern version of the umlaut is somewhat more
complex than the old one, assimilating two features rather than one. Nevertheless, it is
still arguably a phonologically (or phonetically) natural assimilation process, assimilating rounding and backness.
Although the u-umlaut discussion was most lively on the international scene in the
1970s (see e.g. Iverson 1978; Iverson & Anderson 1976; Orešnik 1975; 1977, cf. also Valfells
1967), the topic keeps popping up to this day, e.g. in journals and conferences dedicated
to Scandinavian linguistics (see e.g. Gibson & Ringen 2000; Indriðason 2010; Thráinsson 2011; Hansson 2013) and even in recent master’s theses and doctoral dissertations
(see Markússon 2012; Ingason 2016). The main reason is that while u-umlaut in Modern
Icelandic is obviously very productive, being applied consistently to new words and loanwords, it shows a number of intriguing surface exceptions. These have been discussed
extensively in the literature cited but here I will concentrate on the most common and
widespread one, namely the lack of umlaut before a /u/ that has been inserted between
the inflectional ending /r/ and a preceding consonant. This epenthesis did not exist in
Old Icelandic as illustrated in (4):
Old Icelandic:
dalr ‘valley’, latr ‘lazy’
Modern Icelandic:
dalur, latur
If u-umlaut is a phonological rule in the modern language, this u-epenthesis has to follow it, as it did historically. This is one of the properties of u-umlaut that have been used
to argue for the necessity of relatively abstract phonological representations and derivations (e.g. Anderson 1969b; 1974; Rögnvaldsson 1981; Thráinsson 2011; Hansson 2013)
while others have maintained that u-umlaut is not a phonological process anymore in
Modern Icelandic and the relevant alternations are morphologized and purely analogical (see e.g. Markússon 2012) or at least “morpheme-specific”, i.e. triggered by particular
morphemes that may or may not contain a /u/ (Ingason 2016).4
In this paper I will compare u-umlaut alternations in Modern Icelandic and Modern
Faroese. This comparison will show very clearly that u-umlaut in Modern Faroese has a
Ingason (2016: 220) formulates his umlaut rule as follows:
Realize an underlying /a/ as /ö/ in the syllable which precedes the morpheme which triggers the umlaut.
As can be seen here, no mention is made of a triggering /u/ in the rule. The reason is that Ingason wants
to derive all all paradigmatic /a/ ~ /ö/ alternations the same way, including the ones where /u/ has been
syncopated historically. Thus he argues that the morpheme -ø in feminine nouns like sök ‘guilt,
case’ and the nom./ morpheme -ø in neuter nouns like börn ‘children’ triggers umlaut the same way
that the morpheme -um does in sökum and börnum. But many researchers have wanted to distinguish between morphologically conditioned umlaut, where there is no triggering /u/, and phonologically
conditioned umlaut triggered by /u/, e.g. Rögnvaldsson (1981). One reason for doing so comes from the
behavior of loanwords like the adjective smart ‘smart, chic’. Here the and the nom/ can
either be smart or smört, i.e. a morphologically conditioned umlaut may or may not apply. But once an
umlauting inflectional ending containing /u/ is added to the loanword smart, the u-umlaut becomes obligatory. Thus can only be smört-um and not *smart-um and the form has to be smört-u and
not *smart-u. This suggests that the morphologically conditioned umlaut is more prone to exceptions than
the phonologically conditioned one, which is actually to be expected. Thanks to Eiríkur Rögnvaldsson for
pointing this out to me.
Höskuldur Thráinsson
number of properties (e.g. paradigm levelling, various kinds of exceptions, total absence
from certain paradigms, inapplicability to loanwords …) that are to be expected if the
relevant alternations are no longer due to a synchronic process. In Modern Icelandic, on
the other hand, u-umlaut has none of these properties and behaves more like a phonological rule. This is of general theoretical interest since it illustrates how phonological
rules can survive (in the case of Icelandic) despite reduced transparency (in the sense
of Kiparsky 1973) and how changes in the phonological system can cause the death of a
phonological rule (in the case of Faroese) and what the consequences can be.
The remainder of the paper is organized as follows: In §2 I first illustrate how the
u-epenthesis works in Modern Icelandic and then present a couple of arguments for the
phonological (as opposed to morphological) nature of Modern Icelandic u-umlaut. §⁇
first describes some facts about the Faroese vowel system that must have been important for the development of u-umlaut and then shows that u-epenthesis does not exist
anymore as a phonological process in Modern Faroese. It is then argued that these developments led to the death of u-umlaut as a phonological process in Faroese. §⁇ then
contains a systematic comparison of u-umlaut alternations in Modern Icelandic and Faroese, concluding that the Faroese ones must be analogical (and morphological) in nature
as they do not exhibit any of the crucial phonological properties that Modern Icelandic
u-umlaut alternations show. In Icelandic, on the other hand, u-umlaut does not show the
non-phonological properties listed for its Faroese counterpart. §⁇ concludes the paper.
2 u-epenthesis and u-umlaut in Modern Icelandic
2.1 The epenthesis rule
The phoneme /r/ frequently occurs in Old Icelandic (Old Norse) as a marker of various
morphological categories, including of strong masculine nouns and adjectives
as illustrated in (5). It sometimes assimilated to a preceding consonant, e.g. /s, l, n/ (cf.
5c),5 but it was deleted after certain consonant clusters, such as /gl, gn, ss/ (cf. 5d):
stór-r ‘big’, mó-r ‘peat’, há-r ‘high’
dal-r ‘valley’, lat-r ‘lazy’, tóm-r ‘empty’, harð-r ‘hard’
ís-s ‘ice’, laus-s ‘loose’, stól-l ‘chair’, fín-n ‘fine’
fugl ‘bird’, vagn ‘wagon’, foss ‘waterfall’ (stem foss-)
It is likely that the /r/ in words of type (5b) was syllabic in Old Icelandic. There are
no syllabic consonants in Modern Icelandic, on the other hand. Instead a /u/ appears
between the /r/ and the preceding consonant in the modern version of words of type
(5b). There is historical evidence for this u-insertion from the late thirteenth century and
Assimilation to stem-final /l, n/ only happened in Old Icelandic if these consonants were preceded by
long vowels, i.e. Old Icelandic diphthongs and vowels that are standardly represented by accented vowel
symbols in Old Icelandic orthography, cf. stól-l ‘chair’ vs. dal-r ‘valley’, fín-n ‘fine’ vs. lin-r ‘soft, limp’,
heil-l ‘whole’ vs. hol-r ‘hollow’.
5 U-umlaut in Icelandic and Faroese: Survival and death
onwards (see e.g. Kristinsson 1992 and references cited there) and many linguists have
argued that u-epenthesis is still a productive phonological process in Modern Icelandic
(e.g. Anderson 1969b,a; Orešnik 1972; Rögnvaldsson 1981; Kiparsky 1984).6 This implies
that speakers distinguish between a -ur-ending where the underlying morpheme is #-r#
and the /u/ is epenthetic (and does not trigger u-umlaut) and a -ur-ending where the /u/
is not epenthetic and the underlying morpheme is #-ur# (and the /u/ triggers u-umlaut).
This contrast is illustrated in (⁇) vs. (⁇) (see also the examples in 1 and 4 above):
#dal+r# ‘valley’
#lat+r# ‘lazy’
#sag+ur# ‘sagas’
#tal+ur# ‘numbers’
dal-ur by epenthesis no umlaut
lat-ur by epenthesis no umlaut
Thus the ending #-r#, which is both found in strong masculine nouns like dalur
‘valley’ and in the strong masculine form of adjectives like latur ‘lazy’, does not have the
same properties as the ending #-ur# which is found in feminine nouns like sögur
‘sagas’ and tölur ‘numbers’. Despite their surface similarities in certain environments,
speakers can clearly distinguish these endings. A part of the reason must be that the ending #-r# only shows up as -ur in phonologically definable environments,
i.e. the modern version of words with stems of type (5b), whereas the ending #ur# is not so restricted and always shows up as -ur. This is illustrated in Table 1 (compare
the examples in 5).
Comparison of Table 1 and the Old Icelandic examples in (5) reveals a slight extension
of r-deletion: The /r/ of the morphological ending #-r# is now deleted after /r/ (compare
line d of the table to 5a) and after all instances of /s/, not just /ss/ (compare line d of the
table to (5c,d)). The u-epenthesis illustrated in line b of Table 1 is an innovation, of course.
Otherwise the ending behaves in much the same way as in Old Icelandic. The
different behavior of the morphemes compared in Table 1 can be seen as an argument
for distinguishing them in the underlying form, e.g. for not analyzing the ending
as #-ur#.
2.2 Some phonological properties of Modern Icelandic u-umlaut
In this section I will mention two sets of facts which show that u-umlaut still has certain
properties in Modern Icelandic that are to be expected if it is a phonologically conditioned process.
Orešnik later maintained that u-epenthesis could not be a synchronic rule in Modern Icelandic because
of the existence of exceptional word forms like klifr ‘climbing’ (from the verb klifra ‘climb’), sötr ‘slurping’ (from the verb sötra ‘slurp’), pukr ‘secretiveness’ from the verb pukra(st) ‘be secretive about’, etc.
(Orešnik 1978; see also the discussion in Kjartansson 1984). In words of this kind one would have expected
u-epenthesis to apply. The importance of these exceptions is not very clear since this is a very special class
of words (all derived from verbs ending in -ra) and it is typically possible or even preferred to apply the
epenthesis rule to these forms, giving klifur, sötur, pukur, etc. For the sake of completeness it should be
noted that the final -r in word forms like sötr, pukr has to be voiceless and this may be related to the fact
that there are no syllabic consonants in Modern Icelandic, as stated above.
Höskuldur Thráinsson
Table 1: Phonological realization of the inflectional endings #-r# and #-ur# in
Modern Icelandic.
type of stem
phonological realization of
the ending #-r#
phonological realization of
the ending #-ur#
ending in a vowel
ending in a single consonant
(but see c)
ending in a high
vowel + /l,n/
(mó-r ‘peat’, há-r ‘high’)
(dal-ur ‘valley’, lat-ur ʽlazyʼ)
(ló-ur ‘golden plovers’)
(sög-ur ʽsagasʼ, töl-ur
(spús-ur ‘wives’, súl-ur
ʽcolumnsʼ, dýn-ur
(ýs-ur ʽhaddocksʼ, aus-ur
‘scoops’, hór-ur ‘whores’,
ugl-ur ‘owls’, hrygn-ur
ʽspawning fish’, byss-ur
ending in /s, r/ or
clusters ending
in /l, n/ such as
/gl, gn/
(stól-l ‘chair’, fín-n ‘fine’)
(ís ‘ice’, laus ‘loose’, foss
‘waterfall’bjór ‘beer’, stór
‘big’, fugl ‘bird’, vagn
First, if u-umlaut was morphologically conditioned and not phonologically, we would
expect it to be restricted to certain morphological categories or parts of speech. It is
not. It applies in the paradigms of nouns, adjectives and verbs when a /u/ follows in the
inflectional ending (with the exception of the epenthetic /u/ already mentioned). This is
illustrated in (7):
a. saga ‘saga’, sög-u, nom/ sög-ur, sög-um
b. snjall ‘smart’, snjöll-um, snjöll-u
c. kalla ‘call’, köll-um, kölluð-u
The so-called i-umlaut is very different in this respect. It is clearly not alive as a phonological rule anymore but its effects can still be observed in the modern language in certain
morphologically definable environments. As a result we can find near-minimal pairs of
word forms where i-umlaut has applied in one member but not the other although the
phonological conditions seem identical. Some examples are given in (8):
a. háttur ‘mode’, hætt-i/*hátt-i, hætt-ir/*hátt-ir
b. sáttur ‘satisfied’, *sætt-i/sátt-i, *sætt-ir/sátt-ir
In (8a) we see examples of the paradigmatic alternation /á ~ æ/ (phonetically [au] ~ [ai] in
the modern language, probably [aː] ~ [ɛː] in Old Icelandic) originally caused by i-umlaut.
5 U-umlaut in Icelandic and Faroese: Survival and death
In the we have /á/ in the stem but in the the only acceptable form is hætti
and the “non-umlauted” version *hátti is unacceptable. Similarly, in the only
hættir is acceptable and *háttir is not. At a first glance we might think that an /i/ in
the inflectional ending is still causing this “umlaut” but a comparison with the adjectival forms in (8b) indicates that this cannot be the case. Here the only acceptable weak form is sátti and not *sætti and the only form is sáttir and not *sættir.
So the i-umlaut alternations in Modern Icelandic are clearly morphologically conditioned
and not phonological anymore (see also Thráinsson 2011: 93 for further examples of this
Second, recall that standard generative phonology formulations of u-umlaut in Icelandic of the kind illustrated in (3b) above state explicitly that /u/ only triggers umlaut
of /a/ in the immediately preceding syllable. This is illustrated by examples like the following:
a. bakki ‘bank’ bökk-um/*bakk-um
b. akkeri ‘anchor’ *ökker-um/akker-um
In (9a) the u-umlaut obligatorily applies to the root vowel /a/ in the immediately preceding syllable. In (9b), on the other hand, the /u/ in the (same) inflectional ending cannot
apply to the root vowel /a/ because there is a syllable intervening. An interesting and
much discussed case, e.g. by Anderson in several of the publications cited above, involves trisyllabic words with two instances of /a/ in the stem. Consider the examples in
a. kalla ʽcallʼ kalla-ð-i, *kallö-ð-um/köllu-ð-um/*kallu-ð- um/*kölla-ð-um
b. banan-i ‘banana’ banön-um/bönun-um/*banun-um/*bönan-um
Consider first the conceivable forms of the verb kalla ‘call’. Based on the formulation (3b) of the u-umlaut rule, one might have expected the form *kallöðum, where the
/u/ in the inflectional ending triggers u-umlaut of the /a/ in the preceding syllable. This is
not an acceptable form, however. The reason is that in forms of this sort a “weakening”
of unstressed /ö/ to /u/ is obligatory. This weakening is found in in many words, e.g. the
plural of the word hérað ‘district’, plural héröð or (preferred) héruð, meðal ‘medicine’,
plural meðöl or (preferred) meðul. It is not always obligatory but it seems that in the
past tense of verbs of this sort it is. But once the (umlauted) /ö/ in *kallöðum has been
weakened to /u/ it obligatorily triggers u-umlaut of the preceding /a/ so kölluðum is acceptable but *kalluðum is not. Finally, the form *köllaðum is not acceptable either, since
there u-umlaut would be applied across an intervening syllable, which is not possible,
as we have seen (cf. 9b). The u-umlaut works in a similar fashion in the word banani,
except that here the weakening of the second (and unstressed) syllable from /ö/ to /u/ is
not obligatory. Hence banönum is an acceptable form, with the /u/ in the final syllable
triggering u-umlaut of the preceding /a/ to /ö/. But if this /ö/ is further weakened to
Höskuldur Thráinsson
/u/, then u-umlaut of the first /a/ is obligatory and bönunum is an acceptable form but
*banunum is not.7 As predicted by the formulation of the u-umlaut rule in (3b) a form
like *bönanum is unacceptable because there the u-umlaut would have applied across an
intervening syllable. Facts of this sort have been interpreted as showing that u-umlaut
in Modern Icelandic is of a phonological nature since it depends on syllabic structure
(no syllables can intervene between the umlaut trigger and the target) and it can be applied iteratively (a /u/ which itself is derived by u-umlaut and subsequent independently
needed weakening can trigger u-umlaut).
3 The conditions for u-umlaut in Modern Faroese
3.1 u-umlaut and the Modern Faroese vowel system
Modern Faroese has preserved some u-umlaut-like vowel alternations. A couple of examples are given in (11) (see also Thráinsson et al. 2012: 78, 100, passim):
dag-ur ‘day’, døg-um; spak-ur ʽcalmʼ, spøk-u
At first glance, these alternations seem very similar to the Icelandic ones described in
the preceding sections. But while the u-umlaut alternations are arguably phonologically
(or phonetically) natural in Modern Icelandic (see the diagram in 2b and the formulation
in 3b), it will be claimed below that this is not the case in Faroese. To demonstrate this,
it is necessary to look closely at the Faroese vowel system. Consider first the following schematic representation of Faroese u-umlaut of the type just illustrated, where the
alleged umlaut trigger is encircled (cf. Thráinsson 2011: 98, Thráinsson et al. 2012: 33,
compare Árnason 2011: 248–250):8
u-umlaut in Modern Faroese and the system of monophthongs:
u [uː/ʊ]
ø [øː/œ]
æ [ɛaː/a]
Something like (13) would seem to be a possible formulation of a process of this kind
in traditional generative phonology terms (compare 3b):
It is sometimes claimed that bönönum is also an acceptable form for some speakers. If this is so, it is possible
that the /ö/ in the next-to-last syllable triggers u-umlaut (i.e. ö-umlaut!) of the /a/ in the first syllable.
That would simply mean that the feature [−low] in the definition of the environment of the u-umlaut in
(3b) would be omitted. But since there are no derivational (nor inflectional) morphemes containing an
underlying /ö/ (i.e. an /ö/ that cannot have been derived by u-umlaut), this proposal cannot be tested
independently of the iterative rule application, as pointed out by a reviewer.
8 Vowel length is predictable in Faroese, as it is in Icelandic: Vowels are long in stressed open syllables,
otherwise short. As illustrated in the brackets in (12), there is often a considerable difference in the phonetic
realization of the long and short variants. This will be illustrated below. — It should be noted that Árnason
(2011: 76) assumes a different analysis of Faroese monophthongs.
5 U-umlaut in Icelandic and Faroese: Survival and death
(13) Possible[ phonological
formulation of u-umlaut in Modern Faroese:
/æ/ →
/ C0 V [+round]
Presented this way, u-umlaut in Faroese looks like a fairly natural assimilation rule at
a first glance.9 But the facts are somewhat more complicated.
First, the alleged trigger /u/ is not too stable in Modern Faroese. The reason is that
unstressed /i,u/ are not distinguished in all Faroese dialects. In some dialects they merge
into an [ɪ]-like sound, in others into an [ʊ]-like sound but some dialects distinguish them
as [ɪ] and [ʊ] (see Thráinsson et al. 2012: 27, and references cited there). This situation has
clearly added to the phonological opacity of u-umlaut alternations for speakers acquiring
Second, the target of the u-umlaut in Faroese is arguably a “moving” one. As indicated
in (12), the umlaut affects the phoneme represented there as /æ/. As the orthography
suggests, it is a descendant of Old Norse /a/ in words like dagur, spakur (see 11). It is
realized phonetically as [ɛaː] when long and [a] when short, as shown in (12), cf. spakur
[spɛaːh kʊɹ] ‘calm’, sg.n spakt [spakt] (see Thráinsson et al. 2012: 34 passim). But in the
history of Faroese Old Norse /a/ [a] and /æ/ [ɛː] merged so the phoneme represented here
as /æ/ can also be a descendant of Old Norse /æ/ and then it is represented in the spelling
as ʽæʼ, cf. trælur [th ɹɛaːlʊɹ] ‘slave’, æða [ɛaːva] ‘eider duck’. Words written with ʽæʼ show
the same alternation between long [ɛaː] and short [a] as demonstrated for spakur and
spakt above (e.g. vænur [vɛaːnʊɹ] ‘beautiful’ sg.m vs. vænt [van̥ t], cf. Thráinsson et al.
2012, p. 34). Yet it seems that u-umlaut is rarely if ever found in the ʽæʼ-words. Thus
the of trælur is trælum and not *trølum (compare dølum of dalur ‘valley’)
and although the words æða ‘eider duck’ and aða ‘(big) mussel’ sound the same, i.e. as
[ɛaːva], the of the former has to be æðum [ɛaːvʊn] and øvum [øːvʊn] can only be of aða.10
To further complicate matters, the development of Old Norse /a/ in Faroese has left
“room” for a “regular /a/” in the Faroese vowel system, as shown in the diagram in (12).
It occurs in loanwords and is realized as [aː] when long and [a] when short, cf. Japan
[ˈjaːh pan], japanskur [jaˈph anskʊɹ] ‘Japanese’.11 It does not seem that this vowel ever
undergoes u-umlaut in Faroese (for further discussion see §4).
A reviewer suggests, however, that a process changing rounding and height as formulated for Faroese in
(13), might be less natural from the point of view of acoustic phonetics than a process changing rounding
and backness the way the u-umlaut rule in Modern Icelandic does according to (3): The former affects
both F1 (for the height difference) and F2 (for rounding) whereas the latter affects F2 in opposite directions
(raising it for fronting but lowering it for rounding). Thus the Modern Icelandic u-umlaut rule would
“generate more similar input-output mappings”, which may be preferred to less similar ones.
10 A reviewer points out that the fact that u-umlaut does not apply do ‘æ’-words in Faroese suggests that
“u-umlaut had already taken on a morphological character before /a/ and /æ/ merged.” But since there are
no written records of Faroese from 1400‒1800, the historical development of the language is very murky.
11 In the noun Japan the stress falls on the first syllable, in the adjective japanskur it falls on the second one
as indicated. Hence the quantity alternation in the first vowel.
Höskuldur Thráinsson
Finally, there is no u-epenthesis in Modern Faroese to “explain away” apparent exceptions to u-umlaut as will be shown in the next section.
3.2 The lack of u-epenthesis in Modern Faroese
Now recall that the most obvious surface exception to u-umlaut in Modern Icelandic is
due to the u-epenthesis described above. This rule creates -ur-endings that do not trigger
u-umlaut. It was argued that this epenthesis rule is still productive in Icelandic, witness
the fact that it only applies in phonologically definable environments. Hence there is
a clear distributional difference between -ur-endings produced by the epenthesis rule
(and not triggering u-umlaut) and -ur-endings where the /u/ is a part of the underlying
form (and triggers umlaut). This is not the case in Faroese, where the ending -ur as a
marker of the of strong masculine nouns and adjectives, with a /u/ that was historically inserted by epenthesis, has been generalized to all environments. Hence it has
become distributionally indistinguishable from other -ur-endings. Table 2 compares the
phonological realization of the #-r#-ending in Modern Icelandic to its Modern
Faroese counterpart (see also Thráinsson 2011: 100):
Table 2: Phonological realization of a strong in Modern Icelandic and Modern Faroese.
type of stem
phonological realization of a
strong ending in
Modern Icelandic
phonological realization of a
strong ending in
Modern Faroese
ending in a vowel
(mó-r ‘peat’, há-r ‘high’)
ending in a single consonant
(but see c)
ending in a high
vowel + /l,n/
(dal-ur ‘valley’, lat-ur ʽlazyʼ)
(mó-ur/mógv-ur ‘peat’, há-ur
(dal-ur ‘valley’, lat-ur ʽlazyʼ)
ending in /s, r/ or
clusters like /gl,
(stól-l ‘chair’, fín-n ‘fine’)
(ís ‘ice’, laus ‘loose’, foss
‘waterfall’, stór ‘big’, fugl
‘bird’, vagn ‘wagon’)
(stól-ur ‘chair’, fín-ur ‘fine’)
(ís-ur ‘ice’, leys-ur ‘loose’,
foss-ur ‘waterfall’, stór-ur
‘big’, fugl-ur ‘bird’, vagn-ur
This has clearly made the u-umlaut rule in Faroese more opaque since now the nonumlauting and umlauting ur-endings occur in the same phonological environments. It
seems very likely that this has contributed to the death of u-umlaut as a phonological
process in Faroese.
5 U-umlaut in Icelandic and Faroese: Survival and death
4 Testing the predictions
In the preceding discussions I have described /a ~ ö/ alternations in Modern Icelandic
and their Modern Faroese counterparts. I have argued that the Icelandic alternations are
still governed by a synchronic phonological process. Although these alternations are still
found in Modern Faroese, I have argued that they cannot be governed by a phonological
rule. Instead they must be morphologically governed or analogical. This analysis makes
several testable predictions (see Thráinsson 2011: 100–102).
First, we do not a priori expect phonologically conditioned alternations to be restricted
to particular morphological categories whereas morphologically conditioned alternations obviously are, by definition. As we have already seen, the Icelandic u-umlaut occurs
in the inflectional paradigms of nouns, adjectives and verbs and in various grammatical
categories (cases, numbers, tenses, persons …). Its Faroese counterpart behaves differently. It is found in the inflectional paradigms of nouns and adjectives, as we have seen
(cf. 11), but not in the past tense forms of verbs, where it would be expected on phonological grounds. Thus we have við kölluðum in Icelandic vs. vit kallaðu in Faroese for ‘we called’, and við frömdum vs. vit framdu in Faroese for ‘we did, made’.
Second, a phonological rule should not allow analogical extensions to forms that do
not fit its structural conditions. Such extensions are not found for Icelandic u-umlaut
but in Faroese they are very common. Thus the /ø/ of the oblique cases søgu ‘saga’ has
been analogically extended to the form søga and many other words of a similar
type. The corresponding form *söga is unacceptable in Icelandic.12
Third, a phonologically conditioned rule should apply whenever its structural conditions are met. Thus we would not expect to find inflectional forms in Icelandic where
u-umlaut fails to apply in an appropriate environment. Such examples are very common
in Faroese, on the other hand. Thus the of the noun rakstur ‘shave’ in Faroese
is rakstrum and not the expected *røkstrum, the of spakur ‘calm’ can either be
spøkum or spakum, etc. (see Thráinsson et al. 2012: 79, 100, passim). Corresponding
unumlauted forms are unacceptable in Icelandic.
Fourth, there is evidence for “iterative” application of u-umlaut in Icelandic, with one
application of the u-umlaut rule feeding another. This was discussed above (second
part of §2.2) in connection with forms like kölluðum ‘(we) called’ and
bönunum ‘bananas’. No such evidence is found in Faroese, where the corresponding
forms are kallaðum and bananum.13
As a reviewer reminds me, the Icelandic neologism for computer is interesting in this connection. It was
supposed to be tölva (related to the word tala ‘number’ — this was when computers were mainly used for
computing) in, oblique singular cases tölvu. In Proto-Nordic time /v/ could trigger umlaut of /a/
to /ǫ/ so we have Old Norse words like vǫlva ‘sooth-sayer, witch’. But since /v/ is not a trigger of umlaut
in Modern Icelandic (witness loanwords like salvi ‘salve, cream’), speakers tend to use the form talva for, thus in a way undoing the underlying /ö/ in the nominative as if they are “assuming” that the /ö/
in the oblique cases is derived by a synchronic u-umlaut from /a/, as in words like saga ‘saga’, oblique sögu
(for some discussion see Thráinsson 1982).
13 The latter form may be related to the fact that banan ‘banana’ is a loanword and contains the vowel /a/
(long variant [aː]) and not /æ/, cf. the discussion in §3.1. See also the next paragraph.
Höskuldur Thráinsson
Finally, Icelandic u-umlaut is so productive that it is naturally applied in loanwords,
as we have seen. This is not so in Faroese. Thus the word app (for a small program) has
been adopted into both languages. In Icelandic the has to be öppum whereas the
natural form is appum in Faroese. This can easily be verified by searching for the word
combinations með öppum and við appum ‘with apps’ on Google. For the first variant one
finds a number of Icelandic hits, for the second Faroese ones.
The general conclusion, then, is that u-umlaut in Modern Icelandic has a number of
properties that are to be expected if it is a phonological process but none of the properties
one might expect of morphologically conditioned or analogical alternations.
5 Concluding remarks
While it has often been argued that phonology need not be “natural” (see e.g. Anderson
1981), there must obviously be limits to the “unnaturalness” and opacity of phonological
processes. Once they become too unnatural and opaque, they can no longer be acquired
as such and the phonological alternations originally created by them will be relegated to
morphology. Then their productivity will be limited and it will at best survive to some
extent by analogy, but analogical processes are known to be irregular and unpredictable.
The fate of i-umlaut in Icelandic is a case in point, as described above (see the discussion
of the examples in 8). But whereas we do not have detailed information about how iumlaut died as a phonological process, comparison of the development of u-umlaut in
Icelandic and Faroese sheds an interesting light on how a phonological rule can die and
how it can survive despite changing conditions.
Many thanks to Steve Anderson for introducing me to the wonders of synchronic uumlaut way back when. Thanks are also due to the editors and to two anonymous reviewers for help, useful comments, suggestions and corrections.
Anderson, Stephen R. 1969a. An outline of the phonology of Modern Icelandic vowels.
Foundations of Language 5. 53–72.
Anderson, Stephen R. 1969b. West Scandinavian vowel systems and the ordering of phonological rules. Cambridge, MA: Massachusetts Institute of Technology PhD thesis.
Anderson, Stephen R. 1972. Icelandic u-umlaut and breaking in a generative grammar. In
Evelyn S. Firchow, Karen Grimstad, Wayne A. O’Neil & Nils Hasselmo (eds.), Studies
for Einar Haugen, 13–30. Mouton: The Hague.
Anderson, Stephen R. 1973. U-umlaut and scaldic verse. In Stephen R. Anderson & Paul
Kiparsky (eds.), A Festschrift for Morris Halle, 3–13. New York: Holt, Rinehart & Winston.
5 U-umlaut in Icelandic and Faroese: Survival and death
Anderson, Stephen R. 1974. The organization of phonology. New York: Academic Press.
Anderson, Stephen R. 1976. On the conditioning of icelandic u-umlaut. Language Sciences.
Anderson, Stephen R. 1981. Why phonology isn’t “natural”. Linguistic Inquiry 12(4). 493–
Árnason, Kristján. 2011. The phonology of Icelandic and Faroese. Oxford: Oxford University
Benediktsson, Hreinn. 1959. The vowel system of Icelandic: A survey of its history. Word
15. [Also published in Guðrún Þórhallsdóttir, Höskuldur Þráinsson, Jón G. Friðjónsson
and Kjartan Ottosson (eds.), Linguistic studies, historical and comparative, by Hreinn
Benediktsson, 50–73. Reykjavík, Iceland: Institute of Linguistics, University of Iceland],
Gibson, Courtenay St John & Catherine O. Ringen. 2000. Icelandic umlaut in Optimality
Theory. Nordic Journal of Linguistics 23(1). 49–64.
Gíslason, Indriði & Höskuldur Thráinsson. 2000. Handbók um íslenskan framburð. [‘handbook on Icelandic pronunciation’]. 2nd edn. Reykjavík: Rannsóknarstofnun Kennaraháskóla
Hansson, Gunnar. 2013. Alternations, allomorphy and abstractness: Icelandic u-umlaut revisited. Paper presented at the 25th Scandinavian Conference of Linguistics, Reykjavík.
Indriðason, Þorsteinn G. 2010. Til varnar hljóðkerfisreglu. Nokkrar athugasemdir við umræðugrein. [‘In defence of a phonological rule. Some comments on a discussion article.’] Íslenskt mál 32. 137–149.
Ingason, Anton Karl. 2016. Realizing morphemes in the Icelandic Noun Phrase. Philadelphia: University of Pennsylvania PhD thesis.
Iverson, Gregory. 1978. Synchronic umlaut in Old Icelandic. Nordic Journal of Linguistics
1. 121–139.
Iverson, Gregory & Stephen R. Anderson. 1976. Icelandic u-umlaut: An exchange of views.
Language Sciences. 28–34.
Kiparsky, Paul. 1973. Phonological representations. In Osamu Fujimura (ed.), Three dimensions of linguistic theory, 1–135. Tokyo: TEC.
Kiparsky, Paul. 1984. On the lexical phonology of Icelandic. In Iréne Johansson ClaesChristian Elert & Eva Strangert (eds.), Nordic prosody III, 135–164. Umeå: University of
Kjartansson, Helgi Skúli. 1984. U -innskot, lífs eða liðið? [‘u-epenthesis, alive or dead?’]
Íslenskt mál 6. 182–185.
Kristinsson, Ari Páll. 1992. U -innskot í íslensku. [ʽu-epenthesis in Icelandic.’] Íslenskt mál
14. 15–33.
Markússon, Jón Símon. 2012. Eðli u-hljóðvarpsvíxla í íslenskri málsögu. [‘The nature of uumlaut alternations in the history of Icelandic.’] Reykjavík: University of Iceland Master’s thesis.
Orešnik, Janez. 1972. On the epenthesis rule in Modern Icelandic. Arkiv för nordisk filologi
87. 1–32.
Höskuldur Thráinsson
Orešnik, Janez. 1975. The Modern Icelandic u-umlaut rule. In Karl-Hampus Dahlstedt
(ed.), The Nordic languages and modern linguistics 3, 621–633. Almqvist & Wiksell,
Orešnik, Janez. 1977. Modern Icelandic u-umlaut from the descriptive point of view.
Gripla 2. 151–182.
Orešnik, Janez. 1978. The Modern Icelandic epenthesis rule revisited. Arkiv för nordisk
filologi 93. 166–173.
Rögnvaldsson, Eiríkur. 1981. U -hljóðvarp og önnur a ~ ö víxl í íslensku. [ʽu-umlaut and
other a ~ ö alternations in Icelandic.ʼ]. Íslenskt mál 3. 25–58.
Thráinsson, Höskuldur. 1982. Bölvuð talvan. [‘The damned computer.’] Íslenskt mál 4.
Thráinsson, Höskuldur. 1994. Icelandic. In Johan van der Auwera & Ekkehard König
(eds.), The germanic languages, 142–189. Routledge, London.
Thráinsson, Höskuldur. 2011. Um dauðans óvissan tíma: u-hljóðvarp lífs og liðið. [‘Uncertain time of death: On u-umlaut alive and dead.’] Íslenskt mál 33. 85–107.
Thráinsson, Höskuldur, Hjalmar P. Peterson, Jógvan í Lon Jacobsen & Zakaris Svabo
Hansen (eds.). 2012. 2nd edn. Tórshavn & Reykjavík: Faroe University Press – Fróðskapur & Linguistic Institute, University of Iceland.
Valfells, Sigrid. 1967. Umlaut alternations in Modern Icelandic. Cambridge, MA: Harvard
University PhD thesis.
Part II
Chapter 6
Word-based Items-and-processes
(WoBIP): Evidence from Hebrew
Outi Bat-El
Tel-Aviv University
In his seminal book A-Morphous Morphology, Anderson provides ample evidence supporting the item-and-process approach to morphology, whereby relations between words, and
thus the derivation of one word from another is expressed in terms of processes. Although
Anderson excluded Semitic languages from the paradigm, I argue in this paper for the advantage of item-and-process in the analysis of Modern Hebrew word relations. Under this
approach, the word/stem is the base, and the putative consonant root is just a residue of
phonological elements, which are lexically prominent as are consonants in non-Semitic languages. The empirical basis of the arguments is drawn from natural and experimental data
of adult Hebrew as well as child Hebrew.
1 Introduction
“Items vs. processes in morphology” is the title of §3.4 in Anderson’s (1992) seminal book
A-Morphous Morphology. In this section, Anderson compares two models of morphology – item-and-arrangement and item-and-process (attributed to Hockett 1954) – and
argues in favor of the latter. Taking apophony (or ablaut; e.g. sing – sang) as one of the
many problems encountered with the item-and-arrangement model, Anderson claims
that “what presents {past} in sang is … the relation between sang and sing, expressed as
the process by which one is formed from the other” (Anderson 1992: 62; emphasis original). The process in this case is replacement (or stem modification); “the past form of
sing is formed by replacing /ɪ/ with /æ/.” Crucially, /æ/ is not the morpheme designating
past, and sang is not derived by combining bound morphemes, i.e. s-ŋ and -æ-.
The section which immediately follows in Anderson’s book (§3.5) is titled “Wordbased vs. morpheme-based morphology”. The issues addressed in these two sections
are always considered together, since one is contingent upon the other. A root-based
morphology is usually analyzed within the item-and-arrangement model. However, if
Outi Bat-El. 2017. Word-based Items-and-processes (WoBIP): Evidence from Hebrew morphology. In Claire Bowern, Laurence Horn & Raffaella Zanuttini (eds.),
On looking into words (and beyond), 115–135. Berlin: Language Science Press.
Outi Bat-El
morphology is word-based, the debate between item-and-arrangement and item-andprocess still holds (see §2). This debate is particularly heated in the study of Semitic
morphology, where a consonantal root has been claimed to be the core morphological
unit in the word.
Paradigms like sing – sang are relatively rare in English, but abundant in Semitic
languages, such as Hebrew, where the relation between words is often expressed with
apophony; e.g. χam – χom ‘hot – heat’, limed – limud ‘to teach – learning’, ∫uman –
∫émen ‘fat – oil’, gadol –gódel ‘big – size’ (stress is final unless otherwise specified).
Since item-and-arrangement has been the traditional approach to Semitic morphology,
and has been supported by traditional Semiticists (see, however, §6) and generative linguists, Anderson contemplates whether sing – sang can be analyzed as a root s-ŋ plus
the markers /ɪ/ ‘present’ and /æ/ ‘past’. He, however, rejects this analysis due to the
absence of “substantive evidence in its favor” (Anderson 1992: 62), and adds in parentheses “as there clearly is … for something like McCarthy’s analysis of Arabic and other
Semitic languages” (ibid). That is, Anderson accepts the common view that item-andarrangement is the appropriate model for Semitic morphology.
While I support Anderson’s approach to morphology, I do not agree with the exclusion
of Semitic languages from the paradigm. On the basis of data from Modern Hebrew, I
provide in this paper evidence supporting the word-based item-and-process (WoBIP)
model for Semitic morphology. That is, English is not like Hebrew, but rather Hebrew is
like English.
In the context of Semitic morphology, I outline in the following §2 the possible morphological models that can be derived from the four different approaches: word-based,
morpheme-based, item-and-process, and item-and-arrangement. Then, in §§3–5 I provide supporting evidence for the word-based item-and-process model, but due to space
limitation, I do not dwell on arguments against competing models. Each piece of evidence supports only part of the model, but together we get a well-motivated model of
morphology. Given Anderson’s commitment to the history of linguistics (see, in particular, Anderson 1985), I devote §6 to two principal Semiticists from the 19th century, whose
grammar books support the word-based item-and-process model. Concluding remarks
are given in §7.
2 Models of morphology
Research in morphology often concentrates on two questions: What is listed in the lexicon and how are words derived? Each of these questions is associated with competing
approaches. The what-question is related to the root-based vs. word-based debate, which
is of particular interest in the study of Semitic morphology, where the root is always
bound. The how-question is related to the item-and-process vs. item-and-arrangement
debate. Together, they give rise to three models of morphology, shown in Figure 1: rootbased item-and-arrangement, word-based item-and-arrangement, and word-based itemand-process.
6 Word-based Items-and-processes (WoBIP): Evidence from Hebrew morphology
is listed
in the
Roots and configurations
Words and configurations
How are
Association of roots
and configurations
i. Extraction
ii. Association of
roots and configurations
Item & Arrangement
Imposing configurations on words
Item & Process
Figure 1: Models of morphology.
In this paper I support the word-based item-and-process (WoBIP) model. Before displaying the supporting arguments, a short review of the three models is given in the
three ensuing subsections.1
2.1 Root-based item-and-arrangement
In the context of Semitic morphology, the root-based morphology teams up with itemand-arrangement. According to the traditional approach, the root in Hebrew and other
Semitic languages consists of 2–4 consonants (3 in most cases) and is combined with a
configuration (Bat-El 2011), where the latter, traditionally termed mishkal for nouns and
binyan for verbs, is a shorthand for the grouping of prosodic structure, vocalic pattern,
and affixes (if any).2 In a configuration like miCCéCet, for example, the C-slots host the
root consonants, the specified consonants (m and t) are affixes, and the vowels are part
of the vocalic pattern (e.g. mivʁé∫et ‘brush’, mizχélet ‘sleigh’, mi∫méʁet ‘guard’). Table 1
shows examples of words sharing a root and examples of words sharing a configuration.
The classical studies seem to suggest a lexical representation consisting of morphemes,
as can be inferred from Moscati’s (1980: 71) account of the Semitic morphological system:
I do not consider here the pluralistic approaches, whereby some words are derived from roots and others
from words (McCarthy 1979; Arad 2005; Berman 2012), because all phenomena can be accounted for within
the WoBIP model reviewed in §2.2.
2 Each of these elements (i.e. the prosodic structure, the vocalic pattern, and the affix) is independent (McCarthy 1979; 1981), but here reference to the configuration suffices. In this context, we should note that
the term “Semitic morphology” refers to morphology that employs configurations consisting of at least a
vocalic pattern and prosodic structure. Of course, Hebrew and other Semitic languages employ the more
conventional affixal morphology, but this type of morphology does not concern us here.
3 Some words get additional, idiomatic meaning. For example, siduʁ carries the general meaning ‘arrangement’ and the more specific one referring to ‘a prayer book’. Similarly, sédeʁ carries the general meaning
‘order’ and the more specific one referring to ‘Passover ceremony’ (sédeʁ pésaχ).
4 As in other studies, the exponent of the 3rd person masculine past serves as the citation form because it is
structurally neutral, i.e. it has no affixes. The gloss is still in the infinitive, implying reference to the lexeme.
Outi Bat-El
Table 1: Roots and configurations.
Words sharing the root √sdʁ
‘to arrange’
‘military parade’
Words sharing a configuration4
‘to search’
‘to cancel’
‘to connect’
“The Semitic languages present a system of consonantal roots (mostly triconsonantal),
each of which is associated with a basic meaning range common to all members of that
root: e.g. ktb ‘to write’, qbr ‘to bury’, qrb ‘to approach’, etc. These roots (root morphemes)
constitute a fundamental category of lexical morphemes.” If roots are listed, so are the
configurations, and word formation thus consists of associating roots and configurations,
i.e. item-and-arrangement.
As Hoberman (2006: 139) notes, “students of Semitic languages find the concept of the
root so convenient and useful that one finds it hard to think about Semitic morphology
without it.” However, researchers vary with respect to the definition of the term “root”.
Lipiński (1997: 202), for example, assumes that “Semitic roots are continuous morphemes
which are instrumental in derivation but subject to vocalic and consonantal change …
based on continuous or discontinuous ‘pattern morphemes”’ (emphasis original). The
“continuous morphemes”, which Lipiński calls roots, are not the traditional consonantal
roots, but rather stems consisting of vowels and consonants; the “pattern morphemes”
are what I call configurations. Aronoff (2007) drains the original morphological (structural and semantic) properties from the root, claiming that it does not have to be linked
to meaning and its phonology can be vague. Yet another use of the term “root” is found
6 Word-based Items-and-processes (WoBIP): Evidence from Hebrew morphology
in Frost, Forster & Deutsch (1997) with reference to an orthographic root, which as the
results of their experiments suggest, has no semantic properties.
2.2 Word-based item-and-process (WoBIP)
Within this approach, the word or the stem is the core element to which all the required
processes apply (Aronoff 1976). As a core element, it does not have an internal morphological structure. The processes are operations (Anderson 1992: 72) that modify the basic
form (Matthews 1974: 97). Indeed, the most common process in languages is the one
deriving bats from bat, i.e. affixation, but there are other processes, such as apophony,
which derives teeth from tooth.
Also in the context of Semitic morphology, the input is a word/stem to which several
processes apply (see §3.1.2 for word vs. stem as the base). The processes vary according to the goal, and the goal is that the output fits into a configuration. Such a goalor output-oriented phenomenon, called stem modification (Steriade 1988; McCarthy &
Prince 1990), is best analyzed within the framework of Optimality Theory (Prince &
Smolensky 1993/2004), as shown in analyses of Semitic morphology, such as McCarthy
& Prince (1993); Ussishkin (1999; 2000); Gafos (2003); Bat-El (2003).
The details of the required modification depend on the structural similarity between
the base and the output; the more similar they are, the fewer the required adjustments.
Any element in the configuration can be modified – the vocalic pattern, the prosodic
structure, and/or the affix. The modification, however, is contingent upon the configuration of the output.
Table 2: Stem modification – modifying elements in the configuration.
Base form
‘to take care of’
Derived form
Modified elements
vocalic pattern
vocalic pattern, affix
vocalic pattern, affix,
prosodic structure
‘to soap’
Within this approach, there is no morphological element consisting solely of three
consonants, and the emphasis here is on a “morphological element”. Of course, related
words share consonants, but these are stem consonants, where the stem is a morphological unit (e.g. tapél in me-tapel ‘caretaker’), but the consonants are phonological elements.
2.3 Word-based item-and-arrangement
Item-and-arrangement can also be applied within the word-based approach, but only if
a root is extracted from the base word (Ornan 1983; Bolozky 1978). That is, the base is
the word but the root is an intermediate morphological element in the derivation. The
Outi Bat-El
derivation proceeds in two stages – extraction and association (Bat-El 1986; 1989). For
example, the word sabón ‘soap’ serves as the base for the verb sibén ‘to soap’, which is
derived in two stages: (i) extraction of the consonants s,b,n, which automatically become
the root √sbn (traditionally called a secondary root), and (ii) association of this newly
formed root with the verb configuration CiCeC. The assumption is that the extracted
consonants carry the semantic properties of their base, which are, in turn, carried over
to the derived form.
However, root extraction is necessary only when one is limited to the root-based approach, and thus to item-and-arrangement. In this model, all words are derived via association of a root with a configuration, regardless of whether the base is a word or a
root. Not only is there no independent reason to prefer root extraction to stem modification (§2.2), but also there is empirical evidence refuting root extraction. These are cases
of phonological transfer (§3.1), whereby properties that cannot be carried over by the
consonants are transferred from the base to the derived form.
3 Phonological and morphological relations
3.1 Transfer of phonological structure
The most striking evidence for a direct relation between words, without an intermediate
stage that derives a root, is provided by cases exhibiting phonological transfer (Clements
1985; Hammond 1988; McCarthy & Prince 1990). As shown below, there are cases where
structural information, which cannot be encoded in the consonantal root, is transferred
from the base to the derived form. In the case of Hebrew, the structural information is
both prosodic and segmental (Bat-El 1994).
3.1.1 Prosodic transfer
Prosodic transfer includes transfer of the entire configuration or of a consonant cluster.
Configurations are often assigned a grammatical function (Doron 2003), but the question is whether this grammatical function is a property of the configuration or just a
property shared by many (but not all) words within a morphological class. In general,
words that share meaning are often structurally similar, but it does not necessarily mean
that this shared meaning is a property of a morphological unit. One striking example is
displayed by the nouns in Table 3 below, most of which are creative innovations (drawn
from These nouns share the configuration CoCCa and the
meaning ‘related to a computer program’.
Since these nouns share a configuration and meaning, the traditional Semitic morphology would assign the meaning to the configuration. This is, of course, erroneous because
there are other nouns with the configuration CoCCa that do not carry this meaning; e.g.
jo∫ʁa ‘dignity’ (cf. ja∫aʁ ‘honest’), χoχma ‘wisdom’ (cf. χaχam ‘smart’), ot͡sma ‘strength’
(cf. at͡sum ‘huge’), jozma ‘enterprise’ (cf. jazam ‘to initiate’). In addition, this meaning is
too specific to function as a morpho-semantic feature.
6 Word-based Items-and-processes (WoBIP): Evidence from Hebrew morphology
Table 3: Nouns sharing a configuration.
CoCCa noun
Related word
‘computer program’
‘stolen computer program’
‘computer program with porno pop-ups’
‘illegally burned computer program’
‘old computer program’
‘to steal’
‘to burn’
What we actually have here is a Semitic-type blending. The last four words in the first
column of Table 3 use the first word toχna as a base form, from which the configuration
is drawn, along with the basic meaning. That is, toχna provides the configuration CoCCa
and the meaning ‘relating to a computer program’. The stem consonants are drawn
from the related words in the third column of Table 3, along with some specific meaning
denoted by this word. Crucially, such a derivation must be word-based, and the fact that
these words are creative innovations suggests that this model is active in the Hebrew
speakers’ grammar.
Other creative examples are found in a children’s story written by Meir Shalev (ʁoni
venomi vehadov jaakov ‘Roni and Nomi and the bear Jacob’). Each invented word in the
first column of Table 4 has two bases, one providing the configuration and another the
Table 4: Meir Shalev’s invented words.
Invented word
Source of configuration
Source of consonants
‘she wears’
‘she puts on
‘to hurry’
‘to rush’
‘to run’
‘she wears
‘she puts on
‘to hurry/
‘to run out’
Given that the invented words draw semantic properties from the two base words, as
is usually the case with blends, direct access to the base must be assumed. That is, the
configuration of one of the base words is imposed on the other.
Cluster transfer is often found in denominative verbs like tʁansfeʁ → tʁinsfeʁ ‘to transfer’ and faks → fikses ‘to fax’ (Bolozky 1978; McCarthy 1984; Bat-El 1994). In such cases,
the distribution of the sequential order of vowels and consonants, thus including the
clusters, is preserved in the derived form. For example, fílteʁ ‘filter’ is the base of the
Outi Bat-El
verb filteʁ (preserved cluster – lt), while fliʁt ‘flirt’ is the base of fliʁtet (preserved clusters – fl, ʁt), and not *filʁet. Note that the higher the structural similarity between the
base and the derived form, the closer the semantic relation (Raffelsiefen 1993), and thus,
the fewer the structural amendments required in the course of stem modification (§2.2),
the greater the semantic similarity.
3.1.2 Segmental transfer
Segmental transfer includes vowel transfer as well as the transfer of an affix consonant
to the stem (Bat-El 1994).
In vowel transfer, an exceptional configuration is selected because its vowel is identical
to that of the base (e.g. kod ‘code’ → koded ‘to codify’, ot ‘sign’ → otet ‘to signal’). It
should be noted that in most cases, the regular configuration is also possible (e.g. kided
‘to codify’). However, the exceptional configuration is used only when the base has an
o. That is, there is output-output correspondence between the base noun kod and the
derived form, and koded is segmentally more faithful to kod then kided (Bat-El 2003).
In affix transfer, the consonant that serves as an affix in the base becomes a stem
consonant in the derived form. This is common with the suffix -n, as in paʁ∫an ‘commentator’ → piʁ∫en ‘to commentate’ (cf. peʁe∫ ‘to interpret’) and the prefix m-, as in
maχzoʁ ‘cycle’ → miχzeʁ ‘to recycle’ (cf. χazaʁ ‘to return’). Note that speakers’ morphological knowledge allows them to strip the word of its affixes (more so in regular
forms), and therefore the inclusion of an affix consonant in the derived words has its
purpose, mostly to preserve a semantic contrast, as in χizeʁ ‘to court’ vs. miχzeʁ ‘to recycle’ (from maχzoʁ ‘cycle’). But in the paradigm of ∫amaʁ ‘to guard’ – mi∫maʁ ‘guard’
there is no *mi∫meʁ (though it is a potential verb).
3.2 Semantic distance
One crucial property distinguishing among the three approaches reviewed in §3 is the
semantic “distance” between related words; among these, only the WoBIP model (1c)
allows a direct relation between a base and its derived form.
The distance factor
a. Root-based item-and-process
‘to arrange’
6 Word-based Items-and-processes (WoBIP): Evidence from Hebrew morphology
b. Word-based item-and-process
‘to arrange’
c. Word-based item-and-arrangement
‘to arrange’
The advantage of the direct relation (1c) is that information can be carried over from
input to output, be it structural (§3.1) or semantic. It is often the case that within a
group of words sharing stem consonants, there is 1st, 2nd or higher degree of separation
between words, as illustrated in Figure 2.
Such a network can express different degrees of semantic relations, depending on how
far one word is from another. Needless to say, such a network cannot be expressed if
all words are derived from a single root. Of course, one can claim that the three words
at the middle of the network (takdím, kidómet, and mikdamá), which are not directly
related to one another, are derived from a root, while all other words are derived from
words (McCarthy 1979; Arad 2005). However, this is an unsupported and unnecessary
burden on the system. All words in the network are connected to one another, directly
or indirectly, where some words are basic and others are derived. The fact that all the
words in Figure 2 share the stem consonants is due to the important role of consonants
in conveying lexical information and lexical relations (see §5.2).
3.3 Derivation without a configuration
A fundamental element of the traditional root-based item-and-arrangement model is that
every word consists of a root and a configuration, where every configuration is a function. This is particularly essential in the verbal paradigms, where the configurations are
claimed to carry grammatical categories, such as transitivity (Doron 2003; Arad 2005).
Such a theory predicts that the transitivity relation must involve a change in the configuration. This is true for most cases (e.g. katáv ‘to write’ – hitkatév ‘to correspond’, ∫alaχ
‘to send’ – ni∫laχ ‘to be sent’, laχat͡s ‘to press’ – hilχit͡s ‘to cause to feel pressured’), but
not all.
In an extensive study of labile alternations in Hebrew, Lev (2016: 114–115) lists 91 verbs
where transitivity does not involve a change of the configuration; three of his examples
are provided in Table 5. As Lev argues, a root-based morphology cannot accommodate
labile verbs because under this approach the root has to associate with two different con-
Outi Bat-El
‘ancient time’
‘to be early’
‘in front’
‘to promote’
‘to progress’
Figure 2: Degrees of separation.
figurations in order to derive verbs contrasting in transitivity. The word-based approach,
on the other hand, can incorporate labile verbs, assuming that transitivity in such verbs
is not lexically specified but rather derived from the syntactic context. That is, some
verbs are specified for [± transitive] and others, i.e. the labile verbs, are [ø transitive].
Many of the examples in Lev’s list are recent innovations, i.e. where verbs with transitivity specification become labile. For example, the verb tijel used to have one meaning
only, ‘to walk’, but today it also means (at least for some speakers) ‘to walk someone
(usually a dog)’. This change can be viewed as a loss of transitivity specification, i.e. [–
transitive] >[ø transitive]. Crucially, it is the verb that loses its specification for transitivity, not the configuration. That is, in historical change too, as shown in the ensuing
§4, it is the word that changes, and not some putative consonantal root.
6 Word-based Items-and-processes (WoBIP): Evidence from Hebrew morphology
Table 5: Labile verbs (Lev 2016: 114–115).
‘to make black’
‘to polish’
‘to cure’
‘to become black’
‘to shine’
‘to recuperate’
4 Historical change
4.1 Configuration change
Over the course of time, words may change their meaning or their structure. In his study
of instrumental nouns in Hebrew, Laks (2015) shows that quite a few instrumental nouns
undergo change in their configuration, in particular within a compound, as illustrated
in Table 6. As Laks shows, the change always goes towards the participial configuration,
and it never occurs when the instrumental noun does not have a verbal counterpart. That
is, while both maχded ‘pencil sharpener’ and mazleg ‘fork’ have the same configuration,
only the former adopts the participial configuration meχaded, given its verbal counterpart χided ‘to sharpen’; the latter cannot adopt a participial configuration because it does
not have a verbal counterpart.
Table 6: Change of configuration in instrumental nouns (Laks 2015).
Old configuration
New configuration – participle
menaked (tekstim)
soχet (mit͡sim)
‘pencil sharpener’
‘text vocalizer’
‘juicer (juice squeezer)’
In order for this restriction to hold, an instrumental noun must be linked to its verb,
from which it can draw its participial configuration. Otherwise, as Laks argues, the instrumental noun could adopt any of the five participial configurations, not necessarily
the one associated with its verb. That is, we get the instrumental noun meχaded ‘pencil sharpener’ because meCaCeC is the participial configuration of χided ‘to sharpen’.
Similarly, we get the instrumental nouns soχet (mit͡sim) ‘juicer’ because CoCeC is the
participial configuration of saχat ‘to squeeze’. Such a change is possible only in a wordbased lexicon; a root-based lexicon does not account for the restrictive generalization as
it allows options that are not attested.
Outi Bat-El
4.2 Semantic change
Over the course of time, the meaning of words also changes; crucially, the semantic
change affects words and not putative roots. For example, the verbs nimlat and himlit
used to differ in transitivity only, with the former meaning ‘to escape’ and the latter
‘to help someone to escape’. Nowadays, these verbs are not related, since himlit means
‘to give birth (for non-humans)’. Similarly, kalat and hiklit used to be related, with the
former meaning ‘to absorb’ and the latter ‘to cause to absorb’. However, the meaning of
hiklit is now ‘to record’, and the two verbs are vaguely related, if at all. For the traditional
root-based approach (§2.1), it would be rather strange that the change in meaning does
not affect the element that carries it, i.e. the root. This inconsistency does not arise within
the word-based approach.
It is quite feasible that the root does not undergo semantic change because its meaning is just “a basic meaning range”, according to Moscati (1980) and other Semiticists,
or underspecified, according to Arad’s (2005) analysis within the theory of Distributed
Morphology (Halle & Marantz 1993 and subsequent studies). That is, semantic specification of roots may have at least three degrees of specification: fully specified (e.g. boy),
underspecified (e.g. Hebrew roots), and unspecified (e.g. the roots in refer, remit, and
resume; Aronoff 1976).
The major problem is that the specific meaning of words is derived, according to Arad
(2005), from the morpho-syntactic context, i.e. the configurations. This works nicely
for some words but not for others. Consider, for example, the pairs zaʁak ‘to throw’ –
hizʁik ‘to inject’ and ma∫aχ ‘to pull’ – him∫iχ ’to continue’. It is not clear which semantic
property can be assigned to the configurations CaCaC and hiCCiC such that the relation
within these pairs would be consistent.
4.3 Segmental change
Like semantic change, segmental change also affects words and not consonantal roots,
even when the change is in the stem consonants. This is seen in the case of stop-fricative
alternation, which due to its opacity, suffers from a great degree of change-oriented
variation (Adam 2002).
As shown in Table 7, normative verb inflectional paradigms are changing under the
force of paradigm uniformity. Although the change affects consonants, it certainly does
not affect a consonantal root because derivationally related words are hardly ever affected; nonetheless they change, and sometimes they even undergo independent change.
For example, while χ can change to k in katav – jiktov (normative jiχtov) ‘to write past
– future’, it never changes to k in miχtav (*miktav) ‘letter’. Note also that while the
direction of change in this paradigm is from a fricative to a stop (jiχtov → jiktov), the
change in a related pair is towards a fricative, as in ktav → χtav ‘handwriting’.
6 Word-based Items-and-processes (WoBIP): Evidence from Hebrew morphology
Table 7: Change of configuration in instrumental nouns (Adam 2002).
Old paradigm
New paradigm
‘to cover past – future’
‘to write past – future’
‘to cancel past – future’
5 Other supporting sources
5.1 Children’s words
During the early stages of acquisition, verbs in the production lexicon of children acquiring Hebrew are not derivationally related, i.e. they do not share stem consonants.
Derivationally related verbs start appearing later on, where the new verbs “are learnt as
versions of, and based upon, the verbs known from before” (Berman 1988: 62).
This direct derivation in children’s speech is not surprising given Ravid et al.’s (2016)
study on the distribution of verbs in spoken and written Hebrew corpora: child-directed
speech (to toddlers age 1;8–2;2) and storybooks (for preschoolers and 1st–2nd grade). In
both corpora, the average number of verbs per root was below two: 684 verbs for 521
root types in the spoken corpus (1.3) and 1,048 verbs for 744 roots in the written corpus
(1.4). Only around 30% of the verb types in each corpus share a root with another verb,
and most such verbs share a root with only one other verb.
These results mean, as the authors admit, that at least until the age of 7, the children
have very little input supporting a root-based morphology. Nevertheless, the authors
insist that the children must “eventually construe the root as a structural and semantic morphological core” (Ravid et al. 2016: 126). As argued in the current paper and
elsewhere, starting with Bat-El (1994), Hebrew speakers are free from this burden since
Hebrew morphology (and Semitic morphology in general) is not root-based, but rather
Previous studies that attribute children and adults’ innovations to root extraction (§2.3
– word-based item-and-arrangement) must now reconsider their conclusion at least for
children below the age of 7. In an experimental study reported in Berman (2003), children
at the age of 4–6 years old had a rather high success rate (84–88%) of morphological
innovation (forming novel verbs from nouns or adjectives) with a very high percentage
of well-formed innovations (91–99%). If children can form verbs from nouns/adjectives
at the stage where they still do not have sufficient input that allows them to form a
root-based morphology (Ravid et al. 2016), they probably use another strategy – the
modification strategy employed within the WoBIP model (§2.3). And if they can use
this model successfully until the age of 7, they have no reason whatsoever to shift to a
root-based model later on. Of course, as I have argued here and elsewhere, they do not
– Hebrew speakers employ WoBIP at all ages.
Outi Bat-El
5.2 Experimental studies
There are quite a few experimental studies supporting the consonantal root in Hebrew.
Most notable are Berent’s studies with the acceptability rating paradigm (Berent & Shimron 1997; Berent, Everett & Shimron 2001, inter alia) and Frost’s studies with the priming
paradigm (Frost, Forster & Deutsch 1997; Frost et al. 2000, inter alia).5 However, most
experimental studies supporting the consonantal root in Hebrew morphology adopted
a visual modality. As such, they cannot tease apart the effect of orthography, which is
primarily consonantal (Bat-El 2002 and Berrebi 2017 for a critical view).
A fresh look on the matter is provided in Berrebi’s (2017) auditory priming study,
which controlled semantic relatedness and orthographic identity. Word pairs sharing
phonological stem consonants were either semantically related (e.g. kibel ‘to receive’ –
hitkabel ‘to be accepted’) or semantically unrelated (e.g. ʁigel ‘to spy’ – hitʁagel ‘to get
used to’); and when semantically unrelated, either orthographically identical with respect to the consonants (e.g. ʁigel ‘to spy’ – hitʁagel ‘to get used to’) or orthographically
different (e.g. ∫ikeʁ ‘to lie’ – hi∫takeʁ ‘to get drunk’, where k is spelled differently). The
results showed that all conditions had a priming effect, i.e. whether or not the prime
and the target were orthographically identical or semantically related. As the property
shared by the prime and the target in all conditions was phonological, i.e. stem consonants, the results suggest that there is a phonological priming effect among words
sharing stem consonants. Crucially, the stem consonants are not a morphological unit
since there was also a priming effect when the words were semantically unrelated and
orthographically different (e.g. ∫ikeʁ – hi∫takeʁ).
If we assume that priming effects reflect the organization of the lexicon, then we can
conclude that words are also phonologically organized according to the stem consonant.
As emphasized in §2.2, the stem consonants are phonological elements (consonants)
within a morphological unit (stem); they do not carry meaning and they do not constitute a morphological unit.
Stem consonants, and not vowels, serve to identify relations between words because
consonants are lexically prominent, while vowels have syntactic functions (Nespor, Peña
& Mehler 2003; Berent 2013); this is true not only for Hebrew but also for non-Semitic
languages. In their experimental study, Cutler et al. (2000) asked the participants: “Is a
kebra more like cobra or zebra?”. They found that speakers identify similarity between
a nonce word (kebra) and an existing word on the basis of shared consonants (kebra –
cobra) rather than shared vowels (kebra – zebra). That is, the consonants serve as the
core of similarity between words in English, French, Swedish, and Dutch as much as
they do in Hebrew and other Semitic languages (see also Ooijen 1996; New, Araujo &
Nazzi 2008; Carreiras & Molinaro 2009; Winskel & Perea 2013).
Consonants are lexically prominent from the very early stages of language development. This is reported in Nazzi & New’s (Nazzi & New) study, where French 16–20 month
old infants could learn in a single trial two new nonce words if they differed by one con5
In an additional study, which was design to ask “is it a root or a stem?” (rather than “is it a root?”), Berent,
Vaknin & Marcus (2007) note that although their results do not falsify the root-based account, they strongly
suggest that the stem can account for the restrictions on identical consonants.
6 Word-based Items-and-processes (WoBIP): Evidence from Hebrew morphology
sonant (pize – tize), but not if they differed by one vowel (pize – paze). That is, although
vowels are acoustically more prominent than consonants, when it comes to lexical contrast, consonants are employed. This is true for children and adults, regardless of the
ambient language, whether it is Semitic or non-Semitic.
Consonants are prominent not only in speech perception and lexical relations but also
in the association between sound and shape revealed by the bouba-kiki effect (Köhler
1929), whereby people pair labial consonants with round shapes and dorsal consonants
with spiky shapes. One of the many subsequent studies of the bouba-kiki effect is Fort,
Martin & Peperkamp (2015), which found that the sound–shape association remains constant regardless of the vowels. That is, lomo was associated with a round shape as much
as limi, and toko with a spiky shape as much as tiki. Fort, Martin & Peperkamp (2015)
conclude that consonants have a greater effect than vowels in sound – shape association.
6 19th century Semitic grammarians
The root-based item-and-arrangement model of Semitic morphology has been deeply
entrenched for generations, thus presenting the advocates of the word-based item-andprocess approach as revolutionary (see Horvath 1981; Lederman 1982; Heath 1987; Hammond 1988; McCarthy & Prince 1990; Bat-El 1994; 2002; 2003; Ratcliffe 1997; Ussishkin
1999; 2000; 2005; Laks 2011; 2015; Lev 2016).
However, WoBIP has its seeds in the studies of the orientalists Wilhelm Gesenius
(1786–1842) and William Wright (1830–1889), who wrote the seminal grammar books of
Hebrew and Arabic respectively. It is important to note that both Gesenius and Wright
were not native speakers of a Semitic language (Gesenius was German and Wright
British), and thus not biased like the other Semitic grammarians by the consonantal
script of Hebrew and Arabic.
Gesenius (1813) distinguishes between “primitive” verbs, which consist of a stem only
and are not derived from any other form, and derived verbs, among which there are verbal derivatives and denominative verbs. Gesenius used the term “internal modification”
when addressing the processes involved in the derivation. He indicates two types of
“changes in the primitive form” (Gesenius 1813: 115): internal modification (cf. stem modification; §2.2) and repetition (i.e. reduplication) of one or two of the stem consonants.
Within the internal modification, he includes vowel change like gadal ‘to grow’ – gidel
‘to raise’, and gemination as in Biblical Hebrew ga:dal ‘to grow’ – giddel ‘to raise’ (there
is no gemination in Modern Hebrew). Crucially, Gesenius compares vowel modification
in Hebrew to that in English lie – lay and fall – fell, and does not find them different.
That is, Gesenius finds stem modification to be identical in both Hebrew and English,
but unlike Anderson (1992) who contemplates whether English is like Hebrew, Gesenius
actually claims that Hebrew is like English.
A similar approach is found in Wright’s (1859) grammar book of Arabic, where he
describes the relation between verbs within the WoBIP model. For example, “the third
form … is formed from the first by lengthening the vowel sound after the first radical” (p.
32) or “[T]he second form … is formed from the first … by doubling the second radical”
Outi Bat-El
(p. 31). This is the format of Wright’s description of each and every binyan in Arabic, and
it specifically says that (i) one form is derived from another, i.e. word-based derivation,
and (ii) the derivation involves some process, doubling, lengthening, etc., i.e. item-andprocess. Note that Wright uses the term “radical” to refer to a consonant in the stem,
without reference to the stem consonants being an independent morphological unit.
That is, although it has always been said that the root-based approach is the one assumed by traditional Semiticists, it is important to emphasize that the two great 19th century Semiticists, Gesenius and Wright, were proponents of the WoBIP model of Semitic
7 Concluding remarks
In §3.6, Anderson (1992: 71) concludes: “… the morphology of a language consists of a
set of Word Formation Rules which operates on lexical stems to produce other lexical
stems …” In this paper I extended the scope of this model to Semitic morphology. That
is, in Semitic languages too, words are derived from words/stems via modification of the
The modification in Semitic morphology is output oriented, as the output has to fit
into a configuration. The constraint-based framework of Optimality Theory (Prince &
Smolensky 1993) allows for output-oriented grammar, where the constraints impose certain configurations (structural constraints) as well as output-output identity of consonants (faithfulness constraints). A configuration is imposed by several constraints, referring to syllabic structure (usually a foot), syllable structure, and vocalic patterns (where
the latter ones are language specific). Identity among the stem consonants is imposed by
the faithfulness constraints, where preservation of segmental identity ensures preservation of morphological relation among words.
That is, the stability of the stem consonants is due to phonological faithfulness constraints that require identity among stem consonants. Phonological faithfulness enhances morphological relations. “Any given focal word (that is, a specific word in which
we are interested) is thus surrounded by a vaguely defined family of words which are
more or less acoustically similar to it. The members of the family will in general have the
widest variety of meaning, and yet it may often happen that some members of the family
will resemble the focal word not only in acoustic shape, but also in meaning” (Hockett
1958: 297, 1987: 86). That is, within an acoustic family of words there is a morphological
A reviewer suggested that Gesenius and Wright adopted the word-based approach, which was used for
Latin grammar, because they worked prior to the introduction of the term morpheme. Kilbury (1976) and
Anderson (1985) attribute the term morpheme to Baudouin de Courtenay’s student H. Ułaszyn, in his articles from 1927 and 1931. However, it is possible to have a notion without a specific term. Sibawayhi
(760–796), who wrote the first known Arabic grammar Al-kitab, used the term kalima ‘word’ in the sense
of a morpheme (e.g. the suffix -ta) and referred to the radicals that make up the words (Levin 1986). Gesenius notes that the Jewish grammarians call the stem root and the stem consonants radical letters. That is,
there was a reference to morphological units (stem, affixes), but the stem consonants did not constitute a
morphological unit.
6 Word-based Items-and-processes (WoBIP): Evidence from Hebrew morphology
family, where the words are not only acoustically similar but also semantically related.
For the purpose of membership in a morphological family, the consonants are more important than the vowels. This status does not grant the consonants morphological status,
neither in English nor in Hebrew.
As my advisor and mentor, Steve Anderson set the foundations for my approach to
Semitic morphology presented in this paper.
Adam, Galit. 2002. From Optimal Grammar to Variable Grammar: Evidence from language
acquisition and change. Tel-Aviv University PhD thesis.
Anderson, Stephen R. 1985. Phonology in the Twentieth century: Theories of Rules and
Theories of Representations. Chicago: University of Chicago Press.
Anderson, Stephen R. 1992. A-morphous morphology. Cambridge: Cambridge University
Arad, Maya. 2005. Roots and patterns: Hebrew Morpho-syntax. Dordrecht: Springer.
Aronoff, Mark. 1976. Word formation in generative grammar. Cambridge, MA: MIT Press.
Aronoff, Mark. 2007. In the beginning was the word. Language 83. 803–830.
Bat-El, Outi. 1986. Extraction in Modern Hebrew Morphology. University of California, Los
Angeles MA thesis.
Bat-El, Outi. 1989. Phonology and word structure in Modern Hebrew. University of California, Los Angeles PhD thesis.
Bat-El, Outi. 1994. Stem modification and cluster transfer in Modern Hebrew. Natural
Language and Linguistic Theory 12(4). 571–593.
Bat-El, Outi. 2002. Semitic verb structure within a universal perspective. Joseph Shimron
(ed.). Amsterdam: John Benjamins.
Bat-El, Outi. 2003. The fate of the consonantal root and the binyan in Optimality Theory.
Recherches Linguistiques de Vincennes 32. 31–60.
Bat-El, Outi. 2011. Semitic templates. In Marc van Oostendorp, Colin Ewen, Elizabeth
Hume & Keren Rice (eds.), Blackwell companion to phonology, 2586–2608. Malden, MA:
Berent, Iris. 2013. The phonological mind. Cambridge: Cambridge University Press.
Berent, Iris, Daniel Everett & Joseph Shimron. 2001. Do phonological representations
specify variables? Evidence from the Obligatory Contour Principle. Cognitive Psychology 42(1). 1–60.
Berent, Iris & Joseph Shimron. 1997. The representation of Hebrew words: Evidence from
the obligatory contour principle. Cognition 64. 39–72.
Berent, Iris, Vered Vaknin & Gary Marcus. 2007. Roots, stems and the universality of
lexical representations: Evidence from Hebrew. Cognition 104. 254–286.
Outi Bat-El
Berman, Ruth. 1988. Word class distinction in developing grammars. In Izchak
Schlesinger & Yonata Levy (eds.), Categories and processes in language acquisition, 45–
72. Hillsdale, NJ: Erlbaum.
Berman, Ruth. 2003. Children’s lexical innovations: Developmental perspectives on Hebrew verb-structure. In Joseph Shimron (ed.), Language processing and language acquisition in a root-based morphology, 243–291. Amsterdam: John Benjamins.
Berman, Ruth. 2012. Revisiting roots in Hebrew: A multi-faceted view. In Malka Muchnik & Zvi Sadan (eds.), Studies on Modern Hebrew and Jewish Languages, 132–158.
Jerusalem: Carmel.
Berrebi, Si. 2017. The roots of consonant bias: A psycholinguistic study of phonological facilitation in Hebrew. Tel-Aviv University M.A. Thesis.
Bolozky, Shmuel. 1978. Word formation strategies in Modern Hebrew verb system: Denominative verbs. Afroasiatic Linguistics 5. 1–26.
Carreiras, Manuel Jon Andoni Duñabeitia & Nicola Molinaro. 2009. Consonants and vowels contribute differently to visual word recognition: ERPs of relative position priming.
Cerebral Cortex 19. 2659–2670.
Clements, George. 1985. The problem of transfer in nonlinear phonology. Cornell Working
Papers in Linguistics 5. 38–73.
Cutler, Ann, Nuria Sebastiín-Gallés, Olga Soler-Vilageliu & Brit van Ooijen. 2000. Constraints of vowels and consonants on lexical selection: Cross-linguistic comparisons.
Memory and Cognition 28. 746–755.
Doron, Edit. 2003. Agency and voice: The semantics of the Semitic templates. Natural
Language Semantics 11. 1–67.
Fort, Mathilde, Alexander Martin & Sharon Peperkamp. 2015. Consonants are more important than vowels in the bouba-kiki effect. Language and Speech 58. 247–266.
Frost, Ram, Kenneth I. Forster & Avital Deutsch. 1997. What can we learn from the morphology of Hebrew? a masked-priming investigation of morphological representation.
Journal of Experimental Psychology: Learning, Memory, and Cognition 23(4). 829–856.
Frost, Ram, Avital Deutsch, Oran Gilboa, Michael Tannenbaum & William MarslenWilson. 2000. Morphological priming: Dissociation of phonological, semantic, and
morphological factors. Memory & Cognition 28. 1277–1288.
Gafos, Adamantios I. 2003. Greenberg’s asymmetry in Arabic: a consequence of stems
in paradigms. Language 79(2). 317–355.
Gesenius, Wilhelm. 1813. Gesenius’ Hebrew grammar. Oxford: Clarendon Press.
Halle, Morris & Alec Marantz. 1993. Distributed morphology and the pieces of inflection. In Kenneth Hale & Samuel Jay Keyser (eds.), The view from Building 20: Essays in
linguistics in honor of Sylvain Bromberger, 111–176. Cambridge, MA: MIT Press.
Hammond, Michael. 1988. Templatic transfer in Arabic broken plurals. Natural Language
& Linguistic Theory 6. 247–270.
Heath, Jeffrey. 1987. Ablaut and ambiguity: Phonology of a Moroccan Arabic Dialect. Albany: SUNY Press.
Hoberman, Robert. 2006. Semitic triradicality or prosodic minimality? Evidence from
sound change. In Cynthia Miller (ed.), Studies in Semitic and Afroasiatic linguistics
6 Word-based Items-and-processes (WoBIP): Evidence from Hebrew morphology
presented to Gene B. Gragg, 139–154. Chicago: The Oriental Institute of the University
of Chicago Studies in Ancient Oriental Civilization.
Hockett, Charles F. 1954. Two models of grammatical description. Word 10. 210–234.
Hockett, Charles F. 1958. A course in modern linguistics. London: Macmillan.
Hockett, Charles F. 1987. Refurbishing our foundations. Amsterdam: John Benjamins.
Horvath, Julia. 1981. On the status of vowel patterns in Modern Hebrew: Morphological
rules and lexical representations. Extended Word-and-Paradigm Theory. UCLA Occasional Papers in Linguistics 4. 228–261.
Kilbury, James. 1976. The Development of Morphophonemic Theory. Amsterdam: John Benjamins.
Köhler, Wolfgang. 1929. Gestalt psychology. New York: Liveright.
Laks, Lior. 2011. Morpho-phonological and morpho-thematic relations in Hebrew and Arabic
verb formation. Tel-Aviv University PhD thesis.
Laks, Lior. 2015. Variation and change in instrument noun formation in Hebrew and its
relation to the verbal system. Word Structure 8. 1–28.
Lederman, Shlomo. 1982. Problems in a prosodic analysis of Hebrew morphology. Studies
in the Linguistic Sciences 12. 141–163.
Lev, Shaul. 2016. Hebrew labile alternations. Tel-Aviv University MA thesis.
Levin, Aryeh. 1986. The medieval Arabic term kalima and the modern linguistic term
morpheme: Similarities and differences. In Dāwid Ayyālôn & Šārôn (eds.), Studies in
Islamic history and civilization: In honour of Professor David Ayalon, 423–446. Leiden:
Lipiński, Edward. 1997. Semitic languages: Outline of comparative grammar. Leuven: Uitgeverij Peeters en Department Oosterse Studies.
Matthews, Peter H. 1974. Morphology: An introduction to the theory of word-structure.
Cambridge: Cambridge University Press.
McCarthy, John. 1984. Prosodic organization in morphology. In Mark Aronoff & Richard
Oehrle (eds.), Language Sound Structure, 299–317. Cambridge, MA: MIT Press.
McCarthy, John & Alan Prince. 1990. Foot and word in Prosodic Morphology: the Arabic
broken plural. Natural Language and Linguistic Theory 8. 209–283.
McCarthy, John & Alan Prince. 1993. Prosodic morphology i: constraint interaction and
satisfaction. Ms., University of Massachusetts, Amherst.
McCarthy, John J. 1979. Formal problems in Semitic phonology and morphology. MIT PhD
McCarthy, John J. 1981. A prosodic theory of nonconcatenative morphology. Linguistic
Inquiry 12(3). 373–418.
Moscati, Sabatino. 1980. An introduction to the comparative grammar of the Semitic languages. Wiesbaden: Otto Harrassowitz.
Nazzi, Thierry & Boris New. 2007. Beyond stop consonants: Consonantal specificity in
early lexical acquisition. Cognitive Development 22. 271–279.
Nespor, Marina, Marcela Peña & Jacques Mehler. 2003. On the different roles of vowels
and consonants in speech processing and language acquisition. Lingue e Linguaggio 2.
Outi Bat-El
New, Boris, Veronica Araujo & Thierry Nazzi. 2008. Differential processing of consonants
and vowels in lexical access through reading. Psychological Science 19. 1223–1227.
Ooijen, Brit van. 1996. Vowel mutability and lexical selection in English: Evidence from
a word reconstruction task. Memory & Cognition 24. 573–583.
Ornan, Uzzi. 1983. How is the Hebrew word formed? In Moshe Bar-Asher, Aron Dotan,
Davic Tene & Gad B. Sarfatti (eds.), Language studies, 13–42. in Hebrew. Jerusalem:
Prince, Alan & Paul Smolensky. 1993. Optimality Theory: Constraint interaction in generative grammar. Oxford: Blackwell Publishing.
Prince, Alan & Paul Smolensky. 1993/2004. Optimality Theory: Constraint interaction in
generative grammar. Oxford: Blackwell Publishing.
Raffelsiefen, Renate. 1993. Relating words: A model of base recognition. Linguistic Analysis 23. 3–139.
Ratcliffe, Robert R. 1997. Prosodic templates in a word-based morphological analysis of
Arabic. In Mushira Eid & Robert Ratcliffe (eds.), Perspectives on Arabic Linguistics X,
147–171. Amsterdam: John Benjamins.
Ravid, Dorit, Orit Ashkenazi, Ronit Levie, Galit Zadok, Tehila Grunwald, Ron Bratslavsky
& Steven Gillis. 2016. Foundations of the early root category: Analyses of linguistic
input to Hebrew-speaking children. In Ruth Berman (ed.), Acquisition and development
of Hebrew, from infancy to adolescence, 95–134. Amsterdam: John Benjamins.
Steriade, Donca. 1988. Reduplication and syllable transfer in Sanskrit and elsewhere.
Phonology 5(1). 73–155.
Ussishkin, Adam. 1999. The inadequacy of the consonantal root: Modern Hebrew denominal verbs and output-output correspondence. Phonology 16(3). 401–442.
Ussishkin, Adam. 2000. Root-and-pattern morphology without roots or patterns. In Masako Hirotani, Andries W. Coetzee, Nancy Hall & Ji-yung Kim (eds.), Proceedings of
the 30th meeting of the North East Linguistic Society, 655–670. Amherst, MA: GLSA.
Ussishkin, Adam. 2005. A fixed prosodic theory of nonconcatenative templatic morphology. Natural Language and Linguistic Theory 23(1). 169–218.
Winskel, Heather & Manuel Perea. 2013. Consonant/vowel asymmetries in letter position
coding during normal reading: Evidence from parafoveal previews in Thai. Journal of
Cognitive Psychology 25. 119–130.
Wright, William. 1859. A grammar of the Arabic language. Cambridge: Cambridge University Press.
Chapter 7
Root-based syntax and Japanese
derivational morphology
Brent de Chene
Waseda University
This paper argues that the formation of transitive and intransitive verb stems in Japanese, a
process that has been widely seen as supporting the Distributed Morphology view of derivational stem-formation as performed by the syntax, cannot in fact be analyzed as syntactic.
The Japanese data are thus consistent with Anderson’s (1982) claim that it is precisely that
morphology traditionally classified as inflectional that reflects syntactic operations.
1 Introduction
In a well-known paper, Anderson (1982: 587) proposes that “Inflectional morphology
is what is relevant to the syntax,” where syntactically relevant properties are those “assigned to words by principles which make essential reference to larger syntactic structures.” He claims further that a delimitation of inflection on this basis closely mirrors
the traditional understanding of where the boundary between inflection and derivation
lies. In contrast, the Distributed Morphology literature, in treating syntax as root-based
and stem formation of all types as syntactic, denies significance to the traditional distinction between inflection and derivation and renders vacuous the claim that inflection
is just that portion of morphology that realizes elements and properties manipulated
by the syntax.1 The present paper takes up the formation of transitive and intransitive
verb stems in Japanese, a case that has been widely seen as supporting the DM view of
stem-formation as performed by the syntax, and argues that a closer look reveals that
the derivational processes in question cannot in fact be analyzed as syntactic. In the
The founding paper of the DM framework, Halle & Marantz (1993), presents DM as a theory of inflection
and makes no explicit claims about derivation, but the adoption of root-based syntax and the rejection
of the inflection/derivation distinction are clear at least by Marantz (1997; 2001). See below for further
Brent de Chene. 2017. Root-based syntax and Japanese derivational morphology. In
Claire Bowern, Laurence Horn & Raffaella Zanuttini (eds.), On looking into words (and
beyond), 135–155. Berlin: Language Science Press. DOI:10.5281/zenodo.495442
Brent de Chene
end, then, the Japanese data is consistent with Anderson’s view that there is a fundamental distinction between inflection and derivation and that the criterion of syntactic
relevance picks out just that morphology traditionally classified as inflectional.2
In recent years, the derivational morphology of the Japanese verb has become a standard example (as in Harley 2012) illustrating the DM claim that syntax is root-based – the
claim, that is, that along with functional morphemes, the atoms of syntactic computation
are roots rather than (inflectable) stems or (inflected) words (Embick & Marantz 2008: 5).
In particular, it has become widely accepted (Marantz 2013: 106) that the Japanese suffixes that create transitive and intransitive verb stems are instances of little v, causative
and inchoative, that attach to roots and thus that the verb stems themselves are syntactic
constructions – much like, say, the combination of a verb stem with a tense element or
a main verb with an auxiliary. Here, I note first that these claims about the constituency
of Japanese verb stems rest on a restricted database that masks the fact that a significant number of stems involve sequences of two transitivity-determining suffixes. I then
present the failure of two nested suffixes to interact in the way expected of syntactic
elements – in particular, the fact that an inner suffix must be taken as invisible for purposes of semantic interpretation and argument structure – as the first of several related
arguments casting doubt on the proposal to generate Japanese verb stems syntactically.
The data on which DM theorists base their claim that the verbal derivational suffixes
of Japanese are instances of little v attaching to roots is the appendix of Jacobsen (1992),
which represents a light revision of the appendix of Jacobsen (1982), and in turn appears
lightly revised as Appendix I in Volpe (2005). That appendix consists of roughly 350 pairs
of isoradical intransitive and transitive verbs presented in their citation forms (Imperfect/Nonpast Conclusive) and sorted into sixteen classes depending on the derivational
suffixes that appear at the right edge of their stems. The fact that the Jacobsen/Volpe
appendix is limited to verb stems presented pairwise means that using it as a basis for
the identification of root requires assuming for each transitivity pair that there are neither stems of other lexical categories nor verb stems outside the transitivity pair that
provide information about the relevant root. §2 below, in the context of presenting background information on Japanese derivation, introduces a number of cases in which this
assumption is unjustified. The following three sections, building on the observations of
§2, present reasons for doubting that verb stems are syntactically derived. While for
concreteness I refer throughout to the DM literature cited above and related work, the
argumentation is intended to apply to any proposal to generate Japanese verb stems
§3, first, shows that a substantial minority of verb stems involve two transitivizing (T)
or intransitivizing (I) suffixes (with the four orders TT, TI, IT, II all attested), but that
an outer suffix must be taken to render an inner one null and void for purposes of argument structure and semantic interpretation. §4 shows that the same is true for the suffix
On a personal note, while I have taken the idea that inflection is precisely the syntactically relevant morphology as a guiding principle for many years, it was anything but obvious to me at the time Steve proposed
it. It ranks high in my personal inventory of the many things I have learned from Steve, and I am happy
to have this opportunity to reaffirm it in a volume dedicated to him.
7 Root-based syntax and Japanese derivational morphology
pair -m- (verbal) and -si- (adjectival), with the additional complication that the order in
which those two suffixes appear relative to a root R is an idiosyncratic function of R. §5,
finally, argues against a syntactic account of stem formation on the basis of semantic
change, claiming, for lexical causatives in particular, that the diachronic instability of
the putatively compositional causative interpretation (much as if a phrase like kick the
bucket were to lose its compositional interpretation, retaining only the idiosyncratic one)
shows that that interpretation cannot have been based on a syntactic derivation in the
first place. In all of these cases, the behavior of the derivational suffixes under consideration is contrasted with that of inflectional and uncontroversially syntactic elements.
§6, a brief conclusion, sketches two possible non-syntactic approaches to derivational
morphology and speaker knowledge thereof and suggests that the choice between them
for cases like the one considered here remains a topic for further research.
2 Background
In considering the shortcomings of Jacobsen’s (1982; 1992) appendix as a database for
Japanese verbal derivation, the first thing to note is that the pairwise presentation of the
data does not always adequately represent the relations of isoradicality that hold among
verb stems. This is because a number of roots underlie three or (in at least one case) four
verb stems rather than two; in such cases, Jacobsen either lists two pairs in separate
places or, as we will see below, omits one of the stems. In several cases involving three
stems on a single root, there are two pairs of stems differentiated at least roughly by root
alloseme, with a formal contrast for either transitives or intransitives but not both. For
example, the difference between the allosemes ‘solve’ and ‘dissolve, melt’ of the root tokcorresponds to a formal distinction for transitives but not for intransitives, as shown in
(1) and (2).3
a. tok-e- ‘be solved’
b. tok- ‘solve’
a. tok-e- ‘melt (i)’
b. tok-as- ‘melt (t)’
In other cases, as in (3) and (4), there is no alloseme-dependent pairing, simply a triplet
of isoradical stems.
a. tunag-ar- ‘be connected’
b. tunag-e- ‘connect (t)’
c. tunag- ‘connect (t)’
Below, taking the distinction between inflection and derivation in Japanese to be uncontroversial, I use
stem in the traditional meaning “morpheme (sequence) subject to inflection” and cite bare stems rather
than inflected forms; “(i)” and “(t)” in glosses indicate intransitive and transitive meanings, respectively.
Brent de Chene
a. uk- ‘float (i)’
b. uk-ab- ‘float (i)’
c. uk-ab-e- ‘float (t)’
In these last two cases, the policy of pairwise listing results in one stem of each isoradical
set (specifically, 3b and 4a) being left out of the database.
In fairness to Jacobsen, it must be noted that morphological analysis was not his aim
in compiling his appendix. Most crucially for our purposes, he nowhere refers to the
notion “root”, and it is only with Volpe’s (2005) DM treatment that the root becomes a
central concept in the interpretation of the appendix data. Volpe’s (2005: 121 (note 27))
procedure for root extraction, however, amounts to simply peeling off the outermost
derivational suffix and labeling the residue a root, and he has been followed implicitly
in this practice by other DM theorists.
We should observe before proceeding that there are many cases, illustrated by (5)
below, in which Volpe’s procedure does in fact yield a root.
a. nao-r- ‘get better (illness, injury); get repaired’
b. nao-s- ‘cure; repair’
(5) is clearly the kind of case Marantz (2013: 106) has in mind when he says about Japanese
that “there seems overwhelming support for analyzing the suffixes signaling either the
lexical causative as opposed to the inchoative or the inchoative as opposed to the lexical
causative as realizations of a little v head attaching to the root.” As we will now see,
however, there are a number of respects in which the properties of (5) do not generalize
to the Japanese derivational system as a whole. Most crucially, there is reliable evidence
for a number of Volpe’s “roots” that they are actually morphologically complex, with
the result that many verb stems contain two derivational suffixes rather than one. Given
that, as we have already noted, Volpe’s procedure for root extraction involves no attempt
to compare verb stems with stems of other lexical classes or with verb stems outside the
transitivity pair under consideration, this result is unsurprising. Let us examine a few
representative cases.
Consider the sequence tunag- of (3) above. Comparison of that sequence, roughly
meaning ‘connect’, with the noun tuna ‘rope’ suggests that the former is undersegmented, and in particular that the transitive stem tunag- consists of the noun stem tuna
(or the root that underlies it) suffixed with -g-. This suggestion is confirmed when we
observe that -g- is suffixal in a number of other stems as well, with a core subset ((6–7
below and the three of note 3) displaying a very specific semantics: -g- takes as input a
noun stem denoting a tool T and returns a verb stem with the meaning “to make typical
use of T”. Three examples that occasion resegmentation of entries of the Jacobsen/Volpe
appendix are given in (6) through (8), with both a transitive and an intransitive stem
noted in each case.4
Three further examples whose status in the contemporary language might be thought questionable are
tumu-g- ‘spin (thread)’ (tumu ‘spindle’), ha-g- ‘fletch (arrow)’ (ha ‘feather’), and, with an irregular alternation of t with s, husa-g- ‘cover, stop up’ (huta ‘cover’).
7 Root-based syntax and Japanese derivational morphology
a. tuna ‘rope’
b. tuna-g- ‘tie together, tie up’
c. tuna-g-ar- ‘get connected’5
a. to(-isi) ‘whetstone’
b. to-g- ‘whet’
c. to-g-ar- ‘become pointed’
a. mata ‘crotch, fork’
b. mata-g- ‘step over, straddle (t)’
c. mata-g-ar- ‘straddle (i)’
The derivational relationships postulated in (6–8) appear unimpeachable in both formal
and semantic terms: the roots are nonalternating, and the semantic relationship between
nominal and verbal meanings is unmistakable.
More common as a stem-forming suffix than -g- is -m-, which can be shown to be a
stem formant in several dozen verbs. (9–11) display three cases in which recognition of
suffixal -m- forces resegmentation of strings that Volpe takes to be roots (the (a) items
of (9) and (10) are adjective stems, and that of (11) is an adjectival noun, a stem with
adjectival meaning but essentially nominal inflection).
a. ita- ‘painful’
b. ita-m- ‘be painful, get injured’
c. ita-m-e- ‘injure’
a. yuru- ‘slack’
b. yuru-m- ‘slacken (i)’
c. yuru-m-e- ‘slacken (t)’
a. hiso-ka ‘stealthy, secret’
b. hiso-m- ‘be hidden, lurk’
c. hiso-m-e- ‘conceal, mask’
We have seen that in addition to verb stems formed with the common suffixes -r- and
-s-, illustrated in (5), there are verb stems formed with -g- and -m-. In fact, of the nine
occurring stem-final consonants, all but n can be shown to be suffixal in some stems.
Suffixal -b- has been illustrated in (4b) above; (12) through (14) display one example each
for -k-, -t-, and -w- (w deletes in the phrasal phonology before nonlow vowels; here and
below, I take reference to a suffix -C(V)- to subsume reference to its post-consonantal
allomorph -aC(V)-).
Kunio Nishiyama (personal communication) suggests the possibility that -g- in (6) is a (transitivity-neutral)
verbalizer, with the transitivity of (6b) resulting from a null transitivizer parallel to the intransitive -ar- of
(6c). A fully general form of this proposal will require the postulation of a very large number of morphological zeros.
Brent de Chene
a. na-k- ‘make characteristic sound’ (animal); ‘weep’ (human)
b. na-r- ‘sound (i)’ (inanimate subject)
c. na-r-as- ‘sound (t)’
a. hana-re- ‘move (i) away (from); be released’
b. hana-s- ‘move (t) away (from); release’
c. hana-t- ‘release forcefully, discharge’
muk- ‘face, look (in a direction)’
muk-e- ‘cause to face, turn (t) (in a direction)’
muk-aw- ‘face, proceed toward’
muk-aw-e- ‘(go to) meet, receive (a visitor)’
We see, then, that the inventory of suffixes that create verb stems of determinate transitivity is a good deal larger than envisioned in the Jacobsen/Volpe appendix, where,
apart from idiosyncratic formations, the relevant set is essentially limited to -r-, -s-, -re-,
-se-, -e-, -i-, and zero. In closing this introductory section, let us consider two semantic
issues that arise with respect to the Jacobsen/Volpe appendix data. The first involves the
interpretation of roots, the second the interpretation of suffixes.
Quite apart from the question of whether or not roots are taken to be elements that
are manipulated by the syntax, no attempt to segment stems into roots and suffixes synchronically is a fully grounded project in the absence of a criterion for isoradicality –
a criterion, that is, for determining when two given stems share a root and when they
do not. The semantic lability of individual stems over time that will be illustrated in §5
makes this by no means an idle question. It is, however, a question that neither Jacobsen
nor Volpe engage with seriously; Jacobsen (1982: 38)6 says only that the members of a
transitivity pair must exhibit “a certain degree of semantic affinity”, and Volpe (2005:
32) confines himself to observing that “Root semantics is a wide-open area for further
research”. The question of isoradicality is essentially coextensive with the traditional
problem of distinguishing homophony from polysemy, a problem that may ultimately
be illuminated by psycholinguistic and neurolinguistic research (see Marantz 2013: 103).
It is worth keeping in mind, however, that any program involving the synchronic identification of roots requires innumerable provisional decisions on this matter.
Turning now to the interpretation of the stem-forming suffixes of which we have seen
a number of examples, let us note first that while Volpe (2005) follows Jacobsen (1982;
1992) in referring to the two members of a transitivity pair as “intransitive” and “transitive”, more recent literature such as Harley (2008; 2012) and Marantz (2013) use the more
specific “inchoative” and “causative”. In fact, cases like ka-r- (Western Japan; cf. Eastern
ka-ri-) ‘borrow’ versus ka-s- ‘lend’ and azuk-ar- ‘take on deposit’ versus azuk-e- ‘deposit’
show that even the former pair of terms is too specific to be accurate in general. This is because the first member of each of those pairs shows “intransitive” morphology in spite of
displaying what, under Burzio’s generalization, are the twin hallmarks of causative little
See also note 5, p.34 and the corresponding note 30 of Jacobsen 1992 (pp. 248–249).
7 Root-based syntax and Japanese derivational morphology
v, namely an agentive external argument and accusative case-marking. Cross-linguistic
parallels7 suggest that the treatment of ‘borrow’ as the intransitive counterpart of ‘lend’
is by no means accidental or exceptional. The phenomenon of a stem with causative
meaning but “intransitive” morphology appears to show that if the semantics of the two
morphological types are specified separately, they will have to overlap. Let us briefly
note another type of example that suggests the same conclusion.
The stems too-r- ‘pass through’ and mata-g- ‘step over, pass over, straddle’ (8b above)
are closely parallel in both their semantics and their case-marking. When the subject
is animate, as in (15) (where stem-internal segmentation is suppressed), that subject
(marked nominative but omitted in the examples) is both an agent and a theme moving along a path, and the accusative object is an intermediate point on that path.
a. Syootengai
o toot-te
ni modot-ta.
shopping.district acc pass.through-cj station dat return-pf
‘I passed through the shopping district and returned to the station.’
b. Saku o matai-de hodoo ni hait-ta.
barrier acc step.over-cj sidewalk dat enter-pf
‘I stepped over the barrier and onto the sidewalk.’
In other uses, the agent of examples (15) may be replaced by an inanimate theme, with
matag- in the meaning ‘pass over’, or by a path argument, as in The road passes through
the tunnel/over the train tracks.
In spite of the close semantic parallelism between too-r- and mata-g-, however, the
two stems differ in their transitivity status: too-r- is the intransitive corresponding to
transitive too-s- ‘pass though (t)’, while mata-g- is the transitive corresponding to intransitive mata-g-ar- ‘straddle’ (8c above), the latter differing from mata-g- in taking a
dative rather than an accusative object. Unless too-r- and mata-g- are semantically distinct in a way we have failed to identify, this fact shows that the transitivity status of
a stem cannot be a function of that stem’s semantics alone, and a fortiori cannot be a
function of the semantics of that stem’s suffix. An alternative possibility, which considerations of space preclude developing here, is that there is a continuum of degrees of
transitivity, as suggested by Hopper & Thompson (1980) and subsequent work, and that
what transitivity pairs have in common is that the “transitive” member has a higher degree of transitivity than the “intransitive” member.8 In any case, however, the evidence
we have seen here is sufficient to establish that there is no simple, general account of
the semantics of the suffixes that create transitivity-specific Japanese verb stems, and
that, as was the case regarding the question of a criterion for isoradicality, much work
remains to be done in this area.
Above, we have seen that the data of the Jacobsen/Volpe appendix is a good deal more
complex and irregular, both formally and semantically, than consideration of examples
See Kuo (2015: 59, 84–85, 107) for the Taiwanese languages Amis, Puyama, and Seediq, respectively; other
languages for which the relationship can be easily verified include Tagalog and Swahili.
Jacobsen (1992: 73–74) develops a scalar concept of transitivity but does not suggest that the common point
of transitivity pairs is a transitivity differential in favor of the morphologically transitive member.
Brent de Chene
like (5) might suggest. Nothing in the present section, however, is intended as an argument for or against any particular treatment of that data. Taking our discussion of the
Jacobsen/Volpe appendix as a starting point, we now turn, in Sections 3 through 5, to
arguments against proposals to generate Japanese verb stems syntactically.
3 Sequences of verbal suffixes
As we have already noted, one consequence of the resegmentations that are entailed
by comparing the stems that participate in transitivity pairs with stems of other lexical
categories (as well as with other verb stems) is that many stems can be seen to display a
sequence of two suffixes attached successively to a root rather than a single transitivitydetermining suffix. For example, the (c) examples of (6) through (8) above all involve
the sequence -g-ar-, where the first suffix creates a transitive stem and the second an
intransitive. Similarly, the (c) examples of (9) through (11) all involve -m-e-, where the
first suffix creates an intransitive stem and the second a transitive. Suffix sequences are
also observed in (12c) and (14d).
Sequences of two transitivizing suffixes and two intransitivizing suffixes are observed
as well. For example, (16d) below, where (16) is an expansion of (6), involves the sequence
-g-e-, where both suffixes create transitive stems, and (17c) involves the sequence -m-ar-,
where both suffixes create intransitive stems.
tuna ‘rope’
tuna-g- ‘tie together, tie up’
tuna-g-ar- ‘get connected’
tuna-g-e- ‘tie together, connect’
yasu-raka ‘peaceful, calm’
yasu-m- ‘rest (i)’
yasu-m-ar- ‘become rested, at ease’
yasu-m-e- ‘rest (t)’
Recall now the DM claim that Japanese transitivity-determining suffixes are instances
of little v, with at least an inchoative and a causative “flavor” (Marantz 2013: 107) to be
distinguished. Abstracting away from the fact that (at a minimum) both types of little v
will have to be polysemous, and writing the inchoative version as “vi ” and the causative
version as “vc ”, the structure of the two stems of (5), for example, will be as shown in
(18) (simplified glosses given) .
a. nao-r- [[R]vi ] ‘get better’
b. nao-s- [[R]vc ] ‘make better’
In the same way, the structure of the stems (16c–16d) will be as in (19), and that of the
stems (17c–17d) will be as in (20). (Here and below, I take the fact that -g- and -m- (and
also -b-, -k-, -t-, -w-) in isolation are entirely parallel in function to the suffixes the DM
7 Root-based syntax and Japanese derivational morphology
literature treats as little v (notably -r-, -s-, and -e- (see e.g. Marantz 2013: 108) to license
a parallel treatment for them in the DM framework we are taking as representative of
syntactic treatments of derivation.)
a. tuna-g-ar- [[[R]vc ]vi ] ‘connect (i)’
b. tuna-g-e- [[[R]vc ]vc ] ‘connect (t)’
a. yasu-m-ar- [[[R]vi ]vi ] ‘get rested’
b. yasu-m-e- [[[R]vi ]vc ] ‘rest (t)’
If the representations of (19–20) are constructed in the syntax, in line with the proposal
that roots and functional morphemes are the primitives of syntactic derivation, we will
expect them to be interpreted compositionally, with the meaning of the outer little v
combining with the result of composing the meaning of the inner little v with that of
the root. In fact, no verb stem has an interpretation that involves two units of “little v
meaning”, either two instances of “inchoative” or two instances of “causative” or one of
each; for interpretive purposes, the only little v that matters in representations like those
of (19–20) is the outer one.9 This is as if, when the Perfect auxiliary occurs outside of the
Progressive in English or the Passive outside of the (productive) Causative in Japanese,
as illustrated in (21), the outer auxiliary were to nullify the interpretation of the inner
one rather than composing with it semantically.
a. have been eating [PERF[PROG[V]]]
b. tabe-sase-rare- [[[V]CAUS]PASS] ‘be made to eat’
It would seem that in uncontroversially syntactic constructions like those of (21), this
kind of nullification never occurs, and thus that we can assume that the syntactic computational system includes no mechanism for opting out of compositional interpretation
in this way. The structures of (19–20) therefore pose a major problem for the idea that
the suffixes deriving Japanese verb stems are syntactic elements.
We have seen that the syntactic status of constructions like (19–20) is called into question by their interpretive properties. The representations of (19) pose a second problem
as well, namely that the internal vc will introduce an external argument that must ultimately remain unrealized.10 In the remainder of this section, I concentrate on documenting further instances of the construction (19a), verb stems that introduce no external
argument in spite of containing a transitivizing suffix.
While the vi of (20b) could be taken to be semantically active, the meaning of such causatives would have
to coincide with that of causatives derived from roots, as in (18b). The semantic inertness of the inner little
v thus follows for this case as for the others. (In DM, identification of category-determining elements with
phase heads requires that lexical causatives, being monophasal, be root-based (Marantz 2007).)
10 The causative interpretation and the external argument may in fact be introduced by separate heads
(Pylkkänen 2008: chapter 3); what is important for our purposes is that in the data at hand they are both
present when a transitivizing suffix appears alone but absent when it appears inside another transitivitydetermining suffix.
Brent de Chene
Consider first the isoradical sets (22–25), all of which illustrate the suffix sequence
a. mak- ‘roll up, wind around’
b. maku-r- ‘roll up, tuck up’
c. maku-r-e- ‘get turned up, ride up’
a. nezi ‘screw’
b. nezi-r- ‘twist’
c. nezi-r-e- ‘get twisted’
a. yabu-k- ‘rip (t)’
b. yabu-r- ‘rip (t)’
c. yabu-r-e- ‘rip (i)’
kasu-ka ‘faint, at the limits of perception’
kasu-m- ‘become hazy, dim’
kasu-m-e- ‘cloud (the vision of), deceive; graze, skim over; skim off, steal’
kasu-r- ‘graze (touch lightly in passing)’
kasu-r-e- ‘become faint or discontinuous (printing, writing); become hoarse
The stems of (22–25) are all in common use in contemporary Japanese; a final parallel set
that is particularly transparent semantically but for which the verb stems are obsolete
is kubi ‘neck’, kubi-r- ‘strangle’, kubi-r-e- ‘die by hanging oneself’.
Examples of the construction (19a) involving the suffix sequence -m-ar- can also be
cited, as in (26–28). (26a) reflects the fact, not previously exemplified, that bare roots not
infrequently occur reduplicated as adverbial items of the mimetic vocabulary.
kurukuru ‘round and round (rotation, winding)’
kur- ‘reel in, wind’
kuru-m- ‘wrap by rolling’
kuru-m-ar- ‘be rolled up, wrapped up’
kuru-m-e- ‘lump together’
tuka ‘hilt, handle’
tuka-m- ‘grasp’ (accusative object)
tuka-m-ar- ‘be caught, captured’; ‘hold on to’ (dative object)
tuka-m-aw-e- ‘catch, capture’
Taking the root to be maku- in (22) obviates postulating a new suffix allomorph for the (b) and (c) examples
but requires a rule deleting a root-final vowel in a zero-derived verb stem for (22a). Given also a rule a
+ i > e, mirroring the presumed historical development (see Ono 1953 and subsequent literature), many
apparently consonant-final roots could be reanalyzed along parallel lines; for example, the stems of (1–2)
above could be tok-, toka-i-, toka-s- (√toka) rather than tok-, tok-e-, tok-as- (√tok).
7 Root-based syntax and Japanese derivational morphology
a. haza-ma ‘gap, interstice’ (< hasa-ma (ma ‘interval’))
b. hasa-m- ‘insert between’
c. hasa-m-ar- ‘get caught between’
In (6–8) and (22–28), then, we have seen examples in which intransitivizing suffixes
appear outside transitivizing suffixes, resulting in stems of the shape (19a). These are
structures for which, as a result of the internal vc , both a causative interpretation and an
external argument are predicted, but do not materialize. We have already argued that
the syntactic status of all four constructions (19–20) is called into question by the fact
that the inner little v of those constructions is never interpreted. Regarding the unrealized external argument of stems of the shape (19a), similarly, it is clear that there is
no way, in a system of syntactic derivation based on selectional features and the Merge
operation and restricted by a “no tampering” condition (Chomsky 2008: 138), for a specifier introduced by one head to be deleted or ignored as a consequence of merger of a
higher head. The conclusion seems inescapable, then, that a system of stem-formation
that allows stems of the form (19a), and stems of the form (19–20) more generally, cannot
be the result of the syntactic computational system.
4 Verbal -m- and adjectival -siIn (19–20) above, we saw that transitivizing and intransitivizing suffixes, characterized
as vc and vi respectively, can occur in any of the four logically possible orders following
a root. We have not seen any examples, however, in which the members of an individual
pair of suffixes appear in a given order after one set of roots but in the opposite order
after another set. For example, the suffixes of the sequence -g-e- always occur in that
order regardless of their status as transitivizing or intransitivizing. In fact, there are
three possibilities in that regard: both suffixes can be transitivizing, as in (16d), the first
can be intransitivizing and the second transitivizing, as in yawa-ra-g-e- ‘soften (t)’ (cf.
yawa-ra-g- ‘soften (i)’), or the first can be transitivizing and the second intransitivizing,
as in hisya-g-e- ∼ hisi-g-e- ‘be crushed’ (cf. hisya-g- ∼ hisi-g- ‘crush’). In this section
we will observe two suffixes,one deriving verb stems and the other adjective stems, for
which there are four modes of attachment to a root: direct affixation of each suffix, verbal
suffix preceding adjectival, adjectival suffix preceding verbal, and both orders with the
same root. It will be argued that both the fact that only the outer suffix is interpreted,
parallel with what we saw in §3, and the fact that the relative position of the suffixes is
an idiosyncratic function of the individual root militate against treating the suffixes as
syntactic elements.
Many Japanese roots support both a verb stem in -m-, exemplified in §3, and an adjective stem formed with the suffix -si-. While adjective stems in -si- are not treated
in the DM literature on Japanese derivation, that suffix has a natural DM analysis as a
category-determining little a, where the latter is a stative counterpart of inchoative vi
and causative vc (Marantz 2013: 103). In the examples of (29–30), both suffixes attach
directly to a root, making those examples parallel, as the displayed structure shows, to
Brent de Chene
the verb stems nao-r- and nao-s- that we saw in (5) and (18) (the root of 30 also supports a
stem kuy-i- that is a close synonym of (30b); y deletes before a front vowel in the phrasal
a. suzu-si- [[R]a] ‘cool, refreshing’
b. suzu-m- [[R]vi ] ‘cool off, refresh oneself’
a. kuy-asi- [[R]a] ‘causing chagrin, regret’
b. kuy-am- [[R]vc ] ‘rue, regret’
There are a number of roots supporting both types of stem seen in (29–30), however, for
which the verb stem in -m- is derived from the adjective stem in -si-. This is illustrated in
(31–32) (I take -si- to be suffixal in an otherwise unsegmentable CVCVsi- adjective stem).
a. kuru-si- [[R]a] ‘painful, uncomfortable, difficult’
b. kuru-si-m- [[[R]a]vi ] ‘suffer’
a. kana-si- [[R]a] ‘sad’
b. kana-si-m- [[[R]a]vi ] ‘grieve, sorrow’
And there are roots for which, in contrast, the verb stem in -m-, whether transitive (as
in 33b) or intransitive (as in 34b) serves as the base for derivation of the adjective stem
in -si-:12
a. uto- [[R]a] ‘distant, ill-informed’
b. uto-m- [[R]vc ] ‘shun, ostracize’
c. uto-m-asi- [[[R]vc ]a] ‘unpleasant, repugnant’
a. ita- [[R]a] ‘painful’
b. ita-m- [[R]vi ] ‘be painful; get damaged’
c. ita-m-asi- [[[R]vi ]a] ‘pitiable, pathetic’
Finally, there is at least one root for which both the verb stem in -m- and the adjective
stem in -si- contain both suffixes, in the opposite order in the two cases:
a. tutu-m-asi- [[[R]vc ]a] ‘modest, unpretentious’
b. tutu-si-m- [[[R]a]vc ] ‘be cautious regarding; abstain from’
What conclusions can we draw from the data of (29–35)? First of all, with regard to
interpretation, those examples support the same observation that was made in §3 for
stems of the four types in (19–20), namely that when a stem contains two derivational
suffixes, the inner one is interpretively inert.13 The semantic relations of the two stems
For an English parallel to the three types (29–30), (31–32), (33–34), consider ambigu-ous/ity, duplic-it-ous,
13 While one might imagine for some of the doubly suffixed stems of (31–35) that the interpretation of the
whole depends in some way on that of the inner suffix, there is evidence against this idea in some cases.
With respect to (34), for example, the root-reduplicated adjective itaita-si- ‘pitiable, pathetic’ shows that
the occurrence of that meaning for the stem ita-m-asi- has nothing to do with the inner suffix -m-.
7 Root-based syntax and Japanese derivational morphology
to each other and to the root in (35), for example, are roughly the same as in (29–30),
even though the stems of (35) each contain two suffixes and the stems of (29–30) only
one. This observation, as we have seen, casts doubt on the proposal that the suffixes in
question are syntactic elements.
A parallel argument can be made regarding the relative position of suffixes. (19–20)
have already shown, of course, that if suffixes are divided into transitivizing (“causative”)
and intransitivizing (“inchoative”) types, there are no constraints on their relative order
when two of them occur in the same stem, so that their actual order in particular cases
becomes a function of the individual root. As suggested by the discussion of the suffix
sequence -g-e- at the beginning of this section, though, if we classify suffixes on strictly
distributional grounds, without reference to transitivity value, it is possible to set up
two position classes that will obviate conditioning of suffix order by roots in the great
majority of cases: roughly speaking, the suffixes recognized by the Jacobsen/Volpe segmentation of stems will belong to the outer layer, with the inner layer being composed
of suffixes such as -g-, -m-, -w-, and (transitivity-neutral) -r-.
For the data of (29–35), however, conditioning of suffix order by individual roots is
inescapable. This, then, constitutes a second way, independent of the interpretive inertness of the inner suffix, in which the behavior of -m- and -si- fails to conform to what we
would expect of syntactic elements. Returning to the analogy with auxiliary verbs that
we appealed to in §3 (see 21 above), the positional relations of those two suffixes are as if
the Perfect and the Progressive auxiliaries (say) both appeared adjacent to the stem for
one class of verbs, but the Perfect was formed by placing the Perfect auxiliary outside
the Progressive for a second class of verbs, and the Progressive was formed by placing
the Progressive auxiliary outside the Perfect for a third class. The reason, of course, that
this is difficult to imagine is that we expect unambiguously syntactic elements to appear
in a fixed order with respect to a verbal or nominal stem. Indeed, since the 1990s, a great
deal of work in cartographic syntax (notably Cinque 1999) has developed the idea that
the (hierarchical) ordering of syntactic functional heads is fixed not only internally to a
single language, but universally. From that perspective, the radical failure of Japanese
verbal -m- and adjectival -si- to display a consistent ordering makes it extremely difficult
to view them as syntactic heads.
5 Compositional meanings and semantic change
We have claimed that the syntactic computational system includes no mechanism for
opting out of compositional interpretation, in particular by allowing a higher head to
nullify the interpretation of a lower one. More generally, it seems reasonable to assume
that the compositional interpretation of structures generated by the syntax is automatic,
so that there is no way to block the compositional interpretation of a syntactic constituent.14 We expect it to be true, in other words, that no instance of a syntactically generated structure or construction can idiosyncratically fail to display the compositional
I will assume that this principle is not compromised by the delayed transfer to the interfaces characteristic
of phase-based derivation (Chomsky 2001).
Brent de Chene
semantic interpretation associated with that structure or construction.15 As a result, a
phrase like kick the bucket that is demonstrably generated by the syntax will automatically have the compositional interpretation predicted by its lexical items and its syntactic
structure, independently of whether it has one or more listed interpretations as well. As
a diachronic corollary, we can infer that loss of the compositional interpretation of a
syntactically generated constituent is not a possible change, assuming that the grammar
and the lexicon have remained stable in the relevant respects. Thus, it would not be possible for kick the bucket to lose its compositional interpretation over time, retaining only
the idiomatic one. When a phrase that was once generated by the syntax does have only
a listed interpretation, it is either because the component words have dropped out of the
lexicon, as is probably the case for the phrase to plight one’s troth for most contemporary English speakers, or because the grammar no longer generates phrases of the type
in question, as is the case for the phrase till death do us part.
What is true for manifestly phrasal constituents is true for inflected forms as well. Lexicalization (i.e. idiomatization) of guts in the meaning ‘courage’ and balls in the meaning
‘audacity’ has no effect on the status of those forms as regular plurals as long as the relevant stems and the rules for forming and interpreting plurals are diachronically stable.
In Japanese, many verbal Gerund forms in -te are lexicalized as adverbs: sitagatte, yotte
‘consequently’ (sitagaw- ‘obey’, yor- ‘be due to’), kiwamete, itatte ‘extremely’ (kiwame‘reach, carry to extremity’, itar- ‘reach’). As long as the relevant verb stems remain in
the lexicon and -te remains an inflectional suffix, however, there is no way that these idiomatic meanings can replace the compositional meanings that the forms have by virtue
of their inflectional (ultimately, syntactic) status. The same is true of verbal Conjunctive
forms that have been lexicalized as nouns: nagasi ‘sink’ (naga-s- ‘make flow’), nagare
‘flow, course of events’ (naga-re- ‘flow’).16
If loss of a compositional interpretation is not a possible semantic change, assuming
stability of grammar and lexicon, then demonstrating that the predicted compositional
meaning of a putatively syntactic construction is subject to loss over time will support
the conclusion that the construction in question is not syntactic after all, since if it were,
its compositional meaning should be diachronically stable. In the present section, I will
make this argument with respect to the Japanese lexical causative in -s-, exemplified by
stems like nao-s- ‘cure, repair’, seen in (5b) and (18b) above. Specifically, I will document a
number of cases in which the construction [R[s]] can be shown to have had the predicted
interpretation CAUS(ǁRǁ) (ǁRǁ the interpretation of R) originally but later to have lost that
interpretation in spite of the fact that ǁRǁ itself has remained constant.
Correspondingly, establishing that some phrase P is a counterexample to this principle will require (a)
displaying P’s syntactic structure; (b) displaying the rule of interpretation associated with that structure;
and (c) showing that P idiosyncratically lacks the predicted interpretation.
The semantics of these nouns has been treated in the DM literature since Volpe (2005) as involving selection
of root allosemes by a noun-forming suffix (“special meanings of the root triggered across the little v
head” (Marantz 2013: 107). The extreme semantic distance that separates many of the nouns from their
corresponding roots (abundantly documented by Volpe), however, makes idiom-formation a more plausible
basis for the nominal meanings than alloseme choice (for the distinction between the two mechanisms, see
Marantz 2013: 105).
7 Root-based syntax and Japanese derivational morphology
As a first example, consider the stem yurus- ‘allow, forgive’. In Old Japanese (see
Omodaka et al. 1967), the primary meaning of this stem is ‘slacken (t)’, with secondary
meanings ‘let go of’; ‘allow, comply with, tolerate’; and ‘forgive, exempt’. Yurus-, in other
words, is historically the causative in -s- on √yuru ‘slack’ (see ⁇) above), a root that in
modern Japanese underlies the adjective stem yuru- ‘slack’, the nominal adjective yuruyaka ‘slack, gradual’, and the verb stems yuru-m- ‘slacken (i)’ and yuru-m-e- ‘slacken
(t)’. As is clear from these four stems, the root has been completely stable semantically
over thirteen centuries, and the same can be assumed for causative -s-. There is no trace
in the modern meaning of yurus-, however, of the original concrete primary meaning
‘slacken’. That meaning, in other words, has been completely replaced by the originally
secondary or extended meanings ‘allow’ and ‘forgive’. If yuru-s- had been a syntactic
construction, with the meaning ‘slacken (t)’ the compositional result of a semantic rule of
interpretation, this replacement should have been impossible, just as we have suggested
that it would be impossible for kick the bucket to lose its compositional meaning and
retain only the idiomatic one.
The history of the stem itas- ‘do (humble)’ is broadly parallel. In Old Japanese, it is
the causative corresponding to itar- ‘reach a limit’, as explicitly noted in Omodaka et
al. (1967), and thus means ‘bring to a limit’. In the modern language, while intransitive
itar- has retained its original meaning, itas- is for the most part, bleached of concrete
content, simply a suppletive humble variant of suru ‘do’. A third case in which a s-stem
has lost a putatively compositional causative meaning involves konas- ‘deal with, take
care of; be skilled at’, whose primary meaning was originally ‘break up, pulverize’ and
which is based historically on ko ‘powder’ (Ono, Satake & Maeda 1974). Like many other
original monosyllables, ko has been replaced as a freestanding noun by a bisyllabic form,
in this case kona, which is attested starting around 1700. The only serious proposal for
the origin of kona (see NKD) appears to be that it is a backformation based on konas-. If
the backformation theory is correct, kona and konas- were unquestionably isoradical at
the relevant point in time, so that konas- consisted of √kona ‘powder’ plus causative -s-.
Today, however, while the root noun remains in the language, the meaning ‘break up,
pulverize’ for the verb is extinct.17
Two further stems in -s- for which the predicted causative meaning appears to have
been lost over time are hatas- ‘carry out, perform, accomplish’ and kuras- ‘make a living; live, spend (time)’. The roots appear in the zero-derived noun hata ‘edge, perimeter;
outside’ and the zero-derived adjective stem kura- ‘dark’, respectively, and are semantically identifiable in the intransitives hate- ‘end (i)’ and kure- ‘darken (day), end (i)’ (for
the a ∼ e alternation, see note 11 above). The expected primary meaning ‘end (t)’ of
hatas- appears in the gloss ‘bring to a conclusion’ in Omodaka et al. (1967); for kuras-,
similarly, Omodaka et al. record the expected primary meaning ‘spend the time until
evening’ (i.e. ‘let the day darken’). In both cases, however, this compositional meaning
is absent from the modern stems, neither of which stands in a purely causative relation
to the corresponding intransitive or to the root. The meaning of hatas-, as the above
While dictionaries retain examples like tuti o konasu ‘break up dirt (clods)’, the speakers I have consulted
deny knowledge of such a usage.
Brent de Chene
definition indicates, inherently includes an element of purposive activity (carrying out
a command, achieving a goal, fulfilling an obligation) that is absent from that of hate-.
While the semantic difference between kuras- and kure- is more subtle, the basic fact
preventing the former from functioning as the causative of the latter is that, unlike kure(‘come to an end’), kuras- (‘spend (time)’) is atelic. Both hatas- and kuras-, then, like
yurus-, itas-, and konas-, are cases in which the predicted interpretation CAUS(ǁRǁ) of
the construction [R[s]] has been lost over time.
In this section, we have seen an argument against the syntactic derivation of Japanese
verb stems based on semantic change, using causatives in -s- as a representative stemtype. It goes without saying, we should emphasize, that perhaps the most common type
of semantic change, the addition of idiomatic or extended meanings, does not count
against the hypothesis of syntactic generation: as is well known, linguistic units of any
size can be idiomatized, with the tendency to undergo idiomatization inversely proportional, roughly speaking, to size (Di Sciullo & Williams 1987: 14). But loss of a putatively
compositional meaning, we have claimed, does count against syntactic generation, because there is no reason to take the compositional interpretation of syntactic structure
to be anything but automatic and exceptionless. In order for a compositional meaning
M to be lost, the syntactic structure underlying it would first have to be exempted from
compositional interpretation, with M being lexicalized at the same time; M could then
be lost from the lexicon. If this sequence of events is impossible because exemptions of
the required type are never granted, however, a putatively compositional meaning that
is in fact subject to loss cannot have been based on a syntactic derivation in the first
6 Conclusion
Above, I have attempted to evaluate the proposal that the derivational suffixes that create transitive and intransitive verb stems in Japanese are syntactic heads, in particular
varieties of little v. Crucial evidence in this regard has come from identifying an inner
layer of derivational suffixation (-g-, -m-, etc.) in addition to the well-known outer layer
whose main members are -r-, -s-, -re-, -se-, -e-, -i-, and zero, since this has allowed us to
raise the question of how two derivational suffixes interact when they occur together in
the same stem. We saw in §3 that in such a case, the inner suffix is always inert for purposes of argument structure and semantic interpretation, casting doubt on the position
that the suffixes are syntactic elements. In §4, we saw that the same is true for combinations of the verbal suffix -m- and the adjectival suffix -si-, with the added complication
that the order in which those two suffixes occur is an idiosyncratic function of the root.
Finally, in §5, we argued, without reference to suffix sequences, that the combination
of a root and a transitivity-determining suffix, taking causative -s- as a representative
example, cannot be a syntactic construction because its putatively compositional interpretation is unstable over time. All the evidence we have seen, then, points toward the
conclusion that the derivational suffixes under consideration are not syntactic elements.
Equivalently, if one wishes in the face of this evidence to generate Japanese verb and ad-
7 Root-based syntax and Japanese derivational morphology
jective stems syntactically, one will require relaxation of otherwise well-motivated constraints on structure-building and interpretation precisely for the domain of the stem.
As suggested at the outset, our conclusions in this regard support Anderson’s (1982: 594)
position on the place of morphology in the grammar: derivation is pre-syntactic, and the
units of lexical storage are inflectable stems; inflection, in contrast, is the post-syntactic
spellout of morphological elements and morphosyntactic properties that are treated by
syntactic operations.
The conclusion that Japanese derivational suffixes, in contrast with suffixes like the
Passive and the productive Causative, are not syntactic elements is supported at a more
impressionistic level by the fact that, as is easily confirmed, the two sets of suffixes differ
sharply in their degree of regularity, both formal and semantic. Formally, while variation in the shape of the Passive suffix -(r)are- is limited to phonologically conditioned
alternation of r with zero at the left edge, and variation in the shape of the Causative suffix -(s)as(e)- is limited to phonologically conditioned alternation of s with zero at the left
edge and non-phonological alternation of e with zero at the right, variation in the realization of what under a DM analysis will be vi and vc is highly unconstrained, with multiple
unrelated allomorphs for each of the suffixes and almost complete overlap between the
two allomorph sets. Semantically, while the meaning of Passivepassive stems in -(r)areand (apart from occasional idioms) Causative stems in -(s)as(e)- is both regular and relatively straightforward to characterize, the meaning of stems in vi and vc is in most
cases multiply polysemous and highly idiosyncratic; the glosses we have given above,
while aiming at a marginal increase in accuracy over the labels in Jacobsen (1992) and
Volpe (2005), in many cases only scratch the surface of the problem of specifying stem
meaning. With regard to semantics, it should also be remembered that, as we noted in
§2, morphological analysis internal to the stem proceeds on the basis of an unredeemed
promissory note regarding the criterion for isoradicality and that equally serious questions arise about how the meaning of transitivity suffixes is to be specified, given the
apparent semantic overlap between transitivizing and intransitivizing morphology.
If Japanese verb and adjective stems are not, then, created by the syntactic computational system, how should we conceive of their structure and, crucially, the knowledge
that speakers have about that structure? Broadly speaking, there are two types of answer that could be given to this question. On one of them, derivational morphology of
the type we have seen here would be the result of a combinatory system roughly parallel
to syntax but less regular both in terms of the hierarchical relationships holding among
grammatical elements and the semantic interpretation of complex structures. From the
standpoint of theoretical parsimony, of course, this would seem like an unattractive proposal; surely, if possible, we would prefer to maintain that the language faculty involves
a “single generative engine” (Marantz 2001; 2005). Viewing language as a biological object, however, there would appear to be no grounds for excluding a priori the possibility
that our linguistic capacities include a combinatory stem-formation module of the sort in
question. In evolutionary terms, such a module might have provided a vastly expanded
repertory of named concepts in advance of the emergence of a fully regular and productive syntax, representing a sort of half-way house on the road to discrete infinity.
Brent de Chene
The second type of answer that could be given to the question of the form taken by
speaker knowledge of the relations among isoradical stems, assuming that those relations are not mediated by the syntactic computational system, is that that knowledge is
frankly non-generative – that is, non-combinatory. In this case, all stems will be lexically
listed, with relations among them captured by redundancy rules, for example, those of
the type pioneered by Jackendoff (1975) (see also Jackendoff 2002: 53). What is unsatisfying about this type of answer is that it provides no insight into why derivational
morphology should exist at all – why, that is, stems (setting aside compounds) are not
all atomic. While we have seen evidence that at least some derivational morphology
cannot be syntactic, then, there is no unambiguously attractive alternative account of
the structure of speaker knowledge in this area. As a result, the place of derivational
morphology in our linguistic competence remains very much an open question.
I would like to express my appreciation to Takayuki Ikezawa, in conversations with
whom the idea for this paper emerged. I am also grateful to Kunio Nishiyama and Yoko
Sugioka for comments on a presentation of some of this material at the 152nd meeting
of the Linguistic Society of Japan (Tokyo, June 2016). Finally, I am indebted to several reviewers for comments that have resulted in a number of clarifications and improvements.
Remaining errors of fact or interpretation are my responsibility.
(second or perfective) conjunctive
Anderson, Stephen R. 1982. Where’s morphology? Linguistic Inquiry 13(4). 571–612.
Chomsky, Noam. 2001. Derivation by phase. In Michael Kenstowicz (ed.), Ken Hale: A
life in language, 1–52. Cambridge, MA: MIT Press.
Chomsky, Noam. 2008. On phases. In Robert Freidin, Carlos P. Otero & Maria Luisa
Zubizarreta (eds.), Foundational issues in linguistic theory, 133–166. Cambridge, MA:
MIT Press.
Cinque, Guglielmo. 1999. Adverbs and functional heads. Oxford: Oxford University Press.
Di Sciullo, Anna-Maria & Edwin Williams. 1987. On the definition of word. Cambridge,
MA: MIT Press.
Embick, David & Alec Marantz. 2008. Architecture and blocking. Linguistic Inquiry 39(1).
Halle, Morris & Alec Marantz. 1993. Distributed morphology and the pieces of inflection. In Kenneth Hale & Samuel J. Keyser (eds.), The view from Building 20: Essays in
linguistics in honor of Sylvain Bromberger, 111–176. Cambridge, MA: MIT Press.
7 Root-based syntax and Japanese derivational morphology
Harley, Heidi. 2008. On the causative construction. In Shigeru Miyagawa & Mamoru
Saito (eds.), The Oxford handbook of Japanese linguistics, 20–53. Oxford: Oxford University Press.
Harley, Heidi. 2012. Lexical decomposition in modern syntactic theory. In Markus Werning, Wolfram Hinzen & Edouard Machery (eds.), The Oxford handbook of compositionality, 328–350. Oxford: Oxford University Press.
Hopper, Paul J. & Sandra A. Thompson. 1980. Transitivity in grammar and discourse.
Language 56. 251–299.
Jackendoff, Ray. 1975. Morphological and semantic regularities in the lexicon. Language
51(3). 639–671.
Jackendoff, Ray. 2002. Foundations of language. Oxford: Oxford University Press.
Jacobsen, Wesley M. 1982. Transitivity in the Japanese verbal system. Bloomington: Indiana University Linguistics Club.
Jacobsen, Wesley M. 1992. The transitive structure of events in Japanese. Tokyo: Kurosio
Kuo, Jonathan Cheng-Chuen. 2015. Argument alternation and argument structure in symmetrical voice languages: A case study of transfer verbs in Amis, Puyama, and Seediq.
University of Hawai‘i at Manoa Dissertation (pre-defense draft).
Marantz, Alec. 1997. No escape from syntax: Don’t try morphological analysis in the
privacy of your own lexicon. University of Pennsylvania Working Papers in Linguistics
4(2). 201–225.
Marantz, Alec. 2001. Words. Handout for invited lecture. 20th West Coast Conference on
Formal Linguistics, University of Southern California.
Marantz, Alec. 2005. Generative linguistics within the cognitive neuroscience of language. The Linguistic Review 22(2–4). 429–445.
Marantz, Alec. 2007. Phases and words. In Sook-Hee Choe (ed.), Phases in the theory of
grammar, 191–222. Seoul: Dong In.
Marantz, Alec. 2013. Locality domains for contextual allomorphy across the interfaces. In
Ora Matushansky & Alec Marantz (eds.), Distributed Morphology today, 95–115. Cambridge, MA: MIT Press.
Nihon kokugo daiziten dainihan hensyuu iinkai/Syoogakukan kokugo ziten hensyuubu [Great dictionary of the Japanese language second edition editorial committee/Syoogakukan Japanese dictionary editorial division] (ed.). 2000–2002. Nihon kokugo daiziten [Great dictionary of the Japanese language]. 2nd ed. Tokyo:
NKD (Nihon kokugo daiziten) = Nihon kokugo daiziten dainihan hensyuu iinkai/
Syoogakukan kokugo ziten hensyuubu 2000–2002.
Omodaka, Hisataka, Toru Asami, Teizo Ikegami, Itaru Ide, Haku Ito, Yoshiaki Kawabata,
Masatoshi Kinoshita, Noriyuki Kojima, Atsuyoshi Sakakura, Akihiro Satake, Kazutami Nishimiya & Shiro Hashimoto (eds.). 1967. Zidaibetu kokugo daiziten. Zyoodaihen
[Great historical dictionary of Japanese. Nara period]. Tokyo: Sanseidoo.
Brent de Chene
Ono, Susumu. 1953. Nihongo no doosi no katuyookei no kigen ni tuite [On the origin of
the inflected forms of Japanese verbs]. Kokugo to Kokubungaku [Japanese Language
and Literature] 350. 47–56.
Ono, Susumu, Akihiro Satake & Kingoro Maeda (eds.). 1974. Iwanami kogo ziten [Iwanami
historical dictionary of Japanese]. Tokyo: Iwanami Syoten.
Pylkkänen, Liina. 2008. Introducing arguments. Cambridge, MA: MIT Press.
Volpe, Mark. 2005. Japanese morphology and its theoretical consequences: Derivational
morphology in Distributed Morphology. New York: Stony Brook University Dissertation.
Chapter 8
Morphological complexity and Input
Michael Hammond
University of Arizona
In this paper, we examine morphological complexity through the lens of Input Optimization.
We take as our starting point the dimensions of complexity proposed in Anderson (2015).
Input Optimization is a proposal to account for the statistical distribution of phonological
properties in a constraint-based framework. Here we develop a framework for extending Input Optimization to the morphological domain and then test the morphological dimensions
Anderson proposes with that framework.
The dimensions we consider and the framework we develop are both supported by empirical
tests in English and in Welsh.
1 Introduction
Anderson (2015) lays out a number of dimensions of morphological complexity, ways
that we might evaluate how complex different morphological systems are, e.g. number
of morphemes in the system, complexity of principles governing combinations of morphemes, complexity of exponence, complexity of allomorphy, etc.
These are clearly the right kinds of dimensions for evaluating the complexity of morphological systems, and we might be inclined to use them as part of a typology of morphology. However, if we adopt these as our dimensions for calculating complexity, what
follows? As I learned from Steve Anderson years ago in graduate school, typology without implications is bad typology. (For discussion, see, for example, Anderson 1999.)
In this paper, I consider the implications of these dimensions of morphological complexity for the theory of Input Optimization (Hammond 2013; 2014; 2016). This theory
develops a notion of phonological complexity which languages “attempt” to minimize
statistically. To the extent that different phonological representations are complex, they
are under-represented statistically. This complexity shows up in the markedness of surface representations and in the complexity of input–output mappings. For example,
marked segments or syllable structures are under-represented compared with their less
Michael Hammond. 2017. Morphological complexity and Input Optimization. In
Claire Bowern, Laurence Horn & Raffaella Zanuttini (eds.), On looking into words
(and beyond), 155–171. Berlin: Language Science Press. DOI:10.5281/zenodo.495444
Michael Hammond
marked counterparts. In addition, outputs that are distinct from their inputs are underrepresented with respect to outputs that are identical with their inputs.
Interestingly, there are morphological effects as well, effects that sometimes work
in the opposite direction. For example, initial consonant mutation in Welsh causes a
mismatch between output and input, but is over-represented. I argue that this is because
phonological complexity includes morphological mapping. Specifically, to the extent
that morphological distinctions are not made in the surface form, a representation is
more complex. This is formalized in OT-based terms using a constraint deriving from
work by Kurisu (2001).
This general approach is supported by the facts of haplology, e.g. English adjectives in
-ly like weekly not getting double-marked with adverbial -ly and plurals like kings not getting double-marked with genitive -s. The absence of double-marking means that a morphological distinction is not made on the surface; thus these cases are under-represented
as expected.
These morphological cases beg the larger question: should phonological complexity
be generalized further? Should there be a more general notion of morphological complexity, built on dimensions of the sort cited above, where forms that are more complex
morphologically are statistically under-represented? In this paper, I pursue just this
course, formalizing a notion of morphological complexity and then testing it with cases
from English and Welsh with respect to the dimensions of morphological complexity
identified in Anderson (2015).
The organization of this paper is as follows. I first review some of the dimensions of
complexity presented in Anderson (2015). I then outline the theory of Input Optimization
and a framework for a constraint-based theory of morphology that we can assess Input
Optimization with respect to. With these in hand, we then consider the predictions made
by the Input Optimization framework and turn to the English and Welsh data.
2 Dimensions of complexity
Anderson (2015) discusses a number of dimensions of morphological complexity and we
will not review them all here.
We explicitly set aside those systemic dimensions that cannot distinguish options
within a language. For example, Anderson cites the number of elements in the morphological system as a measure of its complexity. Thus, for example, if one language has
one way of marking noun plurals and another has ten ways, we might think of the first
as less complex. Input Optimization makes no predictions about systemic differences
like these, as we will see in the next section, so we don’t consider them any further.
Anderson cites the number of morphemes in a word as a dimension of complexity.1
This can be taken in several ways. One possibility is that one might compare across
different languages, which seems to be Anderson’s intent. Another possibility though
This, of course, begs the question of what is a morpheme, an issue at the forefront of much of Steve’s own
8 Morphological complexity and Input Optimization
would be to compare across words in the same language and we investigate this possibility below.
Another dimension Anderson identifies is whether the morphemes present in a word
depend on each other in some way. We might think of this in two ways. Some morpheme
may only occur when “licensed” by some other. Gender in Spanish is an example of
this. If gender is marked on a noun, then that gender marking must be present in the
plural as well, e.g. mes+a+s ‘tables’ table + feminine + plural. Another example might be
verbal theme vowels in Romance; person/number marking depends on the presence of
the theme vowel.
The other side of this coin would be cases where the presence of some morpheme
blocks another. Haplology is an example of this. For example, adverbial -ly in English
cannot occur on an adjective that already ends in -ly. Thus we have happily, but not
∗ weeklyly.
Anderson also distinguishes among morpheme types in terms of complexity. Simple
prefixation or suffixation is less complex than circumfixion or infixation. Presumably,
non-concatenative morphology like templatic operations, ablaut, umlaut, truncation are
also more complex.
Lastly, Anderson cites the complexity of allomorphy as an instance of general morphological complexity. We take this to mean that a system is more complex when there
is more allomorphy. We interpret allomorphy as generously as possible to include cases
where the phonology seems to be involved, say, in the different pronunciations of the
English plural -s as [s, z, əz], but also plurals that differ on some other basis, e.g. geese,
criteria, sheep, etc.
Anderson treats some other possibilities as well, but the ones above are quite simple
and can be examined within a single language. We list them together below.
1. Number of morphemes in a word
2. Principles of morphological combination, e.g. scope, haplology, etc.
3. Complexity of exponence, e.g. circumfixes, infixes, etc.
4. Complexity of allomorphy
In the following section, we review the Input Optimization proposal and sketch out the
predictions it makes for these. Our interest in Input Optimization is that it provides a
mechanism by which we can assess the dimensions of morphological complexity we’ve
just considered.
3 Input Optimization
The problem that Input Optimization addresses is that certain phonological configurations occur less often than we might otherwise expect. For example, if we look at the
distribution of stress on two-syllable adjectives, we see that adjectives with final stress
Michael Hammond
like alert [əlɹ̩t́ ] or opaque [òpʰék] are less frequent overall. Strikingly, both are even less
frequent when they occur prenominally.
Hammond (2013) argues that this effect is driven by the English Rhythm Rule (Liberman & Prince 1977; Hayes 1984). Certain stress configurations in English are avoided
by shifting a primary stress leftward onto a preceding secondary. Thus we have alternations like Mìnnesóta vs. Mínnesòta Míke; thìrtéen vs. thírtèen mén; etc. When an adjective
with final stress occurs in prenominal position, the relevant configuration is quite likely
to occur. In addition to following context, there is a restriction on preceding context.
With an adjective like òpáque, stress shift leftward is possible because of the preceding
stress, e.g. ópàque stóry, but with an adjective like alért, such a shift is impossible and the
offending configuration must surface, e.g. alért pérson. Both kinds of cases are statistically under-represented in English. Specifically, these configurations arise significantly
less frequently than we might expect based on the overall distribution of adjectives with
these stress patterns.
Input Optimization is a proposal to account for statistical skewings like these that
occur in the phonologies of languages. The idea is developed in Hammond (2013; 2014;
2016). The basic idea is that markedness and faithfulness violations are avoided in the
phonology so as to reduce the complexity of the phonological system. Input Optimization is a generalization of the notion of Lexicon Optimization Prince & Smolensky (1993):
Lexicon Optimization:
Suppose that several different inputs I 1 , I 2 , . . . , In when parsed by a grammar G
lead to corresponding outputs O 1 , O 2 , . . . , O n , all of which are realized as the
same phonetic form Φ—these inputs are all phonetically equivalent with respect
to G. Now one of these outputs must be the most harmonic, by virtue of incurring
the least significant violation marks: suppose this optimal one is labelled O k .
Then the learner should choose, as the underlying form for Φ, the input Ik .
The idea is that if there are multiple ways to produce an output form consistent with the
facts of a language, the learner chooses the input that produces the fewest constraint
violations. There are no empirical consequences to Lexicon Optimization by itself. In
fact, it is defined to apply only when there are no consequences.
To refine this into something we can use, we define a notion of Phonological complexity
that applies to individual input–output pairings, but also to entire phonological systems.
(The basic logic of this is that the complexity of a phonological system is proportional
to the number of asterisks in its tableaux.)
We define the output/surface forms of a language as a possibly infinite set of forms.
(2) O = {O 1 , O 2 , . . . , O n , . . .}
Each output form has a corresponding input:
(3) I = {I 1 , I 2 , . . . , In , . . .}
The phonology is comprised of a finite sequence or vector of constraints:
8 Morphological complexity and Input Optimization
(4) C = ⟨C 1 , C 2 , . . . , Cn ⟩
Any input–output pairing, (Ii , O i ), then defines a finite vector of violation counts, some
number of violations for each constraint earned by the winning candidate for that input.
⟨nC1 , nC2 , . . . , nCn ⟩
With these notions, Phonological Complexity (PC) is defined as follows:
(6) Phonological Complexity (PC)
The phonological complexity of some set of forms is defined as the vector sum of
the constraint violation vectors for surface forms paired with their respective
optimal inputs.
To produce a relative measure of PC given some set of n surface forms, divide the
PC score for those forms by n.
Hammond (2016) exemplifies this with a hypothetical example of nasal assimilation.
Imagine we have the forms in (7) we wish to compute the PC for. Given the inputs provided in column 2, we have the constraint violations for winning candidates in columns 3
and 4.
/on pi/
/an ba/
/un bo/
/en do/
/on ta/
/un ti/
/an ku/
/in ga/
/on ke/
om pi
am ba
um bo
en do
on ta
un ti
aŋ ku
iŋ ga
oŋ ke
The relative complexity of this first system is: ⟨0, 6⟩/9 = ⟨0, .66⟩. We can compare the
system in (7) with the one below in (8). Here we have a different array of output forms,
but the same logic for inputs and constraint violations.
/on pi/
/an ba/
/en do/
/on ta/
/un ti/
/in di/
/an ku/
/in ga/
om pi
am ba
en do
on ta
un ti
in di
aŋ ku
iŋ ga
Michael Hammond
The relative complexity of the second system is: ⟨0, 4⟩/8 = ⟨0, 0.5⟩, less than the first.
As argued by Hammond (2016), this notion extends obviously to weighted constraint
systems. For example, in a system with strict ranking, ⟨0.1, 0.4⟩ is more complex than
⟨0, 0.5⟩.
The proposal then is that all phonological systems are skewed to be less complex.
Input Optimization
All else being equal, phonological inputs are selected that minimize the
phonological complexity of the system.
Note that (9) alters the frequency of input–output pairings and does not change the
input–output mapping of any particular form. For example, this principle prefers (8)
to (7), though both systems contain the same pairings. The difference is in the relative
frequency of the pairings that occur.
Our goal in this paper is to see if it is profitable to extend this system to include morphology. In point of fact, Hammond (2016) addresses this question partially in response to
statistical effects in Welsh. In particular, Welsh initial consonant mutation is statistically
over-represented when, based on what we have seen so far, we might have expected the
Consonant mutation in Welsh refers to a set of phonological changes that apply to
initial consonants in specific morpho-syntactic contexts. For example, the Soft Mutation
makes the following changes:
Input Output
Input Output
There are many contexts where this occurs, e.g. after certain prepositions, direct object
of an inflected verb, after certain possessives, feminine singular nouns after the article,
etc. The following figure gives some examples after the preposition i [i] ‘to’.
i ben
i gath
i fis
i nai
i siop
[i bɛn]
[i gaθ]
[i vis]
[i naj]
[i ʃɔp]
‘to a head’
‘to a cat’
‘to a month’
‘to a nephew’
‘to a shop’
8 Morphological complexity and Input Optimization
The chart above also includes examples of non-mutating consonants. Note that words
with these occur in mutation contexts with no change.
The Input Optimization framework would seem to predict that mutation should be
under-represented. After all, mutation entails a faithfulness violation and, all else being
equal, the system is less complex to the extent that such violations are avoided. This,
however, is not what occurs. Instead, we get over-representation in mutation contexts.
Words that begin with consonants that can mutate are over-represented in mutation
contexts compared with words that begin with consonants that do not mutate.
To capture this, Hammond (2016) proposes the (revised) Realize Morpheme constraint
(12). This is a slight revision of a constraint that Kurisu (2001) motivates on other (nonstatistical) grounds. This constraint basically militates for the expression of morphological information.
Realize Morpheme (revised) (RM’)
Let α be a morpheme, β be a morphosyntactic category, and F(α) be the
phonological form from which F(α+β) is used to express a morphosyntactic
category β. Then RM’ is satisfied with respect to β iff F(α+β) , F(α)
With this in hand, the reason why Welsh mutation is over-represented is to reduce
phonological complexity by minimizing violations of RM’.
The RM’ constraint is also invoked by Hammond (2016) to account for haplology in
English. We’ve already cited the fact that forms like ∗ weeklyly are blocked. Similarly,
we find overt marking of the genitive in English does not occur on plural forms marked
with -s; the genitive plural of cat is cats’, not something like ∗ catses. Both kinds of cases
are statistically under-represented in English: they are avoided to minimize violations
of RM’.
While RM’ (12) does what’s required, it begs the question of whether a more general
version of PC is appropriate. In other words, beyond the effects of RM’, do we expect
Input Optimization to apply to morphology?
4 Constraint-based morphology
To assess this, we need a constraint-based theory of morphology. There have been a
number of proposals over the years for how to deal with morphology generally in an
OT-like framework. The earliest we know of are Russell (1993; 1995); Hammond (2000),
but see Aronoff & Xu (2010); Xu & Aronoff (2011) for more recent and fuller proposals.
A full-on theory of this sort is well beyond the scope of this paper, but let’s lay out what
such a theory might look like, at least in sufficient detail so we can assess whether it
makes the right predictions about Input Optimization.
Let us assume that morphology—like phonology—is a constraint-based system mapping inputs to outputs. Inputs are denuded of any morphological marking, but have sufficient featural information so that we can evaluate whether morphologically marked
Michael Hammond
candidate forms satisfy relevant constraints. For example, we might imagine that plural
marking in English comes about by taking a stem marked [+plural] and adding various
affixes or performing other operations that do or do not express that feature. The idea is
that the syntax provides a featurally complex object that the morphology can then interpret. Morphological operations like affixation, reduplication, mutation, etc. add features
which do or do not match those required by the syntax. Following is a schematic partial
tableau to give a sense of this.
] [
] [
We would want constraints that force the correct morphological operation to take
place. Presumably there would be one or more constraints that enforce a correspondence between the features required by the stem and the features offered by any affixes
or other changes; to the extent that those don’t match, we would have violations. For
convenience, let’s call this Features (Fs). The RM’ constraint above, or constraints that
get the same effects, should fall in this class.
We also need constraints that militate against gratuitous morphological operations.
Some of this might be achieved by featural correspondence imposed by Fs, but we surely
need something to account for the relative markedness of morphological operations generally. Perhaps something like this:
∗ Ablaut
≫ ∗ Infix ≫ ∗ Prefix ≫ ∗ Suffix
The basic idea is to posit constraints that militate against any morphological operation.
These constraints are ranked with respect to each other, presumably in a universal fashion. This hierarchy would then be interleaved with the Fs constraint. For example, we
might have:
∗ Ablaut
≫ ∗ Infix ≫ Features ≫ ∗ Prefix ≫ ∗ Suffix
The effect of such a ranking is that the featural needs of a stem can be met by prefixation
and suffixation, but not by other operations.
This system is woefully incomplete and, in its present form, cannot do justice to the
full range of effects we see in morphological systems. See, for example, Anderson (1982;
8 Morphological complexity and Input Optimization
1992). It is, in some ways, quite similar to these proposals in treating affixation as an instance of more general morphological operations that interpret syntactically-motivated
features. However, our goal here is not to develop a full-on constraint-based morphological theory. Rather, the point is to build enough of such a theory so that we can test
Input Optimization with respect to the dimensions of morphological complexity identified above.
Let’s now return to our dimensions and consider one by one what our theoretical
skeleton in conjunction with Input Optimization predicts. First, we have the number of
morphemes. All else being equal, the system certainly as developed militates for as few
morphemes, or other morphological operations, as possible. Additional morphology
entails violations of the constraints in (14) and Input Optimization predicts these should
be avoided statistically.
The second dimension of complexity refers to principles of morphological combination. The system we’ve developed says nothing (so far) about the licensing side of this,
but it does address morphological haplology. To the extent that haplology occurs, it entails violations of RM’ (12) and of Fs. Previous work cited above has already established
that Input Optimization applies in these cases.
The third dimension is the complexity of exponence, that certain morphological operations are intrinsically more complex than others. This is captured by the ranking,
e.g. in (14). We expect morphologies to be statistically skewed against violations of the
higher-ranked constraints.
The fourth dimension is complexity of allomorphy, allomorphy that is a consequence
of phonology or morphophonology like English plural [s, z, əz], but also plurals that
differ on some other basis, e.g. geese, criteria, sheep, etc. The phonological cases fall
under the core Input Optimization proposal. In fact, Hammond (2013) shows statistical
skewing for English plural and past allomorphy in just the expected directions. The other
case cited is also accommodated by the proposal. Internal modifications like geese or
truncation+suffixation like criteria violate higher-ranked constraints than simple plural
suffixation; hence they should exhibit under-representation. Similarly, plurals with no
change like sheep should violate RM’ and Fs and be under-represented.
Summarizing, a constraint-based morphological theory of the sort sketched out, in
conjunction with Input Optimization, makes the following predictions:
1 Words should have fewer morphemes.
2 Haplology should be avoided. (This has already been established by Hammond 2016.)
3 More marked morphological operations (per the hierarchy above) should be
4 Morphophonology should be avoided. (This has already been established by
Hammond 2013.)
5 Ablaut, umlaut, truncation, etc. should be avoided. (This is essentially the
same as #3 above.)
6 Zero-marking should be avoided.
Michael Hammond
We must therefore examine #1, #3/5, and #6 empirically. In the next sections, we look
at all three cases with data from English and Welsh.
5 Number of morphemes
The first prediction of Input Optimization applied to our toy constraint-based theory of
morphology is that a form is more complex if it has more morphemes. This is a bit tricky
to test. In many cases, having fewer morphemes is not necessarily the less complex
option. For example, consider the plural form sheep which lacks an overt plural suffix.
Is this less complex than a form like cat+s? Probably not. The most reasonable analysis
given the framework above is that the plural sheep surfaces with an undischarged plural
feature. On that view, it is not clearly less complex than a form like cat+s.
We might also think of strong verb forms like spoke, as compared with look+ed. Here,
however, it would be a mistake to view spoke as having fewer morphemes than look+ed.
Rather, there is some operation, perhaps mostly lexical, for creating or selecting strong
verb forms when available. Presumably, this would add to the complexity of spoke.
To find a case without these alternative analyses, we turn to Welsh plurals. Welsh has
a number of ways of forming plurals. For example:
[kʰənɬɨń jaɰ]
Note that there are different suffixes and stem changes.
There is another class of nouns, however, where it is the singular that is marked rather
than the plural. The singular is always marked with either -yn in the masculine gender
or -en in the feminine gender. For example:
8 Morphological complexity and Input Optimization
[mɔrgrɨǵ ɨn]
[mɛvɨś ɛn]
There are some blended cases as well, where nouns marked for the singular co-occur
with stem changes or take plural suffixes as well. For example:
The existence of the pairs where the singular is marked instead of the plural allows us
to test the number of morphemes prediction without the problems of the English cases
above.2 On one hand, we take nouns which mark the plural with -(i)au, the most frequent
plural suffix, and no associated stem changes. On the other, we take nouns which mark
the singular with -en or -yn, and no associated stem changes or plural marking. In other
words: problem/problemau, etc. vs. coeden/coed, etc. What we’re interested in is whether
there is a difference in the relative frequency of singular and plural forms in the two
classes as a function of whether the form has an extra morpheme. Since we have both
types of marking in Welsh, we can do this independent of the relative frequency of
singulars and plurals in the language.
Individual lexical items have different frequencies of occurrence, so we must equalize
for this. We therefore take the ratio of singular to plural as a measure of the relative
frequency of singular and plural. Since this is a ratio, it abstracts away from the overall
frequency of each pair.
One might counter that it’s possible to treat these as instances of subtractive morphology. There are two
arguments against this. First, the singulatives are always marked with the suffixes -yn or -en (depending
on gender). Second, there are cases where nouns end in these phonetic sequences where they are not suffixes. In these cases, normal plural formation occurs. For example: emyn/emynau ‘hymn’, terfyn/terfynau
‘boundary’, ffenomen/ffenomenau ‘phenomenon’, awen/awenau ‘inspiration, muse’, etc.
Michael Hammond
For this investigation, we use the CEG corpus (Ellis et al. 2001). This is a tagged written
corpus of 1223501 words. For each word form, it also includes lemmas, so it is possible
to determine singular–plural pairs fairly easily. In this corpus, we find 885 distinct pairs
where the plural is marked with -(i)au and 41 distinct pairs where the singular is marked
with -en or -yn. (As above, in both cases, we exclude pairs where stem changes are
When the plural is marked, the ratio of singulars to plurals is 11.08; when the singular
is marked, the ratio is 1.26. Singulars greatly outnumber plurals that are marked, but
singulars occur far less frequently when they are marked instead. This difference is
significant: t(920.287) = −8.267, p < .001. This is consistent with the hypothesis that
forms with more morphemes are more complex.
6 Marked morphological operations
Let’s now consider the question of whether more marked morphological operations are
under-represented. If they are, this would be consistent with Anderson’s typology and
Input Optimization.
To test this, let’s look at the distribution of plurals in English using the tagged Brown
corpus (Kučera & Francis 1967). The Brown corpus is a fairly old written corpus of 928181
words. The advantage of using it here is that it is tagged, so identification of singular and
plural nouns is relatively easy, and it is widely used and available.
Focusing on plural nouns, we can separate them into regular plurals marked with -s
vs. other plural forms, e.g. men, stigmata, radii, oxen, etc. When we pair these up with
their respective singular forms, we get the following overall counts:
Pl. tokens
Sg. tokens
Overall, there are far more regular than irregular forms, but this is, of course, to be
expected by the very definition of “irregular”. It is, however, also consistent with Input
Optimization. The complexity of a system can be enhanced by limiting the number of
forms that exhibit marked properties. It can also be enhanced by limiting the distribution
of forms that do have those properties.
Is there a difference in the likelihood of a plural form given its regularity? If irregular forms are more complex, then we would expect their use to be statistically underrepresented because of Input Optimization. To test this, we calculate the ratio of singular
to plural tokens for each noun pair. This ratio allows us to examine the relative distribution of singular and plural forms, abstracting away from the overall frequency of any
specific lexical item. The difference is plotted in Figure 1.
Strikingly, the difference goes in the wrong difference here: irregular plurals are more
frequent relative to their singular forms than regular -s plurals with respect to their
singular forms. This difference is significant: t(90.318) = 3.151, p = 0.002.
0 10
singular/plural ratio
8 Morphological complexity and Input Optimization
Figure 1: Singular-to-plural ratios for regular and irregular plurals in English
We conclude that the distribution of irregular plurals is ambiguous. In terms of relative
frequency of singulars and plurals, the distribution goes in the wrong direction. In terms
of overall distribution, however, it goes in the right direction. There are 2891 instances of
irregular plurals in Brown and 46083 instances of regular plurals. If we were to assume
that both types were equally likely, the difference is certainly significant: X 2 (4344, N =
48974) = 425346.797, p < .001.
7 Zero marking
0 10
singular/plural ratio
Let’s now turn to zero marking. The claim is that zero marking is more complex and
therefore the prediction is that zero marking should be under-represented.
We examine this with respect to plurals in English in the Brown corpus. Zero-marked
plurals in English includes examples like: deer, aircraft, buffalo, etc. The difference in
ratios between regular plurals in -s and zero plurals is shown in Figure 2.
Figure 2: Singular-to-plural ratios for regular and zero plurals in English
Zero-marked plurals are far more frequent—relatively speaking—than regular plurals. Unfortunately, the variance is quite high—there is a lot of variation within each
Michael Hammond
category—and though the mean difference is large, it is not significant: t(19.002) =
−1.416, p = 0.173. As with the irregular plurals, however, the absolute difference is
significant. There are 184 instances of zero plurals in Brown and 46083 instances of regular plurals. If we were to assume that both types were equally likely, the difference is
certainly significant: X 2 (4285, N = 46267) = 311130.115, p < .001. Again then, though
the relative count is not significant, the absolute count goes in the right direction.
8 Summary
Our goal here has been to test the dimensions of morphological complexity proposed
in Anderson (2015) with the theory of Input Optimization. As reviewed above, Input
Optimization maintains that grammatical complexity, as assessed through constraint
violation, is minimized at the input level of the grammar. Specifically, we should see
under-representation of more marked morphological structures.
We picked out several dimensions of morphological complexity to examine, some of
which have already been treated with respect to Input Optimization. The following list
is repeated from Section 3 and annotated to reflect our results.
Words should have fewer morphemes. This is borne out by the distribution of
marked plurals and marked singulars in Welsh.
Haplology should be avoided. (This has already been established by Hammond 2016.)
More marked morphological operations (per the hierarchy above) should be
avoided. This was tested with respect to English plurals and is borne out in an
absolute comparison, but not in a relative comparison.
Morphophonology should be avoided. (This has already been established by
Hammond 2013.)
Ablaut, umlaut, truncation, etc. should be avoided. (This is essentially the
same as #3 above.)
Zero-marking should be avoided. This was tested with respect to English plurals and is borne out in an absolute comparison, but not in a relative comparison.
First, all else being equal, we expect forms with more morphemes to be dispreferred
to forms with fewer morphemes. We saw that this was borne out in a comparison of
singular–plural pairs in Welsh where in some cases the singular has an extra morpheme
and in others the plural has an extra morpheme.
Second, we predict that morphological haplology should be under-represented. This
was established in previous work with respect to the English genitive plural and adjectives in -ly.
Third, more marked morphological operations should be under-represented with respect to less marked morphological constructions. We saw an overall effect here with
English irregular noun plurals. We also saw that the relative distribution of plurals with
respect to singulars went in the opposite direction.
8 Morphological complexity and Input Optimization
Fourth, we predict that morphophonology should be avoided. This was established in
previous work with respect to morphophonology associated with English past -ed and
plural -s.
The fifth point is the same as the third.
Finally, zero-marking should be under-represented. We saw an overall effect here with
English zero-marked noun plurals. We also saw that the relative distribution of plurals
with respect to singulars went in the opposite direction.
We conclude that the parameters of complexity developed in Anderson (2015) tested
here and in previous work are correct.
We have seen that there is some divergence in the absolute and relative representation
of plural marking, but we leave investigation of that for future work.
Thanks to Nick Kloehn and Diane Ohala for useful discussion. All errors are my own.
Anderson, Stephen R. 1982. Where’s morphology? Linguistic Inquiry 13(4). 571–612.
Anderson, Stephen R. 1992. A-morphous morphology. Cambridge: Cambridge University
Anderson, Stephen R. 1999. A formalist’s reading of some functionalist work in syntax.
In Michael Darnell, Edith A. Moravcsik, Frederick J. Newmeyer, Michael Noonan &
Kathleen Wheatley (eds.), Functionalism and formalism in linguistics, vol. 1, 111–135.
Amsterdam: John Benjamins.
Anderson, Stephen R. 2015. Dimensions of morphological complexity. In Matthew Baerman, Dunstan Brown & Greville G. Corbett (eds.), Understanding and measuring morphological complexity, 11–26. Oxford: Oxford University Press.
Aronoff, Mark & Zheng Xu. 2010. A Realization Optimality-Theoretic approach to affix
order. Morphology 20. 381–411.
Ellis, N. C., C. O’Dochartaigh, W. Hicks, M. Morgan & N. Laporte. 2001. Cronfa Electroneg o Gymraeg (CEG): A 1 million word lexical database and frequency count for
Hammond, Michael. 2000. There is no lexicon! Coyote Papers 10. ROA #43 (1995), 55–77.
Hammond, Michael. 2013. Input optimization in English. Journal of the Phonetic Society
of Japan 17. 26–38.
Hammond, Michael. 2014. Phonological complexity and input optimization. Phonological
Studies 17. 85–94.
Hammond, Michael. 2016. Input optimization: Phonology and morphology. Phonology
33. 459–491.
Hayes, Bruce. 1984. The phonology of rhythm in English. Linguistic Inquiry 15. 33–74.
Michael Hammond
Kurisu, Kazutaka. 2001. The phonology of morpheme realization. University of California,
Santa Cruz PhD thesis.
Kučera, Henry & W. Nelson Francis. 1967. Computational analysis of present-day American English. Providence: Brown University Press.
Liberman, Mark & Alan Prince. 1977. On stress and linguistic rhythm. Linguistic Inquiry
8(2). 249–336.
Prince, Alan & Paul Smolensky. 1993. Optimality Theory. University of Massachusetts
and University of Colorado.
Russell, Kevin. 1993. A constraint-based approach to phonology. University of Southern
California PhD thesis.
Russell, Kevin. 1995. Morphemes and candidates in Optimality Theory. ROA #44.
Xu, Zheng & Mark Aronoff. 2011. A Realization Optimality Theory approach to blocking
and extended morphological exponence. Journal of Linguistics 47(3). 673–707.
Chapter 9
Multiple exponence in the Lusoga verb
Larry M. Hyman
University of California, Berkeley
Sharon Inkelas
University of California, Berkeley
with Fred Jenga
University of Texas, Austin
In this paper we address an unusual pattern of multiple exponence in Lusoga, a Bantu language spoken in Uganda, which bears on the questions of whether affix order is reducible
to syntactic structure, whether derivation is always ordered before inflection, and what motivates multiple exponence in the first place. In Lusoga, both derivational and inflectional
categories may be multiply exponed. The trigger of multiple exponence is the reciprocal
suffix, which optionally triggers the doubling both of preceding derivational suffixes and
of following inflectional suffixes. In these cases, each of the doubled affixes appear both
before (closer to the root) and after the reciprocal. We attribute this pattern to restructuring,
arguing that the inherited Bantu stem consisting of a root + suffixes has been reanalyzed as
a compound-like structure with two internal constituents, the second headed by the reciprocal morpheme, each potentially undergoing parallel derivation and inflection.
1 Introduction
Among the most important contributions of Steve Anderson’s realizational approach
to morphology have been his early insistence that morphology is not reducible to syntax, his argument that formal theoretical models of morphology need to take different
approaches to derivation and inflection (“split morphology”), his development of morphological rule ordering as the mechanism of ordering affixes, and his postulation that
redundant (inflectional) morphological exponence is actively avoided by grammars. According to Anderson (1992), derivational morphology takes place in the lexicon, while
Larry M. Hyman & Sharon Inkelas with Fred Jenga. 2017. Multiple exponence
in the Lusoga verb stem. In Claire Bowern, Laurence Horn & Raffaella Zanuttini
(eds.), On looking into words (and beyond), 171–189. Berlin: Language Science Press.
Larry M. Hyman & Sharon Inkelas with Fred Jenga
inflectional morphology takes place in the syntax. Inflectional morphology is realized by
the application of ordered rules which spell out features supplied by syntactic principles
such as agreement. The best evidence that the ordering of inflectional affixes cannot simply be read off of syntactic structure comes from morphotactics which have no analogue
or simple justification in syntax.
In this paper we address some rather unusual facts from Lusoga, a Bantu language
spoken in Uganda, which bear on the questions of whether affix order is reducible to
syntactic structure, and whether derivation is always ordered before inflection, particularly as concerns multiple exponence. In §2 we introduce the Bantu verb stem and briefly
summarize what has been said about the ordering of derivational suffixes within it. After
reviewing the findings that much of this ordering is strictly morphotactic, not following
from syntactic scope or semantic compositionality, in §3 we discuss multiple exponence
among the Lusoga derivational verb extensions. In §4 we then turn to the original contribution of Lusoga, which shows multiple exponence of inflectional agreement as well
as unexpected intermingling of inflectional and derivational affixation. We present our
analysis in §5 and conclude with a few final thoughts in §6.
2 The Bantu verb stem
Most overviews of the Bantu verb stem assume a structure with an obligatory verb root
followed by possible derivational suffixes (“extensions”), and ending with an inflectional
final vowel (FV) morpheme. As shown in (1), the verb stem may in turn be preceded by
a string of inflectional prefixes to form a word:
inflectional prefixes
While this structure has been reconstructed for Proto-Bantu (Meeussen 1967), there is
much variation on how the different derivational “verb extensions” are ordered. As
shown in Hyman (2003b), most Bantu languages show at least a tendency to favor the
“CARP” template in (2), for which we give reflexes in several Bantu languages:
The arguments for recognizing the CARP template include the following:
9 Multiple exponence in the Lusoga verb stem
(i) Certain pairs of co-occurring suffixes must appear in a fixed surface order. This
is true of the causative + applicative (CA), which can co-occur only in this order, independent of their relative scope. Compare the following two examples from Chichewa,
in which applicative -ir- introduces an instrument (Hyman & Mchombo 1992; Hyman
2003b). Scope (schematized on the right) varies across the two examples, but surface
order is the same:
applicativized causative:
lil-its-ir‘cause to cry with’
causativized applicative:
takas-its-ir- ‘cause to stir with’
[ [ cry ] -cause-with ]
[ [ stir-with ] -cause ]
(ii) Non-templatic orders which are driven by scope can occur with certain sets of suffixes, but are typically limited and show a “compositional asymmetry”: The a-templatic
order is restricted to the reading in which the surface order corresponds to relative scope,
while the templatic order can be interpreted with either possible scope relations (e.g. reciprocalized causative, causativized reciprocal). The two orders of causative and reciprocal (CR, RC) illustrate this property in (4), again from Chichewa:
(4) a.
templatic CR:
mang-its-an- ‘cause each other to tie’
‘cause to tie each other’
a-templatic RC:
mang-an-its- ‘cause to tie each other’
*‘cause each other to tie’
[ [ tie ] -cause-e.o. ]
[ [ tie-e.o. ] -cause ]
[ [ tie-e.o. ] -cause ]
As seen in (4a), the templatic CR order allows either scope interpretation, while the atemplatic RC order in (4b) can only be used to express a causativized reciprocal. The same
facts are observed in cases where the a-templatic order of applicative and reciprocal is
reinforced by an A-B-A “copied” sequence:
(5) a.
templatic AR:
‘tie (sth.) for each other’
‘tie each other for (s.o.)’
[ [ tie ] -for-e.o. ]
[ [ tie-e.o. ] -for ]
a-templatic RAR:
mang-an-ir-an- ‘tie each other for (s.o.)’
* ‘tie (sth.) for each other’
[ [ tie-e.o. ] -for ]
Again, as seen in (5a), the templatic AR order can have either scope (reciprocalized
applicative, applicativized reciprocal), while in (5b) the a-templatic (RA) + copy (R) sequence can only be compositional, hence an applicativized reciprocal. (We will see such
A-B-A sequences in Lusoga in §3.)
(iii) A third argument for CARP is that at least one language, Chimwiini, allows only
this order, whereas no Bantu language allows verb extensions to be freely ordered by
scope. Thus, Abasheikh (1978: 28) writes:
Larry M. Hyman & Sharon Inkelas with Fred Jenga
“In Chimwi:ni, unlike some other Bantu languages, the order of the extensions is
restricted. The following ordering of the extensions mentioned above is as follows:
- Verb Stem - Causative - Applied - Reciprocal - Passive. It is not possible to put
these extensions in any other order.”
Other than stative -ik-, which is more restricted in its co-occurrence with other suffixes,
the above summarizes the general picture for the productive extensions which are involved in valence. Even given the occasional variations, e.g. Kitharaka (Muriungi 2003),
which reverses the applicative and reciprocal, hence the order CRAP, the evidence points
unequivocally to the fact that extension order is determined primarily by template.
The importance of templaticity is also seen from the existence of one other valencerelated suffix, the short causative -i- (I) which typically occurs between the reciprocal and
passive, hence CARIP (see also Bastin 1986, Good 2005). Although both *-ɪc- (> -ɪs-, -is-)
and *-i- were present in Proto-Bantu, *-ɪc- occurred only in combination with *-i-, hence
*-ɪc-i- (cf. Bastin (1986). However, as summarized in (6), the current distribution of the
two extensions (as well as the productivity of -i-) varies considerably across different
Bantu languages (Hyman 2003b: 261):
-is-i- and -i: Kinande, Luganda, Lusoga
-is- only
: Chichewa, Shona, Zulu
-i- only (or almost only) : Nyamwezi, Nyakyusa
The fact that -is- is the linearly first extension and -i- a quite later extension in CARIP, for
reasons not motivated by scope, presents one more reason to accept a templatic, rather
than compositional approach to Bantu verb extensions. However, this conclusion is not
without interesting complications. As shown in such studies as Hyman (1994; 2003a)
and Downing (2005), -i- frequently produces frication of a preceding consonant (a.k.a.
Bantu spirantization) with potential multiple (cyclic) effects, as seen from the following
examples in which -i- co-occurs with the (non-fricativizing) applicative -il- suffix in (7)
from Cibemba:
‘be lost’
‘lose for/at’
‘make cry’
‘make cry for/at’
Morphology (I)
Morphology (A)
In both outputs, the applicative and short causative exhibit the expected surface AI
order. However, the frication of lub- ‘be lost’ and lil- ‘cry’ to luf- and lis- suggests that at
some level of representation, -i- is root adjacent. Hyman (1994) adopts the above cyclic
analysis in which morphology and phonology are interleaved (see e.g. Kiparsky 1982):
-i- combines with the root on the first morphological cycle, triggering a phonological
application of frication on the root. When the applicative is added on the next cycle of
morphology, it is “interfixed” between the root and the short causative, in conformity
9 Multiple exponence in the Lusoga verb stem
with the AI order required by the CARIP template. This example illustrates the surface
nature of the template.
Although it is not part of the CARIP template of valence-changing derivational suffixes, the “final vowel” (FV) inflectional ending position is also templatic in that it is
required in most Bantu languages. The set of suffixes that may appear in the FV position
includes past tense *-ɪ-, subjunctive *-ɛ, and (in most other contexts) default *-a. The -ɛ
portion of perfective *-il-ɛ, which we will encounter in §4, is also in this slot, even as
the -il- portion is sometimes considered to be part of the extension system. The customary reason for assuming bimorphemic *-il-ɛ is that the short causative (I) and passive
(P) occur between the two parts, hence *-il-i-ɛ and *-il-ʊ-ɛ (Bastin 1983). If we assumed
that *-il-ɛ was monomorphemic, we would have to assume some kind of exfixation or
metathesis of the causative and passive with the [il] portion of -ilɛ. There is a second
argument from Lusoga (and Luganda): Whenever causative -i- or passive -u- is present,
the FV of the perfective complex is -a (see (9) and note 5 below). We assume that -iloccurs in the template ordered before -I-P- with the function of perfectivizing the extended derivational base so it can accept -ɛ or -a (cf. §4).) With this established, we are
ready to go on to the issues that arise in Lusoga.
3 Lusoga verb extensions
As mentioned above, Lusoga is spoken in Uganda and is the Bantu language most closely
related to Luganda. The data cited in this study were contributed by Fr. Fred Jenga, a
native speaker from Wairaka (Jinja District).
3.1 Long and Short Causatives
Lusoga exhibits the CARIP template discussed above, where C refers to the long causative -is- and I refers to the short causative extension -i-. In fact, Lusoga uses both -is-i- and
-i- productively and often interchangeably, to express both causation and instrumentals:
-lim-is-i-, -lim-i- ‘cause to cultivate, cultivate with (sth.)’. As indicated, -is- cannot occur
without -i-, while the reverse is possible. The two causative morphs are quite consistent
in their CARIP templatic ordering with respect to the applicative, namely, -is-il-i- (CAI),
-il-i- (AI), which are realized as -is-iz- and -iz- by the following processes:
‘make cultivate for/at’
lim-is-iz-i-a lim-iz-i-a
lim-is-iz-y-a lim-iz-y-a
Larry M. Hyman & Sharon Inkelas with Fred Jenga
3.2 Reciprocal + Short Causative
Challenges to the CARIP template arise with the reciprocal suffix, which in Lusoga has
the long reflex -agan- of Proto-Bantu *-an-.1 In the next few subsections we will consider
how the reciprocal combines with its fellow extensions in the CARIP template, including
both ordering flexibility as well as affix doubling.
We begin with the short causative, -i-. When used alone, without the long causative,
we observe flexible ordering possibilities, well beyond what would be expected from the
CARIP template. In these and subsequent examples, a left bracket indicates the boundary
between inflectional prefixes and the beginning of the verb stem:
‘they make each other sew’
a. bà-[tùùng-ágán-y-á /tùùng-agan-i-a/
b. bà-[tùùnz-ágán-á
c. bà-[tùùnz-ágán-y-á /tùùng-i-agan-i-a/
In none of (9a-c) does the short causative -i- surface as a vowel. Nonetheless, its presence
is clearly felt. In (9a) it glides, preceding a following vowel; in (9b) and (9c) it spirantizes
the final /g/ of /-tùung-/ ‘sew’ to [z] by a general process in the language, and is otherwise
deleted before the following vowel (of the reciprocal). The reciprocal suffix -agan- does
not trigger compensatory lengthening when vowels glide or delete before it, as also seen
in the examples with root-final vowels immediately followed by -agan-, below:
‘they shave each other’
‘they fear each other’
Note that (9c) appears to exhibit two instances of the short causative: root spirantization
indicates a following short causative, and the glide following the reciprocal also indicates
a following short causative. These two surface reflexes of the short causative could result
from input suffix doubling, something that is attested elsewhere in Lusoga, as shown in
the UR given for (9c). Alternatively, the double reflex of the short causative could be the
result of a-templatic IR order, in which the single short causative spirantizes the root
and then the reciprocal is interfixed inside of it, an analysis Hyman has supported for
Chibemba (7). On this account, short causative doubling (IRI) is illusory. We leave open
for now whether the IRI ordering is required; what is clear is that both RI and IR are
3.3 Reciprocal + Long Causative
We turn next to the long causative -is-, which, as we have seen, must co-occur with the
short causative -i-. The most common realization when reciprocal and long causative
While it is marginally possible for the reciprocal and passive to co-occur in some Bantu languages, typically
with an impersonal subject, e.g. Ndebele kw-a-sik-w-an-a ~ kw-a-sik-an-w-a ‘there was stabbing [stabbed]
of each other’ (Sibanda 2004: 66), we have thus far not been able to get the two to co-occur in Lusoga and
will therefore ignore the passive extension in what follows.
9 Multiple exponence in the Lusoga verb stem
are both present is for -agan- to appear between -is- and -i-, as in (11a), exhibiting the
CRI order expected given the CARIP template. However, two other surface realizations
are also possible:2
(11) ‘they make each other sew’
a. bà-[tùùng-ís-ágán-y-á /tùùng-is-agan-i-a/
b. bà-[tùùng-ís-ágán-á
c. bà-[tùùng-ágán-ís-á
In (11b), -agan- follows -is-i- (CIR). In (11c) -agan- precedes -is-i- (RCI). This variation reveals the same freedom with respect to the ordering of the long causative and reciprocal
extensions that we observed in §⁇ with respect to the ordering of the short causative
and reciprocal extensions.
Note that for phonological reasons, it is impossible to distinguish between the inputs
-is- and -is-i- before -agan-. The reason is that, sandwiched between long causative -isand following vowel-initial -agan-, short causative -i- would glide to -y- and then get
absorbed into the preceding [s], without leaving a trace. As was seen in (10), compensatory lengthening is not expected before -agan-. However, it can be detected between
-i- and a FV when an enclitic such as locative class 17 =kò ‘on it, a little’ is added:
‘they make each other sew a little’
a. bà-[tùùng-ís-ágán-y-áá =kò /tùùng-is-i-agan-i-a =kò/
b. bà-[tùùng-ís-ágan-á =kò
/tùùng-is-i-agan-a =kò/
c. bà-[tùùng-ágán-ís-áá =kò
/tùùng-agan-is-i-a =kò/
CIRI + encl
CIR + encl
RCI + encl
In (12a), the final length on -aa can be directly attributed to the gliding of the preceding
-i-, since there is a surface [y], as can be the final length in (12c), where the glide has
been absorbed into the preceding [s]. Although (12b) does not show a surface reflex of
the internal -i-, we continue to assume that -is- must be accompanied by -i-, as also
reconstructed for Proto-Bantu (Bastin 1986).
While there are three possible realizations when reciprocal -agan- combines with the
long and short causative suffixes, the preferred surface orders are IRI in (9c), and CRI,
in (11a). RI and CRI are of course predicted straightforwardly from CARIP, while the
IR of IRI is not. Both early placement of C (-is-) in the CARIP template and the early
realization of the first -i- of the hypothesized a-templatic IRI ordering discussed in this
section are consistent with a generalization that Hyman (2003b: 272) has characterized
as “causativize first!”: Both -is- and -i- are spelled out early, but later affixation may
result in two surface reflexes of -i-, either because of interfixation of subsequently added
extension suffixes or because of outright morphological -i- doubling of the kind seen in
the Chichewa RAR case illustrated in (5b).
Since Lusoga has a /L/ vs. Ø tone system (Hyman 2016), only L(ow) vowels are marked with a grave accent
in underlying forms. Vowels without an accent receive their surface tones by specific rules. H(igh) tone is
marked with an acute in output forms.
Larry M. Hyman & Sharon Inkelas with Fred Jenga
3.4 Reciprocal + Applicative
The CARIP template is complicated further by the behavior of the applicative, represented by “A” in CARIP. In all three of the following examples, the transitive verb kùb‘beat’ is both reciprocalized ‘beat each other’ and applicativized. Applicative -ir- licenses
a locative argument, expressed by the enclitic =wà ‘where’. Here again we observe alternative affix orders:
‘where do they beat each other?’
a. bà-[kùb-ír-ágán-á =wà
b. bà-[kùb-ágán-ír-á =wà
c. bà-[kùb-ír-ágán-ír-á =wà ARA
(13a) represents the expected AR order of CARIP, while the RA order of (13b) represents
an order which is closer to the compositional interpretation of the resulting verb. In
(13c) -ir-agan-ir- has both the AR and RA orders. The variation between AR, RA and
ARA orders represents a competition between the demand of the CARIP template for
one order and the requirement for affixes to appear in a surface order that reflects their
relative scope. The AR order (13a) is templatic; the RA order in (13b) is a scope-based or
compositional override. As suggested by Hyman (2003b), ABA affix doubling can thus be
interpreted as a means of satisfying both template and compositionality considerations;
if the template wants AR and scope wants RA, then ARA, in some manner, satisfies both.3
An illustrative pair of examples is presented in (14), based on the transitive verb bal‘count’, which is reciprocalized and applicativized. In this instance, applicative -ir- licenses a benefactive object:
‘they count them [inanimate class 8] for each
‘they [animate] count each other for us’ ~
‘they count us for each other’
By varying the animacy of the object pronouns in (14), it is possible to bias the scope
interpretation of reciprocal and applicative in opposite directions. In (14a) the object prefix -bi- ‘them’ (class 8) represents an inanimate object such as èbitabo ‘books’ or èbikopò
‘cups’, hence animate ‘each other’ (referring back to bà- ‘they’) claims the benefactive
rôle over inanimate -bì- ‘them’. In this sentence the AR order -ir-agan- satisfies both the
CARIP template and scope: [[count them] for each other]. In (14b), animate first person
object -tù- ‘us’ preferentially claims the benefactive role over third person -agan-, again
The questions in (13) unambiguously ask where the action took place and could therefore be answered “in
Jinga” or “in the house”. The absence of the applicative in the corresponding question bà-[kùb-agan-a =wà
‘where do they beat each other?’ more narrowly asks what spot or area of the body was hit. An appropriate
answer would therefore be “on the head”. Finally, the double reflex of applicative -ir- of ARA -ir-agan-irin (13c) is reminiscent of the double reflex of RAR -an-ir-an- in Chichewa in (5c): the sequence -ir-aganis licensed by CARIP, while -agan-ir- represents the scope override. Concerning ABA suffix ordering, one
might note that Lusoga (13c) violates Hyman’s 2003 generalization, observable in Chichewa (5c), that AB
always reflects the scope, while BA is templatic.
9 Multiple exponence in the Lusoga verb stem
referring back to bà- ‘they’. The -ir-agan- order in this sentence is also templatic, but
this time need not reflect scope: Although the preferred interpretation is [[count each
other] for us], the other scope ([[count us] for each other]) is also possible, though pragmatically less likely. It is thus not surprising that the two alternatives are also possible
in (15) with the same meaning:
‘they [animate] count each other for us’
a. bà-tù-[bál-ágán-ír-a
b. bà-tù-[bál-ír-ágán-ír-á ARA
Parallel to (12b,c), (15a) is a scope override, while -ir-agan-ir- satisfies both CARIP and
scope in (15b). What is surprising is that the same possibilities are at least marginally
acceptable in (16), both sentences having the same meaning.
(16) ‘they count them [inanimate cl. 8] for each other’
a. bà-bì-[bál-ágán-ír-á
b. ⁇bà-bì-[bál-ír-ágán-ír-á ARA
As in (15a,b), the RA sequence occurs perfectly well in (16a), while the doubled RAR
sequence in (16b) was judged as sounding “Lugandish,” perhaps OK to use, but seems
a little funny, “like a foreigner learning Lusoga.” While we have an explanation for the
variation in (13b,c) and (15a,b), neither CARIP nor scope predicts that (16a,b) should be
possible. We thus arrive at a major divergence from the template + scope approach that
accounts for the variations considered above in Lusoga, as well as Chichewa, Chibemba,
and other Bantu languages. We now address why this may be so in the next section.
4 Inflectional FV suffixes in Lusoga
In §3 we were largely able to account for surface variations in verb extension order in
Lusoga by appealing to a tradeoff between the CARIP template and scope considerations:
While the templatic CARIP is always available and represents the default order of affixes,
conflicting orders may be licensed by scope, and template-scope interactions can even
result in ABA sequences. The one major exception concerns cases of atemplatic (A)RA
-(ir)-agan-ir-, in which a-templatic RA -agan-ir- cannot be said to be a compositional
override. In this section we show that this unexpected ordering likely owes its existence
to an optional restructuring of reciprocal -agan-.
To illuminate this hypothesis, we now turn to the interaction of reciprocal -agan- with
the set of complementary inflectional “final vowel” (FV) suffixes. Every verb must end in
one of these. While most verbs end in the default FV -a, specific TAM categories require
one of two other finals, the FV -e or the FV complex -ir-e, which have the following
Larry M. Hyman & Sharon Inkelas with Fred Jenga
: hortative/subjunctive, affirmative imperative singular
with an object prefix, affirmative imperative plural,
negative near future (F1)
-ir-e : perfect/today past (P1), yesterday past (P2)
-a : elsewhere
As summarized in (17a) and exemplified in (18), what unifies the uses of -e is its use in a
subset of irrealis constructions:
‘count them!’
‘count (pl.)!’
‘let’s count!’
‘they will not count’
(singular imperative with an object
prefix; cf. bàl-à ‘count!’)
(plural imperative)
(negative near future F1)
As per the general Bantu stem structure in (1), the FV follows the verb extensions, e.g.
applicative -ir- in (19).
(19) a.
‘count them for us!’
‘count (pl.) for us!’
‘let’s count for them!’
‘they will not count for us’
However, two options are attested when the extension is -agan-:
(20) a.
‘count each other!’
‘let’s count each other!’
‘they will not count each other’
‘count (pl.) each other!’
‘let’s count each other!’
‘they will not count each other’
The expected forms are in (20a), where reciprocal -agan- is followed by FV -e. Surprisingly, the alternatives in (20b) show the FV -e occurring both before and after the reciprocal. In these forms we have segmented off the first FV as -e-, which means that the
reciprocal allomorph is -gan- in this context. The alternative would be to recognize a reciprocal allomorph -egan- which is used whenever there is an upcoming FV -e.4 We will
see in the discussion of perfective -ir-e below that the first -e- is correctly interpreted as
a copy agreeing with the final -e.
The same variation obtains when the applicative suffix is present:
It is important to note that -e-gan- cannot be used if the FV is -a: ò-kú-[bál-ágán-á ‘to count each other’,
bà-[bàl-ágán-á ‘they count each other’ vs. *ò-kú-[bál-é-gán-á, *bà-[bàl-é-gán-á.
9 Multiple exponence in the Lusoga verb stem
‘count (pl.) them for each other!’
‘let’s count them for each other!’
‘they will not count them for each other’
‘count (pl.) them for each other!’
‘let’s count them for each other!’
‘they will not count them for each other’
In (21), the applicative -ir- precedes the reciprocal, showing the AR order predicted by
the CARIP template, but the presence of the FV between the two in the forms in (21b) is
highly unusual from a Bantu point of view.
Exactly the same phenomenon of FV doubling occurs with the perfective -ir-e FV
complex. As in Luganda, Lusoga -ir-e has several allomorphs. These are presented in
(22) in the form they take prior to the application of phonological rules:5
-ir-e : after a CV- verb root
-i- … -e : when fused (“imbricated”) into a longer verb base
: after a labial consonant and /n/
: after a fricated consonant [s] or [z], where -i- → y → Ø
The above four allomorphs are illustrated in the perfect/today past (P1) tense below:
(23) a.
‘we feared’
‘we ran into (s.o./sth.)’
‘we sent’
‘we counted’
In (23a), the /-ir-e/ allomorph is realized after the CV verb /-tì-/ ‘fear’. In (23b), longer
verb bases that end in a coronal consonant undergo imbrication whereby -i- metathesizes
with the consonant. We will see in further examples that the reciprocal -agan- extension
also undergoes imbrication to become -again-e. In (23c), the /-i-/ of /-i-e/ glides to [y].6 In
(23d), -i- fricates the preceding /l/ to [z], yielding the same derivation as in (⁇): /-bal-i-e/
→ baz-i-e → baz-y-e → baz-e, the [y] being absorbed into the preceding fricative.
We will now illustrate each of the above allomorphs of -ir-e in (23) as they are realized
with the reciprocal extension. We start with the reciprocalized version of (23b), which
exhibits the imbricating -i-e perfective FV allomorph. The historically conservative variant, in which the root is followed directly by the reciprocal suffix and then the -i-e FV, is
As was discussed at the end of §2 with respect to Proto-Bantu, we represent -ir-e as bimorphemic. In (22)
we omit the passive and causative forms that occur with final -a, thereby providing even more allomorphs,
e.g. the perfective of the lexicalized passive verb /-lùm-u-/ ‘be in pain’ is tù-[lùm-íír-w-à ‘s/he was in pain’,
while the perfect of the lexicalized causative verb /-tèm-i-/ ‘blink’ is tù-[tèm-ííz-à ‘we blinked’, where r → z
is triggered by the causative suffix /-i-/. Both occur with a long -iir- morph followed by -a. As seen in these
examples, the fact that -a is used with passive -u- and causative -i- provides additional evidence that -iris a separate morpheme from -e or -a.
6 The following -e actually lengthens, but then is shortened by a rule of final vowel shortening (FVS), which
converts à-lím-y-èè to à-lím-y-è. Thus compare the long vowel in à-[lím-y-↓ éé =kò which is realized when
an enclitic follows. (↓ indicates a downstepped high tone).
Larry M. Hyman & Sharon Inkelas with Fred Jenga
shown in (24a). However, the preferred alternative is (24b), in which the perfective -i-e
appears, imbricated, both immediately following the root and immediately following
the reciprocal. URs showing both a single and a doubled FV complex are provided for
each form:
(24) ‘we ran into each other’
a. /tù-[tomer-agan-i-e/
b. /tù-[tomer-i-e-agan-i-e/
A parallel situation obtains in (25), which corresponds to (23c):
(25) ‘we sent each other’
a. /tù-[tùm-agan-i-e/
b. /tù-[tùm-i-e-agan-i-e/
Example (26), based on (23d), shows similar facts, the main difference being the frication
triggered by causative -i- on the verb root -bal- ‘count’:
‘we counted each other’
a. /tù-[bal-agan-i-e/
b. /tù-[bal-i-e-gan-i-e/
Finally, in (27), we see a reciprocalized version of the root in (23a), which, on its own,
would take the -ir-e FV allomorph. The historical variant is shown in (27a), but the
preferred variant, with doubled FV, is given in (27b):
(27) ‘we feared each other’
a. /tù-[tì-agan-i-e/
b. /tù-[tì-ir-e-gan-i-e/
As before there are two instances of the perfective in (27b), vs. one in (27a). In this case
of doubling, however, the allomorphy of the perfective is different in the two copies. The
first copy of the FV follows a CV root and assumes the expected -ir-e form; the second
copy, following the longer -agan-, assumes the imbricating -i-e form. The fact that the
allomorphs are different suggests that the two copies are generated independently.
In sum, both the irrealis -e FV and the perfective FV allomorphs can appear once in a
reciprocalized verb, or twice, with the double spell-out being clearly preferred. We now
turn to an analysis of these facts in §5.
5 Towards an analysis
From the perspective of familiar cross-linguistic principles of affix ordering (derivation
closer to the root than inflection; prohibition on multiple exponence), Lusoga presents
two interesting puzzles: (i) derivational and inflectional suffixes both double; (ii) when
inflectional suffixes double, they do so on either side of derivation, violating the “split
9 Multiple exponence in the Lusoga verb stem
morphology” hypothesis. Thus, in a form like tù-[bàl-é-gàn-é ‘let’s count each other’
from (20b), the irrealis FV -e occurs both before and after the derivational reciprocal
suffix -gan-. While doubling of derivational suffixes has been previously discussed in the
Bantu literature (Hyman 2003b), the doubling of inflection has not. This is the final focus
of this study. Given that the doubling occurs in verbs containing the reciprocal suffix
-agan-, the question we face is what it is about this suffix that triggers the phenomenon.
Why is it only the reciprocal that does this?
Our hypothesis is that the phonological form of the reciprocal has led to a reanalysis
of the internal morphological structure of the reciprocalized Lusoga verb stem. The reciprocal suffix -agan- is the only Lusoga derivational suffix which is both disyllabic and
a-initial. Taken together, these phonological facts are consistent with a reanalysis of the
verb stem in which the reciprocal suffix is bimorphemic, -a-gan. Because of its phonological identity, the -a- portion became identified with the default FV -a. At the same time
this permitted the reanalyzed reciprocal suffix, -gan-, to conform to the default -CVCverb root structure.
As a result of this reanalysis, the verb structure in (28a) became reinterpreted as in
(28b), where we use # to indicate the internal stem boundary:
Expected (inherited)
Unexpected (innovated)
From this step, the following analogical reanalyses follow straightforwardly, with allomorph variation in (29b) conditioned by the phonological size and shape of the root:
(29) a.
Expected (inherited)
Unexpected (innovated)
In (29a) inflectional -e and -ir-e are suffixed after derivational -agan-. (We show the
perfective as -ir-e in the above, although its exact allomorph will vary, as pointed out
in (22).) In (29b) we see the reanalysis brought on by analogy. As a result, from the
simple right-branching suffixing construction in (29a), reciprocal verb stems became
reanalyzed, optionally, as compounding, with two roots: the verb root, and -gan-. Both
are inflectable (29b), though it is possible also to inflect only the verb stem as a whole
As indicated, the compounding account allows us to account for the apparent affixation of the inflectional suffixes -e and -ir-e inside of a derivational suffix, the restructured
reciprocal -gan-. These suffixes also potentially precede the short causative -i-. The inflection of stems containing both -(a)gan- and the short causative is seen in the following
six alternants, based on the causative verb -lùm-i- ‘injure’, where -i- glides to [y] before
the following vowel:7
Although the verb root -lùm- means ‘bite’, the semantics of the lexicalized causative verb -lùm-ì- ‘injure,
cause pain’ is most clearly seen in the corresponding lexicalized passive verb -lùm-ù- ‘to ache, be in pain’.
Larry M. Hyman & Sharon Inkelas with Fred Jenga
(30) ‘let’s injure each other’
a. tù-[lùm-y-ágàn-é
b. tù-[lùm-y-é-gàn-é
The options in (30a) all follow the expected parsing, with -agan- treated as a derivational
suffix. Those in (30b) represent the claimed restructuring in which the FV -e occurs both
before and after reciprocal -gan-. In each set, causative -i- appears immediately after the
root in the first example, after the reciprocal in the second, and both before and after
in the third. In the last two examples of (30b), the first (inflectional) -e occurs not only
before -gan-, but also before the (derivational) causative -i- suffix. Parallel cases could be
illustrated in which -i- combines with the various perfective allomorphs. Our analysis,
which assumes a double or compound stem structure, each of which is independently
inflected, thus nicely accounts for the above (and other) cases where the inflectional
FV linearly precedes (restructured) reciprocal -gan- and potentially other derivational
Root-(exti )-(FVj )
gan-exti -FVj
Before moving on to our conclusion, we briefly cite phonological evidence for our
analysis from closely related Lulamogi, which also optionally realizes the inflectional
FV both before and after reciprocal -gan- (Hyman In press). In this language, there are
two facts concerning vowel length and (pre-)penultimate position that are relevant to
the analysis of the reciprocal. First, a word-initial V- prefix lengthens if it is followed by
a monosyllabic stem (i.e. if it is in penultimate position). This is seen in (32a):
(32) a.
‘s/he fears’
‘they fear’
‘s/he laughs’
As seen in (32b), if the word-initial prefix has the shape CV-, its vowel doesn’t lengthen,
while in (32c) /a-/ fails to lengthen because it is in pre-penultimate position. The second
length-related phenomenon is exemplified in (33):
9 Multiple exponence in the Lusoga verb stem
‘we will fear’
‘we will laugh’
In (33a), the prefix sequence /tu-á-/ (1pl-fut) undergoes gliding + compensatory lengthening to be realized [tw-áá-] in penultimate position. In (33b), on the other hand, the
same gliding process applies, but the result is short [tw-á-], since prefixal vowel sequences are realized short in pre-penultimate position.
A systematic exception to both penultimate prefixal V-lengthening and pre-penultimate prefixal V+V shortening occurs when reciprocal -agan- is suffixed to a monosyllabic
verb root:
‘s/he often fears’
‘we will fear each other’
In (34a), where -agan- is used as a frequentative suffix, the initial subject prefix à- lengthens even though it is in pre-penultimate position. In (34b), the [tw-áá-] sequence remains
long even though it too is in pre-penultimate position. Note also that the first vowel of
the -ty-àgàn- sequence is short, i.e. compensatory lengthening appears not to apply. All
of these observations can be accounted for if we assume the same analysis as in Lusoga:
‘s/he often fears’
‘we will fear each other’
In (35) the # symbol again represents the boundary between the two stems. The result in
(35a) is that the initial /a-/ is now in penultimate position in the first stem and is thus free
to lengthen. In (35b) the /tu-á-/ is now also in penultimate position, and so [tw-áá-] fails
to shorten. Taken alone, either our Lusoga analysis or this Lulamogi analysis of Hyman
(In press) might seem overly speculative—and especially surprising from a traditional
Bantu perspective. However, taken together, the two sets of facts support each other. In
fact, Lulamogi is the only other Bantu language we are aware of that allows the option
of spelling out the FV both before and after the reciprocal extension. Thus compare the
following with Lusoga (20a,b):
‘let’s count each other’
a. tú-[bàl-àgàn-é
b. tú-[bàl-è-gàn-é
As we stated earlier, we think this reconceptualization is due to the fact that -agan- is the
only highly productive suffix that could be re-interpreted in the way we have suggested.
It is significant that the historical Bantu reciprocal suffix *-an- often joins with other
suffixes to make a -VCVC- conglomerate (cf. Bostoen & Nzang-Bie 2010: 1289–91 for
further discussion). In Lusoga, Lulamogi, Luganda, and many other Bantu languages,
*-an- has joined with an archaic *-ang- or *-ag- extension which likely had an original
Larry M. Hyman & Sharon Inkelas with Fred Jenga
pluractional interpretation.8 As we have suggested, the shape and “weightiness” of the
resulting -agan- has led to multiple exponence and inflectional “entrapment” within the
derivational morphology of the verb stem in Lusoga (and Lulamogi). We consider further
implications in the next section.
6 Conclusion
In the preceding sections we have documented multiple exponence of derivational suffixes (§3) and inflectional suffixes (§4) in Lusoga, and have proposed a restructuring
analysis of *-agan- > -a-gan- in §5 to account for the multiple copies of the inflectional
FV in -e-gan- sequences. Harris & Faarlund (2006) discuss instances in which grammaticalization of an outer affix “traps” an inner one, with the result that the two affixes occur
in an unexpected order. Loss of the trapped affix is an attested diachronic repair for this
“entrapment” situation; doubling (by addition of an outer inflectional affix) is another.
Lusoga, however, appears to illustrate reanalysis of a different kind, in which an existing affix is reanalyzed as a root, and doubling represents agreement in a compoundinglike structure of the sort proposed by Inkelas & Zoll (2005) for reduplication, in which
doubled morphemes can also show divergent allomorphy of the kind displayed by the
perfective complex in Lusoga. If correct, the Lusoga facts are important both from a
synchronic and diachronic point of view. An historical change of *affix > root would
contradict the more broadly attested grammaticalization pattern *root > affix (but see
Norde 2009). Synchronically, multiple exponence of the inflectional ending is quite different from the doubling of derivational suffixes. While the latter has been interpreted as
the resolution of a template-scope mismatch, perhaps spelled out cyclically, this cannot
work for inflectional doubling. In the examples in (30) above, it was seen that the derivational causative -i- can appear once or twice: It can appear either before the reciprocal
(-i-agan-), after it (-agan-i-), or both before or after (-i-agan-i-). However, we have thus
only shown two possibilities concerning inflectional FVs such as subjunctive -e. In (20),
repeated as (37a,b), we saw that -e can appear either after -agan- or both before and after
‘count each other!’
‘let’s count each other!’
‘they will not count each other’
‘count (pl.) each other!’
‘let’s count each other!’
‘they will not count each other’
‘count each other!’
‘let’s count each other!’
‘they will not count each other’
While the most general realization of the reciprocal is -agan- in Luganda, the form is regularly -aŋŋanafter CV verb roots, e.g. mw-aŋŋan- ‘shave each other’. Since -aŋŋan- derives from *-angan- via Meinhof’s
Law (Katamba & Hyman 1991: 192–193), this provides evidence that the earlier bimorphemic form was
likely *-ang-an- in all three closely related languages.
9 Multiple exponence in the Lusoga verb stem
However, (37c) shows that it is not possible to express the inflection only on the first
stem. These facts motivate the compounding structure we have offered for the Lusoga
verb stem, and suggest that the second member, on which inflection is obligatory, is the
head, and agreement in derivational and inflectional properties is optionally enforced,
explaining the presence of duplicate morphology on the first constituent. The structures
in (37) are not amenable to a cyclic analysis proceeding bottom-up from the verb root.
In Lusoga, compounding, derivation and inflection are intermingled in typologically
unusual ways. The complexities of the system – and of multiple exponence in general
(Anderson 2015: 21) – give credence to views in which morphology is a component of
grammar with its own internal morphotactic organization; it does not mirror syntax
directly and thus cannot be reduced to syntactic principles. This is a result of which we
think Steve would approve.
Abasheikh, Mohammad. 1978. The grammar of Chimwi:ni causatives. University of Illinois
PhD thesis.
Anderson, Stephen R. 1992. A-morphous morphology. Cambridge: Cambridge University
Anderson, Stephen R. 2015. Dimensions of morphological complexity. In Matthew Baerman, Dunstan Brown & Greville G. Corbett (eds.), Understanding and measuring morphological complexity, 11–26. Oxford: Oxford University Press.
Bastin, Yvonne. 1983. La finale -IDE et l’imbrication en bantou. Annales, Série IN-8, Sciences Humaines 114. Tervuren: Musée Royal de l’Afrique Centrale.
Bastin, Yvonne. 1986. Les suffixes causatifs dans les langues bantoues. Africana Linguistica 10. Tervuren: Annales du Musée Royale de l’Afrique Centrale. Série IN-8, Sciences
Humaines 121, 55–145.
Bostoen, Koen & Yolande Nzang-Bie. 2010. On how “middle” plus “associative/reciprocal”
became “passive” in the Bantu A70 languages. Linguistics 48(6). 1255–1307.
Downing, Laura J. 2005. Jita causative doubling and paradigm uniformity. In Laura J.
Downing, T. Alan Hall & Renate Raffelsiefen (eds.), Paradigms in phonological theory,
122–144. Oxford: Oxford University Press.
Good, Jeff. 2005. Reconstructing morpheme order in Bantu: the case of causativization
and applicativization. Diachronica 22(1). 55–109.
Harris, Alice & Jan Terje Faarlund. 2006. Trapped morphology. Journal of Linguistics 42.
Hyman, Larry M. 1994. Cyclic phonology and morphology in Cibemba. In Jennifer Cole
& Charles Kisseberth (eds.), Perspectives in phonology, 81–112. Stanford: C.S.L.I.
Hyman, Larry M. 2003a. Sound change, misanalysis, and analogy in the Bantu causative.
Journal of African Languages and Linguistics 24(1). 24–55.
Hyman, Larry M. 2003b. Suffix ordering in Bantu: A morphocentric approach. In Geert
Booij & Jaap van Marle (eds.), Yearbook of morphology 2002, 245–281. Amsterdam:
Larry M. Hyman & Sharon Inkelas with Fred Jenga
Hyman, Larry M. 2016. The autosegmental approach to tone in Lusoga. Tech. rep. Ms. University of California, Berkeley.
Hyman, Larry M. In press. Prefixal vowel length in Lulamogi: a stratal account. Journal
of African Languages and Linguistics.
Hyman, Larry M. & Sam Mchombo. 1992. Morphotactic constraints in the Chichewa verb
stem. In Proceedings of the Berkeley Linguistic Society. Berkeley Linguistic Society.
Inkelas, Sharon & Cheryl Zoll. 2005. Reduplication: Doubling in morphology. Cambridge:
Cambridge University Press.
Katamba, Francis & Larry M. Hyman. 1991. Nasality and morpheme structure constraints
in Luganda. In Francis Katamba (ed.), Lacustrine Bantu phonology (Afrikanistiche Arbeitspapiere 25), 175–211. Cologne: Institut für Afrikanistik, Universität zu Köln.
Kiparsky, Paul. 1982. Lexical morphology and phonology. In Linguistic Society of Korea
(ed.), Linguistics in the morning calm: Selected papers from SICOL–1981, 3–91. Seoul:
Meeussen, Achille E. 1967. Bantu grammatical reconstructions. Africana Linguistica 3(1).
Muriungi, Peter Kinyua. 2003. Phrasal movement inside Bantu verbs. University of
Tromsø PhD thesis.
Norde, Muriel. 2009. Degrammaticalization. Oxford: Oxford University Press.
Sibanda, Galen. 2004. Verbal phonology and morphology of Ndebele. University of California, Berkeley PhD thesis.
Chapter 10
Romansh allomorphy (Again!)
Martin Maiden
Oxford University
This essay resumes a debate which has continued for some years between me and Stephen
Anderson regarding the correct analysis of a complex set of data from the verb morphology of the Romansh dialect of Savognin. Anderson believes that the data are an example
of “phonologically conditioned allomorphy”, whilst I maintain that they exemplify “morphomic”, or autonomously morphological, alternation patterns, whose only phonological
motivation lies in diacrhrony. I reply below to Anderson’s most recent analysis of the data,
by discussing reasons in support of my “morphomic” account. I conclude, however, by considering the possibility that our two accounts may be too exclusivist in their respective
“phonologizing” and “morphologizing” stances, and that they are not necessarily wholly
1 Introduction
Some readers might regard this essay as an example of chutzpah, or downright impertinence, but it is a sincere mark of respect for Steve Anderson that I feel able to disagree
with him even in a collection published in his honour. One would not do this with a
less intellectually generous scholar. What follows is, in fact, a further instalment in an
amicable difference of opinion I have had with him for some years (see Anderson 2008;
Anderson 2011b; Maiden 2011b; Anderson 2013) concerning the analysis of a set of data
from some Romansh dialects, and principally, the Surmiran variety spoken in Savognin.
Readers, and not least the honorand himself, may be feeling that there is little left to say.
That there is reason to continue the debate is suggested, for example, by Andrea Sims’
deft review of the issues (Sims 2015: 202–206), to which I return in my conclusion.
Anderson’s analysis displays not only his characteristically penetrating theoretical
rigour, but also a quite formidable grasp of the data. Reasons of space, and the fact
that the data are lucidly laid out by him in previous publications (e.g., Anderson 2008,
2011) permit me to do no more here than summarize them: Savognin has a recurrent pattern of vocalic alternations] in the verb root (or “stem”), such that one alternant occurs
in the root of the singular and third person forms of the present indicative imperative,
throughout the present subjunctive, and also in third conjugation infinitives, while another occurs in the root in the remainder of the paradigm. What is involved originates
Martin Maiden. 2017. Romansh allomorphy (Again!) In Claire Bowern, Laurence
Horn & Raffaella Zanuttini (eds.), On looking into words (and beyond), 189–211. Berlin:
Language Science Press. DOI:10.5281/zenodo.495446
Martin Maiden
as alternation in vowel quality, phonologically conditioned by stress. The distinctions
between stressed and unstressed positions may additionally manifest as differences in
the number of syllables in the root, and in sundry consonantal alternations, including
metathesis. In fact, what Savognin (and Romansh generally) exhibits is an unusually
florid manifestation (in respect of the range of different alternation types involved) of a
pattern of alternation recurrently attested across Romance languages, and I have argued
extensively (e.g., Maiden 2005; 2011c) that it arose historically because of stress-related
vowel differentiation, but then became “autonomously morphological” or “morphomic”
in nature, being no longer determined by stress, but simply by the heterogeneous set of
paradigm cells, one property of which is that they are “rhizotonic” (i.e., stressed on the
lexical root). This set I label (for reasons that are here unimportant) the “N-pattern”. Important diachronic proof that the N-pattern is independent of phonological causation is
the rise in various Romance languages of alternation patterns whose phonological content cannot possibly ever have been determined by stress (including lexical suppletion),
but which replicates the pattern originally “etched out” by the effects of stress. The
distribution of alternation found in Savognin constitutes a widespread variant of this
“N-pattern”. 1 There is no space to recapitulate all the data and arguments, but crucially
Anderson (e.g., 2010: 25;2011: 34f;2013: 10,16,23;2011: 173) accepts the “morphomic” nature of the “N-pattern” and its importance in determining morphological change in the
Romance verb generally. He believes, however, that modern Savognin (and other Romansh varieties: cf. Anderson 2013) is, in effect, a “special case”, indeed a prime example
of ‘phonologically conditioned allomorphy’. He convincingly shows that the alternation
types involved have often become so disparate, and refractory to unique underlying representation, that they must be represented directly in the grammar, but he also argues
that the alternant-pairs characteristically contain one member more suited to appear under stress, and another more suited to appear in unstressed position (Anderson 2011b: 18),
and that the alternants are accordingly conditioned by the presence vs absence of stress.
He also shows that the position of stress in Savognin is systematically predictable from
the segmental content of the word-form. It follows that the alternations are fully predictable from the position of stress, and that appeal to the “N-pattern” is inappropriate.
My response, in nuce, has been that there exist nonetheless within Savognin morphological phenomena that really are irreducibly “morphomic” and follow the N-pattern. Given
this, I say that Anderson misses an important generalization by divorcing the vocalic
alternations from clear-cut cases of the N-pattern. Anderson’s response is, in effect, that
my alleged N-pattern examples are secondary effects of the principle of stress-related
allomorph selection, and that invocation of the morphome risks missing another significant generalization, namely that the alleged stress-related alternations found in the verb
are found across the grammar, outside the verb.
I need to comment briefly on a methodological assumption. Steve Anderson objected
to me orally some years ago (cf. also Anderson 2011b: 13f.;17) that I could not draw in1
Savognin is among a number of Romance dialects where the N-pattern alternant also appears in the first
and second persons plural present subjunctive. The reasons are complex and not immediately relevant here
(cf. Maiden 2012).
10 Romansh allomorphy (Again!)
ferences about Savognin from apparently parallel developments in other Romance languages for which my “morphomic” analysis seemed correct, observing quite reasonably
that the inhabitants of Savognin were “not Romance linguists” (and therefore could not
know about what happens in other Romance languages). My very delayed response is,
in effect: “Yes, but they are still Romance speakers”. Let us suppose that in some other
Romance variety, distant both historically and geographically (a real case is Romanian),
virtually identical patterns of alternation are found, except that this time they are clearly
morphomic; let us further assume that the analysis appropriate to, say, Romanian is perfectly possible for Romansh, even though rival possible accounts exist. Obviously native
speakers of Romansh are not native speakers of Romanian, nor are they Romance linguists enjoying Olympian vistas over comparative Romance morphology: nothing we
could say about Romanian could ever be definitively probative for Romansh. What does
not follow, however, is that the comparative evidence can be ignored. Speakers of both
languages obviously have the same mental endowment, and both languages have inherited much that is structurally common, particularly with regard to the organization of
the inflexional morphology of the verb. What this means is that an analysis which is
justified for Romanian deserves consideration as plausible for Romansh. The Romance
languages ought not to be treated as hermetically isolated entities: rather, the analysis
of one variety should always be allowed to inform that of another. That, in fact, is one of
the reasons for doing Romance linguistics from a comparative perspective (in fact, there
is no other way of doing it), and in the following pages the analysis will frequently be
guided, with all due caution, by comparisons and inferences across cognate but separate
varieties. I now examine the facts which seem to me to require acknowledgement of the
morphomic N-pattern in Savognin.
2 Suppletion: dueir and deir
The verb dueir ‘must, be obliged to (cf. German sollen)’ (from Latin debere) plays a
central role in the debate, because it has a suppletive root allomorph in precisely that set
of cells which, in other verbs, displays the “stressed” alternant (Table 1):
Table 1: Dueir in Savognin
Martin Maiden
The suppletive forms are taken from another verb, stueir ‘must, be necessarily the case
(cf. German müssen)’, which unlike dueir continues to have its own complete paradigm
(Table 2):
Table 2: Stueir in Savognin
prs. ind
prs. sbjv
For Anderson (2008; 2010) dueir is a defective verb and its pattern of alternation is
a matter of phonologically conditioned allomorphy: he believes that the explanation of
the suppletion is that dueir lacks a stressed stem-alternant, having only unstressed /dʊ/.
Since /dʊ/ contains a vowel whose phonological characteristics debar it from occurring
under stress, speakers in effect plug the resultant phonological gap by borrowing appropriate stressed forms from a near synonym of dueir, namely stueir. My view, from
comparative and diachronic evidence, is that it is highly unlikely that dueir could ever
have been, in any relevant sense, “defective”, and that even if it were, the filling of the alleged gap could have nothing to do with phonology. Indeed, any explanation in terms of
phonological conditioning crucially fails to account for the fine details of the allomorphy.
If I am correct, and what we observe is a pattern of allomorphy identical in distribution
to the vocalic alternations, yet independent of phonology, then in principle whatever
explains the paradigmatic distribution of forms of dueir should also be available to explain the vocalic alternations. Indeed, considerations of economy would lead us to prefer
that single explanation. This is a view that I have expounded before (e.g., Maiden 2011b:
46–49), while Anderson, in his latest discussion 2013: 8 states that there are no new
facts, and that we simply disagree. I think that the facts remain very important, and I
(re-)present them below in a slightly revised form.
The ‘defectiveness’ of dueir is the effect, not the cause, of the suppletion. All suppletive verbs whose morphology reflects the incursion of forms of the paradigm of one
lexeme on the paradigm of another are, in one sense, “defective”. If, for example, Grisch
(1939: 89f.n5), DRG s.v. dovair, p.378, Decurtins (1958: 155;158), or Signorell, WuethrichGrisch & Simeon (1987: 165f), describe Romansh reflexes of debere as “defective”, this
simply means that there are parts of the paradigm occupied by forms which are patently
unconnected (diachronically or synchronically) with dueir, and which obviously are connected with stueir. This does not mean that the paradigm of the lexeme meaning ‘must’
somehow has “holes” in it.
10 Romansh allomorphy (Again!)
One might object that the independent existence of stueir as a verb with its own full
paradigm (and indeed with its own distinctive meaning, as I explain shortly) is grounds
to view those forms of it which appear inside dueir as synchronic interlopers drafted in
to occupy otherwise “empty territory”. Such reasoning would force us into the highly
counterintuitive position of claiming, for example, that Savognin esser ‘to be’ is “defective” in respect of its past participle, because the latter has the form sto which is also
(and transparently) the past participle of a different verb, star ‘to stand’. This is actually
a case where there was, historically, defectiveness: the Latin ancestor of esser had no
past participle, for semantic reasons. It is only with the rise in early Romance of verbal
periphrases comprising auxiliary + past participle, that the verb ‘be’ needs to fill the past
participle “slot”, and it does so (in some regions) by supplying the missing form from the
past participle of stare ‘stand’. But the idea that, in modern Romansh, the verb ‘be’
lacks a past participle and “borrows” it from star seems peculiar, given that the verb ‘be’
itself, and the use of forms of it involving its past participle, are utterly basic. Indeed, to
the best of my knowledge no grammarian of the Romance languages has ever described
the wholesale suppletion of ‘be’ in the Romance languages as involving “defectiveness”.
If one can analyse Savognin esser as suppletive but not defective, then surely the same
analysis should be available for dueir.
My principal difficulty with Anderson’s analysis of dueir is that I see absolutely no
motivation to view this verb as defective, beyond the morphological facts which are the
explanandum. All its cells are well and truly filled — and it is almost inconceivable that
a subset of present-tense cells of a verb expressing such a basic meaning could ever be
empty. Here, again, a comparative perspective is useful. Virtually all Romance languages
conserve reflexes of debere, with a full inflexional paradigm, and no Romance languages
show any sign of defectiveness in the verb meaning ‘must’, whatever its origin: there is
no reason for parts of its paradigm to be missing. Many Romansh dialects indeed have
a full paradigm all of whose forms still continue debere (see Decurtins 1958: 152f. DRG
s.v. dovair), and the rhizotonic forms (usually dé-) are robustly attested from the earliest
records, including in central dialects (of which Savognin is one); cf. also Anderson (2010:
30;32). There is simply nothing in the phonology of these dialects, either synchronically
or diachronically, which could have determined deletion of such stressed forms, and the
defectiveness certainly cannot be explained as a phonological effect of stress.2
Anderson (2010: 32) suggests that “the primary factor in the emergence of defectiveness in Surmiran dueir, as well as the complementary pattern in the Engadine languages,
was the morphologization of the vowel alternations in Swiss Rumantsch. If we hypothesize that this was combined with reduced use of the verb due to competition with others such as stueir, it could well have led to the present situation with only one stem
Could a stress-based account be salvaged if, unlike Anderson, one said that any kind of alternation, including an alternation where one of the alternants was zero, could be effected by stress? Given that the
position of stress in Savognin is predictable on grounds of segmental phonological content, one can hardly
invoke the case of zero alternants which would, by definition, lack any segmental content. The best one
could say is that zero forms appear in those parts of the paradigm where stress would normally be expected
to appear. But then one would have to ask: “Where would stress normally be expected to appear?”, and
the answer would be purely morphological: “the N-pattern”.
Martin Maiden
conserved”. I discuss later the nature of the “competition” from stueir, which involves
overlap and replacement, not defectiveness. As for the alternation, regular sound change
would indeed have given rise to a unique alternation between a stressed alternant /de/,
and unstressed alternant /du/ (in the latter the back vowel is the result of an adjustment of the unstressed front vowel triggered in the environment of a following labial
consonant: cf. Italian 3sg.prs.ind déve vs inf dovére). But the notion that Romansh
would eliminate an alternation type because it was “morphologized” (or, perhaps better, idiosyncratic and unpredictable), especially in such a high-frequency verb, seems
unlikely, particularly given that Romansh is notable for retaining extreme and idiosyncratic patterns of vocalic alternation, even in isolated verbs which surely have a much
lower frequency of use. Rogers (1972), in an analysis of Surselvan, lists no fewer than
eleven sets of vocalic alternation each apparently limited to just one verb, with meanings
such as ‘vomit’, ‘scythe’, ‘drivel’ - all without sign of resort to defectiveness; see also my
discussion of Savognin deir ‘say’, below. One might add that the most natural response
to any idiosyncratic type of alternation would surely be not to create a “gap”, by jettisoning one alternant, but to attempt some kind of “repair” by analogically remodelling
the alternants to be less different.3 The last thing one expects is a reaction resulting in
an alternation which, by virtue of being suppletive, is even stranger than the rejected
Viewed in a comparative-historical perspective, the notion that the Savognin reflexes
of debere could be in any significant sense “defective” seems most unlikely. And even if
it were defective, the suppletive filling of the alleged gaps would take place because the
gaps needed filling, not specifically because of the lack of a “stressed” alternant. What
that perspective does suggest, however, is a different scenario (see also Maiden 2011b:
46–49), which involves not “gaps” but “overabundance” (cf. Thornton 2011), in which
more than one realization became available for certain cells of the paradigm of dueir,
and in which one of the alternative realizations ultimately prevails. This situation arises
from particular discourse-related circumstances. The reflexes of debere are subject in
Romansh and beyond to intensive competition from other nearly synonymous alternatives.4 I have no space to detail the facts or the mechanisms (see Maiden 2011b),5 but
essentially what appears to be involved is “face-saving”: speakers avoid the charge of
moral obligation inherent especially in present indicative forms of dueir by resorting to
alternatives such as expressions equivalent to ‘ought’ (e.g., conditional forms of the verb),
or expressions meaning “absolute” (rather then “moral”) necessity, which is exactly expressed by stueir. This tendency created, I suggested, a situation in which the frequent
use of stueir alongside dueir in the present tense led to effective synonymy between the
two forms, eventually resolved by replacing dueir with stueir according to the familiar
pattern of alternation (the “N-pattern”) associated with vocalic allomorphy (a type of
In any case, Savognin stems do sometimes have “inappropriate” forms. Thus Anderson (2011b: 32) discusses verbs such baitár ‘babble’ which has a stem suitable for stress only, but which is is nonetheless used
throughout the paradigm in unstressed environments as well.
See further Stürzinger (1879: 49); Tagliavini (1926: 84); Kramer (1976: 64); Maiden (2011a)
In addition to the sources cited by Maiden (2011a), see also Pult (1897: 166f.) for the suppletive introduction
of forms of the verb ‘want’ in the dialect of Sent.
10 Romansh allomorphy (Again!)
reaction to synonymy attested elsewhere in Romance: cf.Maiden 2004; 2006). The same
paradigmatic distribution, reflexes of debere alternating this time with those of conuenire (originally meaning ‘be fitting’), emerges from ALDII (maps 829–833; 836–838),
for Peio (point 54) and S. Michele all’Adige (point 66). The disappearance of reflexes of
debere from certain cells of the present indicative and the present subjunctive of dueir
never involved “defectiveness”, and has never had anything to do with phonology. The
perception of “defectiveness” is a synchronic effect of the suppletion.
Crucially, the suppletive fusion of dueir and stueir in Savognin is of a significantly
different kind from the alleged phonologically conditioned allomorphy of the vocalic
alternations. The latter is a binary correspondence between alternants and stress: one
alternant is selected under stress, the other elsewhere. The putative relation between the
dueir - stueir alternation and stress can only be described as binary at a level which is in
fact lacking any phonological content. For what, in the case of the suppletion, is allegedly
selected by stress is not a form correlated with stress, but a whole array of phonologically and morphologically different forms. As Anderson himself points out 2008: 124,6
“it is not just a single stem, but the full range of irregular forms of stueir (ia stò, te stost, el
stò, els ston; Subjunctive ia stoptga, etc.) that replaces those of dueir where stress would
fall on the stem”.7 More exactly: “the first and second person singular, and third person,
forms of the indicative of stueir are mapped onto the first and second person singular,
and third person, forms of the indicative of dueir, and the present subjunctive cells of
stueir are mapped onto the corresponding cells of dueir”. Only this way do we get the
observed distribution. In effect, it is not “a stem”, but an entire, morphomically defined,
“slab”, of the paradigm of stueir, a set of full word-forms replete with their own internal
allomorphic idiosyncrasies, that has been mapped onto dueir. A rule of phonologically
conditioned allomorphy involving stress could in principle select a stressed root allomorph of stueir and introduce it into dueir, but it could not necessarily insert the right
root allomorph in the relevant cell. A rule that identifies a morphologically-defined portion of the paradigm as that in which the replacement of one lexeme by the other can
do just that.
Dueir exemplifies lexical suppletion (“incursion”, in the terminology of Corbett 2007),
where one historically distinct lexeme obtrudes on the inflexional paradigm of another
lexeme. Another diachronic route to suppletion is regular sound change so extreme
in its effects that the synchronic result is allomorphy such that the alternants bear no
phonological relation to each other. Savognin has at least one case of phonologically
induced suppletion, namely deir ‘say’, which inflects as in Table 3:
Anderson (2010: 29) even calls this verb “suppletive”.
As mentioned earlier, if we ask “where stress would fall on the stem”, the answer is ineluctably morphological: in the cells of the singular and third person present and imperative and in the present subjunctive.
It seems to me useless to say, instead, “wherever the endings would be unstressed” because the third person singular has no ending, and the distribution of the remaining, unstressed, endings turns out to be the
morphomic N-pattern.
Martin Maiden
Table 3: Deir in Savognin
prs. ind
prs. sbjv
I cannot here retrace the phonological history of this verb in detail. Suffice to say
that the historically underlying root was *dik-, and that sound changes involving, in particular, deletion of unstressed vowels and assimilation of resulting consonant clusters,
created the modern situation. The rather unusual present subjunctive of this verb happens to show the effects of analogical levelling in favour of the originally arrhizotonic
first and second person plural form stems togther with the associated stressed inflexional
ending (cf. Decurtins 1958 for the reflexes of ambulare / ire, dicere, uenire, habere, or
sapere in Samedan, Parsons, and Razen for other Romansh examples of this kind; further
discussion also in Maiden, in progress). That aside (and there is every reason to believe
that the déi- root originally occurred as expected in the present subjunctive), this verb
shows N-pattern suppletion. Anderson (2011b: 17) acknowledges that it is “genuinely
suppletive” and that “the choice of stem is determined by morphosyntactic features”. He
defines in the same way some other, less radically suppletive verbs, for which I give here
just the present-tense forms, e.g, (vu)léir ‘want’, néir ‘come’ (Table 4):
Table 4: (Vu)leir and neir in Savognin
prs. ind
prs. sbjv
prs. ind
prs. sbjv
If we acknowledge that deir and some other verbs have (near-)suppletive patterns determined synchronically by morphosyntactic features, then we have to admit the presence of the morphomic N-pattern in Savognin. Yet if we say that the vocalic alternations
10 Romansh allomorphy (Again!)
are still a matter of “phonologically conditioned allomorphy”, then the fact that they
show exactly the same paradigmatic distribution becomes uncomfortably coincidental.
3 The “augment”
The “augment” is a functionally empty formative which, in certain cells of the inflexional
paradigm of the verb, occurs between the lexical root and desinences denoting tense,
mood, person, and number (for detailed discussions of its nature and origins, which lie
in Latin and proto-Romance Aktionsart suffixes, see especially Maiden 2003;2011: 249–
53;2016: 715f.). In Latin, the relevant affixes were restricted to imperfective-aspect forms,
but had no restrictions for person, number, or tense. In most Romance languages, augments are associated with particular inflexion classes (in Romansh, usually the first and
fourth conjugations), and have become restricted to certain cells of the inflexional paradigm defined by tense, mood, person, and number. In Savognin, the augment occurs
solely in the singular and third person forms of the present indicative, and in all forms
of the present subjunctive. That is to say, of course, that it has exactly the same paradigmatic distribution as the “stressed” vocalic alternants, a fact which clearly suggests a link
between them. Thus first conjugation luschardár ‘strut’, and fourth conjugation tradéir
‘betray’ (Table 5):
Table 5: The augment in Savognin first and fourth conjugation verbs
First conjugation
prs. ind
prs. sbjv
Fourth conjugation
prs. ind
prs. sbjv
It is undisputed that the distribution of the Romance augment cannot be explained,
diachronically or in modern varieties, as the output of any kind of phonological process.
The view that I have developed (see, e.g., Maiden 2005; 2011b,c) is that the redistribution of the alternant from Latin to Romance is purely morphologically determined, and
reflects sensitivity to a paradigmatic pattern created, originally, as an effect of vocalic
alternations between stressed and unstressed vowels: the same pattern can be shown
to have provided a “template” for the suppletive merger of distinct lexical verbs in various Romance languages (notably, the verb ‘go’). I see no reason why what we see in
Savognin, and more generally in Romansh, should be viewed any differently: the distribution of the augment appears a matter of pure morphology, and given that the vocalic
alternations have the same distribution as the augment, they too can be accounted for
in the same, purely morphological, terms.
Martin Maiden
Anderson views the facts, in effect, in terms of a kind of “defectiveness”: verbs showing the augment lack a stressed vocalic alternant, and the augment is inserted wherever
this occurs. Since the augment is inherently stressed, the preceding root-form must be
unstressed, and the lack of a stressed root allomorph is thereby resolved. My view is
that this analysis inverts cause and effect: it is not the case that the augment is applied
because there is no stressed root allomorph but, rather, that there is no stressed root allomorph because the relevant cells of the paradigm are specified as being filled by forms
containing the augment. This latter analysis has the immediate advantage of avoiding
the problem of arbitrary stipulation of defectiveness in one set of cells only. After all, if
stressed alternants can be defective, why should not unstressed alternants too? Why do
we not also find, that is, verbs with a stressed alternant but not an unstressed one? And
if the distribution of the augment is dictated by the need to plug a phonological “gap”,
how is it that such gaps only occur in first and fourth conjugation verbs, precisely the
inflexion classes to which the augments are historically restricted across the Romance?
Discussion of the Savognin augment has tended to focus on first conjugation verbs,
where it is most productive, but where it still only constitutes a subset (and apparently a
minority) of such verbs. We should not forget that the augment also appears in the great
majority of fourth conjugation verbs (characterized by infinitives in -eir), a class comprising dozens of lexemes and endowed with some productivity. If we follow Anderson,
this means that almost all of the fourth conjugation is characterized by lack of a stressed
alternant. Nothing logically excludes this, but it seems counterintuitive to say that a
major, semi-productive, inflexion class is, in effect, “defective” in most of its present
tense. Nobody would countenance such an analysis for the cognate Romance varieties
(Daco-Romance, Italo-Romance, Catalan) where the fourth conjugation behaves in this
Anderson (2011b: 22) points out that the augment frequently appears in neologisms,
including where speakers feel doubt about the identity of the stressed allomorph. This
does not mean, however, that the augment is usually a response to perceived lack of a
stressed alternant. Using the Savognin first conjugation verb luschardár ‘strut’ (exemplified above), Anderson (2008: 122) observes that: “The use of this pattern […] has the
advantage that the speaker does not need to retrieve any information about the specific
alternation pattern of the stem in order to produce all of the correct forms. Otherwise,
it would be necessary to choose […] among a variety of possibilities such as *luscharda,
*luscheirda, *luschorda, *laschurda, *laschorda, etc. Each of these patterns is more or less
secure with reference to at least some verbs in the Surmiran lexicon, but the availability of the paradigm [given above] makes it possible to avoid the choice when positive
evidence is not readily available.” The problem here is that that there is unequivocal evidence for the stressed vowel. This verb is transparently and directly derived from the
nominal luschárd ‘dandy, fop, vain, proud’, which actually contains, moreover, a highly
frequent stressed pejorative suffix -árd. In this case, the identity of the “right” stressed
alternant is patent. This is in fact true of a large number of other verbs that take -esch,
all transparently derived from nouns or adjectives whose stressed vowel is known (examples from Signorell 2001), such as those give in Table 6:
10 Romansh allomorphy (Again!)
Table 6: Nouns, adjectives, and derived verbs in Savognin
Basic noun/adjective
‘dwell (-ing)’
‘sign (-ature)’
‘paint (-brush)’
What such derived forms lack is not a “stressed” alternant but an unstressed one: what
appears in the verb is simply the root of the corresponding rhizotonic nominal form.
Yet there is no sign of attempts to invent a predictable “unstressed” counterpart for the
stressed vowel of the base-form (e.g., inf **udiiér from the noun ódi after the model
of 3sg.prs.ind dórma ‘sleeps’ vs inf durméir) and quite simply the derived verb-forms
preserve the segmental identity of the base form. Many scholars have suggested that in
Romance generally the augment serves to obviate root allomorphy that might otherwise
occur in the lexical root. There are various reasons why this view does not account
for the facts (cf. Maiden 2011c: 251f.), but note that in any case this kind of “solution”
comports a paradox: one type of alternation is merely replaced by another, that between
the augment and its absence, the augment itself retaining an irreducibly “N-pattern”,
morphomic, distribution.
Anderson (2013)broadens his survey beyond Savognin, arguing for a similar analysis
for other Romansh varieties. In fact, in dialects where stress has a somewhat different
distribution from Savognin, the augment duly follows that distribution. Thus Anderson
(2013: 21f.), in Vaz (Valbella; data from Ebneter 1981) the first person plural present indicative, in addition to arrhizotonic forms, as in Savognin, also has rhizotonic forms with unstressed desinence -an, and the predicted “stressed” stem: e.g., inf amblidár ‘forget’, 1sg
In fact, dimóra and fírma below may be derived from the corresponding verbs (cf. Thornton 2004: 517, and
Cortelazzo & Paolo 1988 ss.vv. dimorare, firmare for this type in Italian). If this holds for Savognin, then
these verbs do possess a ‘stressed’ alternant.
Martin Maiden
amblóid, 1pl amblidáin, but inf amprastár ‘lend’, 1sg amprést, 1pl ampréstan. Sometimes,
the same verb has two possible first person plural present indicative forms, e.g.: inf
gudáir ‘enjoy’, 1sg giód, 1pl gudáin / giódan. In verbs taking the augment, there are, correspondingly, forms in 1pl -áin without augment (e.g., inf adorár ‘adore’, 1sg adorésch,
1pl adoráin), and forms in 1pl unstressed -an duly showing it (e.g., inf habitár ‘inhabit’,
1sg habitésch, 1pl habitéschan). According to Anderson (2013: 23) such behaviour poses a
“problem” for the morphomic account, because “it is fairly clear that the stem alternation
and the appearance of -esch […] are tied directly to the position of stress, even where this
is potentially variable, and not to a fixed set of morphological categories”. With respect
to Anderson, I think that it poses no such problem: all it shows is that whatever principle
governs the stressed root allomorph also governs the augment. In this particular case,
in fact, we are not dealing with a change in position of stress at all: rather, we have a
syncretism such that the first person plural form tends to be “taken over” by the third
person plural. This is quite systematic in Vaz (and elsewhere), and occurs independently
of stem stress (for example, in non-present forms).
More revealing is the case presented in Maiden (2011b: 45f.), of distributional discrepancy between augment and stressed vocalic alternant. The dialects of the Val Müstair (see
Schorta 1938: 132) tend to place stress on the root of the infinitive in all conjugations.9
Indeed, they are unique among Romance languages in generally having rhizotonic infinitives even in the first conjugation, as shown in Table 7:10
Table 7: Rhizotony in Val Müstair first conjugation infinitives
ˈlai ̯dər
ˈjai ̯ntər (or janˈtar)
‘spread dung’
Schorta states, however, that root stress in first conjugation infinitives systematically
fails to happen in that class of first conjugation verbs having the element -aj-, which I
Some second conjugation verbs are exceptions.
This phenomenon is mentioned by Stürzinger (1879: 35) and Huonder (1901: 518f.), and is amply confirmed
by data from Val Müstair covered by ALDI /II. See also Grisch (1939: 222) for Vaz, Candrian (1900: 51) for
Stalla, Solèr (1991: 135) for Schams.
10 Romansh allomorphy (Again!)
identify as the augment.11 Here, stress always remains on the ending of the infinitive:
e.g., inf bɐˈtjar ‘baptize’, 3sg.prs bɐˈtjaja; inf biˈar ‘build’, 3sg.prs biˈaja; inf guˈjar ‘dare’,
3sg.prs guˈjaja. The same holds of fourth conjugation infinitives, but apparently only
if they belong to that minority of verbs that lack the augment: Schorta cites inf fiˈnir
’finish’, a verb which takes the augment; compare this with, e.g., inf ˈbwɔʎər ‘boil’ a
verb that does not show the augment). Now the most likely explanation of why the
augment does not appear in the infinitive here is that, in Romance languages generally,
root-stress in infinitives is limited to third conjugation verbs (cf. Maiden 2011c: 201f.
2016: 509), all other classes having non-rhizotonic infinitives. The augment, however, is
characteristic solely of the fourth and first conjugations, not the third. The third, while
a relatively small and unproductive class, contains some of the semantically most basic,
and highest frequency, verbs, and the appearance of root stress in the Val Müstair fourth
and first conjugations is almost certainly modelled, therefore, on the third conjugation.
If no augment can appear in first conjugation infinitives, it is because the distribution
of the augment is morphologically specified, and that specification happens to exclude
In Maiden (2011b) I argued as follows: if the function of the augment is in effect to
supplete the absence of a stressed root-alternant, and if infinitives in the Val Müstair are
generally root-stressed, then we should expect verbs with the augment duly to show that
augment in the infinitive, in lieu of root-stress (e.g., inf **bɐˈtjajər rather than the actually occurring bɐˈtjar). The fact that the augment never appears in the infinitive therefore
also suggests that its distribution is independent of stress, and purely morphologically
specified. The only way to “save” the stress-based account (and this is what Anderson
2013 does) is to claim that the augment is inherently limited to “tensed” forms, and is
therefore not available for the infinitive. He observes, in support of his view, that it does
not occur, either, in participles or in “related non-verbal forms”. Since the augment also
appears in second person singular imperatives, it is not clear that “tensed” is quite the
right term: it might be more accurate to say that the domain of the augment involves
cells with values for person and number. But now Anderson’s claim must be that the
augment is selected in those parts of the paradigm specified for person and number
where stress would otherwise fall on the root, while my account can easily be reformulated, if we wish, as also applying to those parts of the paradigm specified for person and
I must acknowledge here a different, and very careful, analysis of these facts by Kaye (2015: 291–310),
for whom -aj- is not the ’augment’ but part of the stressed lexical root of the verb, whose historically
regular unstressed counterpart is -j- (*bateˈdjare > bɐˈtjar; *baˈtedja > *bɐˈtaja). In a form such as bɐˈtjaja, on
Kaye’s account, the unstressed element -j- has been analogically generalized into the root of the stressed
alternant, originally of the type *bɐˈtaja < *baˈtedja (Kaye 2015: 307); resistance of bɐˈtjar (< *bateˈdjare)
to stress retraction is then, Kaye suggests2015: 309, a function of the degree of phonological difference
between stressed and unstressed root. I would observe that -aj- is exactly the expected reflex of the protoform of the augment (although rarely attested in the rest of Romansh, where is has been supplanted by
-eʃ-), and that it is not clear why the root found in the root-stressed present-tense forms of the verb would
be phonologically disfavoured in root-stressed infinitives. In fact, even if -aj- turns not to be, in origin,
an “augment”, such an analysis suggests that speakers have effectively analysed bɐtj- as the lexical root,
treating -aj- as a kind of excrescent element following it, and one that occurs just in the N-pattern cells.
That is to say that its synchronic status is equivalent to that of the augment in other verbs.
Martin Maiden
number, except for the first and second persons plural present indicative. Both accounts
acknowledge that the phenomenon is heavily morphologized, and applies over a domain
whose definition corresponds to no natural phonological or morphosyntactic class. Even
Anderson’s account is, I submit, implicitly “morphomic”.
4 The generality of the alternations: derivational
Anderson’s analysis gains support from the fact that the vocalic alternations that occur
within the verb also occur outside it: nouns and adjectives with stressed derivational
affixes show the corresponding “unstressed” vocalic alternants in the derived forms. The
sensitivity of these alternations to stress is thereby argued to be a general property of the
grammar, and not a peculiarity of verb morphology.morphology Take, for example, the
behaviour of the vocalic alternants in derivational morphology (Table 8), as presented
by Anderson (2011: 28–30;2013: 13–17):
Table 8: Vocalic alternants in Savognin derivational morphology
guttár ‘drip’
liiér ‘bind’
néiver ‘snow’
Basic noun
Derived nouns
gót ‘drop’
léia ‘union’
néiv ‘snow’
gutélla ‘drip’
liadéira ‘binding’
naváglia ‘big snowfall’
In reality, this might be no more than the residual, and synchronically more or less
accidental, effect of historical differentiation of vowel quality according to stress. This is
suggested by the fact that there are derived forms with stressed suffixes (and therefore
with unstressed roots) where, nonetheless, the stressed alternant occurs. Thus Table 9:
Table 9: Discrepancy between vowel alternation in verbs and derived forms
Derived forms
satgér ‘dry’
accumpagnér ‘accompany’
durméir ‘sleep’
10 Romansh allomorphy (Again!)
Similar phenomena recur elsewhere in Romansh, as Anderson (2013: 20) points out:
thus Vallader has scóula ‘school’, scolár ‘to school’ (3sg.prs.ind scóula), scolaziún ‘education’, yet diminutives scoulína ‘kindergarten’, scoulétta ‘craft school’. Anderson (2011b:
28) deals with such apparent counterexamples by assuming an architecture of the grammar in which morphology and phonology “interact cyclically (with some appropriate
subsystem) of the phonology applying to adjust the results of each stage of morphological elaboration of a form”. The selection of the stressed or unstressed stem alternant
operates only during the “first cycle”; once the stem-shape is determined ‘the decision
is not revisited on subsequent cycles’. A “stressed” base will then persist through later
cycles, even if it is itself no longer stressed: the derivational counterexamples can now
be explained in terms of the cycle on which they occur.
One immediate objection is that saying that an apparently phonological phenomenon
is confined to a particular “cycle” is in fact to concede that it is “morphologized” (the
cycles being defined, precisely, as stages of “morphological elaboration of a form”), and
restriction of a phonological rule to some morphologically defined domain is to introduce into the analysis a considerable degree of arbitrariness (precisely one of the things
that for Anderson consitutes an objection to the purely “morphomic” analysis). Things
look even more arbitrary, and “morphological”, if we consider that we now have to say
that the domain of the phonologically conditioned allomorphy is defined over two quite
disparate sets of forms: “tensed” forms of the verb (at least in Vallader) and the “first
cycle” in derivation. A more fundamental difficulty is that it is not always true that the
“stressed” stem persists unchanged after the “first cycle”: why do we have, say, derived
forms accumpagnedér or durmiglión with “unstressed” alternants, yet accumpognamáint
or dormulént with “stressed” alternants?12 Actually, the predicted selection of the “unstressed” alternant usually occurs in words belonging to inherited vocabulary, but not
in neologisms, which led me to conclude (Maiden 2011b: 41) that what we have is evidence of the “death” of phonological selection of the allomorphs, now reflected only in
traditional vocabulary. This claim tends to be reflected in the behaviour of adjectives
showing reflexes of Latin -abilis, -ibilis (equivalents of English -able, -ible): the “popular” reflex by direct inheritance from Latin (-evel) displays the “unstressed” alternant
(e.g., ludével ‘praiseworthy’), while the “learned” -ábel/-íbel displays the “stressed” alternant (e.g., accumodábel ‘accommodatable’). One might, perhaps, want to assign -evel to
the “first cycle”, and -abel to the second, but even this does not work too well, for we
find occasional examples of the distinctive “unstressed” alternant with -ábel/-íbel: e.g.,
schliíbel ‘soluble’, bavábel ‘drinkable’, purtábel ‘portable’, duvrabel ‘usable’.
Anderson (2013: 15) observes an “asymmetry”, in that the counterexamples to his claim
all involve the appearance of a “stressed” stem that does not bear stress, while no examples exist in which an “unstressed” stem appears under stress. But this is not proof that
that the stem alternants are sensitive to stress. On such evidence as Anderson presents
(and from Signorell 2001), the small inventory of possibly derived forms involving a
stressed stem displays that stem simply because it is the phonologically regular result
of their etyma (e.g., prescháint ‘present’ (adj.) < praeséntem). In any case, it is perfectly
For an inconclusive discussion of these data, see Wolf (2013: 171).
Martin Maiden
possible that some forms such as prescháint are not “derived”, but are the base forms
from which the corresponding verbs are derived. Anderson also observes 2013: 18 that
in Surselvan (in fact, more widely) only “unstressed” root alternants appear in factive
verbs formed with the suffix -ent- or -ant-,13 where the lexical root is systematically unstressed (e.g., inf béiber ‘drink’, 1pl.prs buéin, but factive inf buentár ‘cause to drink’,
never **beibentár). He cites similar phenomena in Puter (Anderson 2013: 10), such as
stanglantér from stáungel ‘tired’, but says that “[m]ore research is needed to establish
the generality of the phenomenon”. Here I concur: we need to be sure that selection of
the “unstressed” stem is synchronically productive and therefore psychologically real.
Otherwise, all we may have is the regular, lexicalized, outcome of old sound changes in
the relevant derived forms. In any case, there do seem to be examples of factive verbs in
-entar that bear the “stressed” root allomorph. Jaberg (1939: 291f.) lists Surselvan examples most of which indeed bear the “unstressed” allomorph, but also gives dormentár ‘put
to sleep’ (cf. inf durmír ‘sleep’, 1sg.prs.ind dorm), and a case in which the unstressed
vowel of the derived form, while phonologically plausible, does not correspond to that of
the base verb (scumpentár ‘cause to be saved, heal’, from inf scampár ‘save’, 3sg.prs.ind
Even leaving counterexamples aside (and I acknowledge that they are not many), it
is not obviously necessary to invoke stress: given that the vast majority of cells of the
inflexional paradigm of any Romansh verb are arrhizotonic, one might equally say that
what we call the “unstressed” stem is the default, on which affixally derived forms are
usually built. An apparent counterargument to such an approach might come (cf. Anderson 2013: 17) from the fact that, in cases of derivation where the stress falls on the root,
it is always the “stressed” allomorph that appears: e.g., Surmiran cumónd ‘order’ (cf. inf
cumandár, 1sg.prs.ind cumónd), clóm ‘call’ (cf. inf clamár, 1sg.prs.ind clóm), gartétg ‘success’ (cf. inf gartagér, 1sg.prs.ind gartétg), dórma ‘narcotic’ (cf. inf durmír, 3sg.prs.ind
dórma), stéma ‘esteem’ (cf. inf stimár, 1sg.prs.ind stéma). Significantly, however, Signorell, Wuethrich-Grisch & Simeon (1987: 51) describe such forms, pre-theoretically, as
“deriving from a finite form”: what we may have here is simply nominalization of first or
third person singular verb-forms (which happen to contain stressed roots), not a derivational process which specifies a stressed root, and thereby must select the “stressed”
alternant. In short, while it is indeed true that many patterns of root-alternation found
in the verb recur across the grammar, this is largely a historical residue, not necessarily
evidence of an active synchronic phonological principle.
5 Conclusion
Anderson (2013: 23) accepts that “morphological categories play a role (e.g. in constraining the appearance of -esch to tensed forms of first and fourth conjugation verbs)”, but
asserts that ‘there is no warrant for invoking the further step of complete and arbitrary
morphological categorization that would be implied by associating the variation with
Cf. Signorell, Wuethrich-Grisch & Simeon (1987: 103f.).
10 Romansh allomorphy (Again!)
a morphome’. I suggest that the data are in fact already inextricably permeated with
“arbitrary” morphological specifications: in Savognin, and in Romansh at large, the morphomic N-pattern is really present. The need to specify that the alleged phonological
principle only applies to “tensed” verb forms (for Val Müstair), or to certain levels of
derivational morphology makes that principle itself “arbitrary”. Given that the behaviour
of suppletion in dueir and the distribution of -esch are, as I have argued, incompatible
with the “phonological” account, attempts to account for the identically distributed vocalic alternations in phonological terms become superfluous. Finally, given that countless Romance varieties do have genuinely morphomic patterns of the kind attested in
Savognin, treatment of Savognin as a special case is what may be “unwarranted”. And
I do not think that Savognin can be presented as the perfect example of “phonologically conditioned allomorphy” that Anderson claims, and yet one must ask whether
Anderson’s insight – that right across the grammar there is a close correlation between
stress and the selection of alternants, might be at risk of being abandoned too lightly. My
criticism has been that there exist some cases where such an analysis does not “work”,
and that since we need independently to invoke the N-pattern even for Savognin, we
should do so for all types of alternation which follow that pattern. But there is an unspoken assumption here which may need to be challenged, and it involves what might
be described as the “ghettoization of the morphomic”. The classic examples of “morphomic” phenomena as adduced by Aronoff (1994) make the case for the existence of
a “morphomic level” of linguistic analysis precisely because they are not plausibly explicable as effects of phonological, syntactic, or semantic conditioning: they are cases
of “morphology by itself”. In morphologists’ enthusiasm to assert the existence of genuinely morphological phenomena, much weight has been placed on the notion of the
“autonomy” of morphology (witness the titles of Maiden 2005, or Maiden et al. 2011).
While there are very good reasons to proclaim loudly that “autonomously morphological” phenomena exist, the search for them should not become a reductivist obsession,
nor is there any good reason to suppose that there cannot exist phenomena which contain a very high degree of purely morphological determination, while yet also possessing
some degree of phonological or other conditioning.
The seeds of a possible compromise appear in Maiden (2013) (actually, in the same
volume as, and immediately following, Anderson 2013). This deals with patterns of consonantal alternation in Italian verbs historically caused by two different kinds of palatalization. Synchronically, the result is that phonologically quite disparate types of alternant tend overwhelmingly to conform to a common distribution such that one alternant
occurs in all (or most) forms of the present subjunctive, and in the first person singular
and third person plural forms of the present indicative, but nowhere else in the paradigm. There are powerful arguments (see, e.g., Maiden 1992,20112011: 205–63) to say that
this pattern has lost all phonological causation and is genuinely morphomic. In Maiden
(2009) I had been extremely critical of attempts by Burzio (2004) to force a synchronic
phonological analysis of the modern Italian facts, quite often by what is, in effect, the
illegitimate resurrection of long dead phonological conditioning environments. For the
Martin Maiden
most part, these are criticisms I stand by: it cannot be said too often that morphology
suffers from a kind of “phonologizing bias” which too readily dismisses morphological
analyses of the data, and is far too prone to give credence to phonologically-oriented accounts, even at the expense of postulating conditioning environments lacking plausible
synchronic justification. Burzio’s analysis appeared to me an example of this kind, but
observation of some of the fine details of the diachrony of the alternations at issue later
caused me to moderate my view.
One type of alternation involved an opposition between velar consonants and palatal
affricates, the latter arising, historically, through palatalization and affrication of velars
before front vowels. Now it is beyond reasonable doubt that there has existed no productive process of palatalization/affrication of velars before front vowels for over a millennium, such a putative process being massively counterexemplified by the existence of
unmodified velars before front vowels from the time of the earliest documents, including within the paradigm of the verbs at issue. The principal fact14 which made me revise
(Maiden 2013) the “morphological exclusivism” of my earlier treatments of the subject,
however, was the observation that in medieval Italian a certain type of analogical innovation affecting verbs displaying the relevant types of alternation, whereby the root of the
present subjunctive was optionally extended into gerund forms with the ending -endo,
was strikingly, and systematically, blocked just where the result would have been a velar
consonant followed by a front vowel. Informally: “don’t allow a velar alternant before
a front vowel if an alternative (and more phonologically “natural”) palatal alternant is
also available”. Thus Table 10 (where “**” means “not occurring”):
Table 10: Analogically reformed subjunctives in old Tuscan
inherited gerund
gerund analogically reformed on subjunctive
possa ‘can’
ve/ dʤ/a ‘see’
te/ ɲɲ/a ‘hold’
abbia ‘have’
pia/ tʧ/a ‘please’
di/ k/a ‘say’
pian/ɡ/a ‘weep’
pia/ ʧ/endo
di/ ʧ/endo
ve/ dʤ/endo
te/ ɲɲ/endo
pia/ tʧ/endo
**di/ k/endo
While it would have been impossible to explain the distribution of the alternants in
purely phonological terms (for the reasons, see Maiden 2013: 25–31), this behaviour
clearly suggests residual sensitivity to phonologically plausible environments for the distribution of certain alternants.
In short, while Anderson’s analysis of the Savognin data seems to me too “phonologizing”, it may be that my own approach, insisting on purely morphological aspects
But see also Maiden (2013: 31–35).
10 Romansh allomorphy (Again!)
of the phenomenon, has been too “morphologizing”, and both approaches seem to be
subject to the questionable assumption that redundancy must be eliminated from the
analysis. I do not think that Savognin is as “pure” an example of “phonologically conditioned allomorphy” as Anderson believes, but the possibility that speakers are sensitive
to the recurrent correlation between certain types of alternation and stress should not
be sacrificed too hastily on the altar of formal economy. As Sims (2015: 205f.) observes,
our two approaches need not in fact be mutually exclusive. We have probably reached
the point where only appropriately devised psycholinguistic experimentation would be
able to tell us more about the Savognin data. However that may be, morphologists and
Romance linguists are truly in Steve Anderson’s debt for having focused our attention
so sharply on these fascinating data.
Jaberg & Jud (1964)
Goebl et al. (1998)
Goebl et al. (2012)
I would like to thank Hans-Olav Enger for some very helpful remarks on an earlier draft
of this chapter.
Anderson, Stephen R. 2008. Phonologically conditioned allomorphy in the morphology
of Surmiran (Rumantsch). Word Structure 1(2).
Anderson, Stephen R. 2010. Failing one’s obligations: Defectiveness in Rumantsch reflexes of DĒBĒRE. In Dunstan Brown Matthew Baerman Greville G. Corbett (ed.), Defective paradigms: Missing forms and what they tell us, 19–36. Oxford: British Academy
& Oxford University Press.
Anderson, Stephen R. 2011a. Romansh (Rumantsch). In Adam Ledgeway & Martin
Maiden (eds.), The oxford guide to the Romance languages, 169–184. Oxford: Oxford
University Press.
Anderson, Stephen R. 2011b. Stress-conditioned allomorphy in Surmiran (Rumantsch). In
Maria Goldbach, Marc-Olivier Hinzelin, Martin Maiden & John Charles Smith (eds.),
Morphological autonomy: Perspectives from Romance inflectional morphology, 13–35.
Oxford: Oxford University Press.
Anderson, Stephen R. 2013. Stem alternations in Swiss Rumantsch. In Silvio Cruschina,
Martin Maiden & John Charles Smith (eds.), The boundaries of pure morphology: Diachronic and synchronic perspectives (Oxford Studies in Diachronic and Historical Linguistics 4), 8–23. Oxford: Oxford University Press.
Martin Maiden
Aronoff, Mark. 1994. Morphology by itself: Stems and inflectional classes. Cambridge, MA:
MIT Press.
Burzio, Luigi. 2004. Paradigmatic and syntagmatic relations in Italian verbal inflection.
Julie Auger, Clancy Clements & Barbara Vance (eds.). 17–44.
Candrian, Johann Paul. 1900. Der Dialekt von Bivio-Stalla. Zürich: Halle.
Corbett, Greville. 2007. Canonical typology, suppletion, and possible words. Language
83(1). 8–42.
Cortelazzo, Manlio & Zolli Paolo. 1988. Dizionario etimologico della lingua italiana.
Bologna: Zanichelli.
Decurtins, Alexi. 1958. Zur Morphologie der unregelmässigen Verben im Bündnerromanischen. Bern: Francke.
Ebneter, Theodor. 1981. Vocabulari dil rumantsch da Vaz. Tübingen: Niemeyer.
Goebl, Hans, Hans Böhmer, S. Gislimberti, D. Kattenbusch, E. Perini, T. Szekely, I. Dautermann, S. Heissmann, U. Hofmann, A. Kozak, H. Pamminger, H.-M. Rössler, R. Bauer &
E. Haimerl. 1998. Atlant linguistich dl ladin dolomitich y di dialec vejins, 1a pert. Wiesbaden: Dr. L. Reichert Verlag.
Goebl, Hans, Ilaria Adami, Helga Böhmer, Axel Heinemann, Frank Jodl, Liza Klinger,
Daniele Rando, Brigitte Rührlinger, Walter Strauß, Tino Szekely, Paul Videsott, Heidemarie Beer, Gertraud Klingler, Agnes Staudinger, Edgar Haimerl, Bernhard Schauer,
Fabio Tosques & Andreas Wagner. 2012. Atlant linguistich dl ladin dolomitich y di dialec vejins, 2a pert. Strasbourg: Éditions de Linguistique et de Philologie.
Grisch, Mena. 1939. Die Mundart von Surmeir: Ober-und Unterhalbstein. Paris/Zurich:
Huonder, Josef. 1901. Der Vokalismus der Mundart von Disentis. Romanische Forschungen
11. 430–566.
Jaberg, Karl. 1939. Considérations sur quelques caractères généraux du romanche. In
Université de Genéve Faculté des lettres (ed.), Mélanges de linguistique offerts à Charles
Bally, 283–292. Geneva: Georg.
Jaberg, Karl & Jakob Jud. 1964. Sprach- und Sachatlas Italiens und der Südschweiz. Zofingen: Ringier.
Kaye, Steven. 2015. Conjugation class from Latin to Romance: Heteroclisis in diachrony and
synchrony. University of Oxford PhD thesis.
Kramer, Johannes. 1976. Historische Grammatik des Dolomitenladinischen. Formenlehre.
Gebrunn bei Würzburg: Lehmann.
Maiden, Martin. 1992. Irregularity as a determinant of morphological change. Journal of
Linguistics 28(2). 285–312.
Maiden, Martin. 2003. Verb augments and meaninglessness in Romance morphology.
Studi di grammatica italiana 22. 1–61.
Maiden, Martin. 2004. When lexemes become allomorphs: On the genesis of suppletion.
Folia Linguistica 38. 227–256.
Maiden, Martin. 2005. Morphological autonomy and diachrony. In Geert Booij & Jaap
van Marle (eds.), Yearbook of morphology 2004, 137–175. Dordrecht: Springer.
10 Romansh allomorphy (Again!)
Maiden, Martin. 2006. Accommodating synonymy: How some Italo-Romance verbs react
to lexical and morphological borrowings. In Arturo Tosi & Anna Laura Lepschy (eds.),
Rethinking languages in contact: The case of Italian, 87–98. Oxford: Legenda.
Maiden, Martin. 2009. From pure phonology to pure morphology: The reshaping of the
Romance verb. Recherches delinguistique de Vincennes 38. 45–82.
Maiden, Martin. 2011a. Instabilità e variazione dialettale nella storia di un verbo trascurato: Dovere nelle varietà italoromanze e “retoromanze”. In Marcato Gianna (ed.), Le
nuove forme del dialetto, 97–104. Padua: Unipress.
Maiden, Martin. 2011b. Morphomes and ‘stress-conditioned allomorphy’ in Romansh. In
Maria Goldbach, Marc-Olivier Hinzelin, Martin Maiden & John Charles Smith (eds.),
Morphological autonomy: Perspectives from Romance inflectional morphology, 36–50.
Oxford: Oxford University Press.
Maiden, Martin. 2011c. Morphophonological innovation. In John Charles Smith Anna
Laura Ledgeway Martin Maiden (ed.), The Cambridge history of the Romance Languages,
216–267. Cambridge: Cambridge University Press.
Maiden, Martin. 2012. A paradox? The morphological history of the Romance present
subjunctive. In Sascha Gaglia & Marc-Olivier Hinzelin (eds.), Inflection and word formation in Romance languages, 27–54. Amsterdam.
Maiden, Martin. 2013. Semi-autonomous morphology: A problem in the history of the
Italian (and Romanian) verb. In Silvio Cruschina, Martin Maiden & John Charles
Smith (eds.), The boundaries of pure morphology, 216–267. Cambridge: Oxford University Press.
Maiden, Martin. 2016. Inflectional morphology. In Adam Ledgeway & Martin Maiden
(eds.), The Oxford guide to the Romance languages, 497–512. Oxford: Oxford University
Maiden, Martin, John Smith, Maria Goldbach & Marc-Olivier Hinzelin (eds.). 2011. Oxford: Oxford University Press.
Pult, Caspar. 1897. Le parler de sent. University of Lausanne PhD thesis.
Rogers, Kenneth. 1972. Vocalic alternation in the Sursilvan Romansh verb. In Fisher John
& Gaeng Paul (eds.), Studies in honor of Mario A. Pei, 173–181. Chapel Hill: University
of North Carolina Press.
Schorta, Andrea. 1938. Lautlehre der Mundart von Müstair. Paris: Droz.
Signorell, Faust. 2001. Vocabulari surmiran-tudestg. Chur: Departamaint d’educaziun digl
Signorell, Faust, Mena Wuethrich-Grisch & Gion Pol Simeon. 1987. Normas surmiranas.
Chur: Tgesa editoura cantunala per stampats e meds d’instrucziun.
Sims, Andrea. 2015. Inflectional defectiveness. Cambridge: Cambridge University Press.
Solèr, Clau. 1991. Romanisch im Schams. Züurich: Phonogrammarchiv der Universität
Stürzinger, Jakob. 1879. Über die Conjugation im Rätoromanischen. Zurich: Winterthur,
Buchdr. von Bleuler-Hausheer & cie.
Tagliavini, Carlo. 1926. Il dialetto del Comelico. Archivum Romanicum 10. 1–200.
Martin Maiden
Thornton, Anna. 2004. Conversione V > N. In Maria Grossmann & Franz Rainer (eds.),
La formazione delle parole in italiano, 517–525. Tübingen: Niemeyer.
Thornton, Anna. 2011. Overabundance (multiple cells realizing the same cell): A noncanonical phenomenon in Italian verb morphology. In Martin Maiden, John Smith,
Maria Goldbach & Marc-Olivier Hinzelin (eds.), Morphological autonomy: Perspectives
from Romance inflectional morphology, 358–381. Oxford: Oxford University Press.
Wolf, Matthew. 2013. Candidate chains, unfaithful spell-out, and outwards-looking
phonologically-conditioned allomorphy. Morphology 23(2). 145–178.
Chapter 11
How to wake up irregular (and
Charles Yang
University of Pennsylvania
I suggest that morphological defectiveness arises when the learner fails to discover a productive/default process in a morphological category. The detection of productivity, or lack
thereof, can be accomplished by the Tolerance Principle, a simple mathematical model of
language learning and generalization. In this paper, I show that the absence of *amn’t, the
negative contracted form of am, in most English dialects can be predicted on purely numerical basis. Implications for language acquisition, variation, and change are also discussed.
1 From Irregular Verbs to Productivity
In my first linguistics talk, which was also my job interview at Yale, I proposed that
English irregular past tense is not learned by forming associations between the stem and
the inflected form, contrary to the dominant view in the psychological study of language
(Rumelhart & McClelland 1986; Pinker 1999). Rather, irregular past tense is generated
by morpholexical rules. These rules do not generalize beyond a fixed list but are rules
nevertheless, in the sense that they take the stem (e.g., think) as the input and generate
an output (e.g., thought), the inflection, via a computational process of structural change
(e.g., “Rime → /ɔt/”). I was approaching the problem as a computer scientist: rules
are most naturally realized as a list of if-then statements, for regulars and irregulars
alike, which turns out to be the approach taken throughout the history of linguistics
(Bloch 1947; Chomsky & Halle 1968; Halle & Marantz 1993) including Steve’s own work
(1973; 1992). There is in fact developmental evidence for the rule-based approach when
I reanalyzed the past tense acquisition data purportedly confirming the associationist
account (Yang 2002b).
I supposed Steve was at least somewhat persuaded by the argument; a few months
later I got the job. But he did wonder aloud after the talk, with a quizzical frown-cumsmile that only he can manage: “But how does a rule wake up in the morning and decide
to be irregular?”
Charles Yang. 2017. How to wake up irregular (and speechless). In Claire Bowern,
Laurence Horn & Raffaella Zanuttini (eds.), On looking into words (and beyond), 211–
233. Berlin: Language Science Press. DOI:10.5281/zenodo.495447
Charles Yang
Indeed. Since words do not wear tags of (ir)regularity, any morphological theory that
recognizes regularity and irregularity, which is pretty much everything on the market,
must say something about how a rule or process wakes up to be irregular. In fact, theories that reject such a categorical distinction (e.g., Hay & Baayen 2003; McClelland
& Patterson 2002) ought to be off market. Children’s morphological productivity is
strongly discrete; see Yang 2016: Chapter 2 for a cross-linguistic review. Their errors
are almost exclusively over-regularizations of productive rules. This is quite well known
thanks to the past tense debate: for example, the past tense of hold sometimes surfaces
as holded, with the “-d” rule (Marcus et al. 1992). What is not widely known and even
less appreciated is the near total absence of over-irregularization errors, despite frequent
anecdotes to the contrary (e.g., bite-bote, wipe-wope, think-thunk, etc.; Bowerman 1982;
Bybee 1985; Pinker 1999). These errors are sufficiently rare, occurring in about 0.2% of
English-learning children’s past tense use, that Xu & Pinker (1995) dub them “weird past
tense errors”. Not a single instance of bote, wope, thunk, or many conceivable analogical patterns can be found in the millions of child English words in the public domain
(MacWhinney 2000). The distinction between regular and irregular rules was in fact
observed in Berko’s (1958) original Wug test. While children were quite happy to add
“-d” to novel verbs such as rick and spow, only one out of eighty six subjects irregularized bing and gling, although adults are often prone to form irregular analogies in an
experimental setting.1
So Steve’s question sent me on a long quest. To maintain that both regulars and irregulars are computed by rules, I needed a story of how children separate out productive
and unproductive rules so precisely and effortlessly. Although a solution was worked
out shortly after (Yang 2002a), it took me many years to fully recognize the scope of the
productivity problem – one of the “central mysteries” in morphology (Aronoff 1976: 35)
– and the challenges it poses.
At a first glance, it doesn’t seem difficult to give an answer for English past tense. The
rule “add -d” covers most verb types in the language and can thus be deemed regular, as
“statistical predominance” has always been regarded as the hallmark of the default (e.g.,
Nida 1949: 14). But this is surely too simplistic when crosslinguistic and psychological
factors are taken into account. More concretely, at least four empirical problems, each of
which is illustrated with a familiar example in (1), fall under the productivity problem.
a. English past tense: That a default rule is learned abruptly and results in overregularization, after a protracted stage of rote memorization (Marcus et al.
1992; Yang 2002b).
b. English stress: That the grammar of English stress (Chomsky & Halle 1968;
Hayes 1982; Halle & Vergnaud 1987) is not trochaic with a list of lexical exceptions despite an overwhelming majority of English words bearing stress on
the first syllable (Cutler & Carter 1987; Legate & Yang 2013).
This suggests that the Wug test and similar methods such as rating have task-specific complications and
should not be taken as a direct reflection of an individual’s morphological knowledge; see Schütze 2005
and Yang 2016 for discussion.
11 How to wake up irregular (and speechless)
c. German noun plurals: That a suffix (“-s”) can be the productive default despite
coverage of fewer nouns than any of its four competitors (Clahsen et al. 1992;
Wiese 1996).
d. Russian gaps: That morphological categories needn’t and sometimes do not
have a default, as illustrated by the missing inflections of certain Russian verbs
in the 1st person singular non-past (Halle 1973).
In Yang (2016), I propose a model of productivity, the Tolerance Principle, which provides a unified solution for the problems in (1), as well as similar problems that involve
inductive learning in phonology, syntax, and language change. In this paper, I revisit
Steve’s question which, in a significant way, drove this project forward. My focus is on
a topic that has featured prominently in Steve’s recent research: morphological gaps and
the nature of defectiveness in word formation (e.g., Anderson 2008; 2010b).
2 The Tolerance Principle
The development of the Tolerance Principle started as a purely formal conjecture: How
would one represent a rule (R) and the exceptions of that rule (e.g., a set of words w 1 , w 2 ,
…, w n )? If one is committed to a mechanistic account of the matter – like a computer
programmer, for instance – perhaps the only way to encode rules and exceptions is
through a set of conditional statements:
(2) If w = w 1 Then …
If w = w 2 Then …
If w = w e Then …
Apply R
This of course immediately recalls the Elsewhere Condition, ever present in linguistics
since Pāṇini (Anderson 1969; Aronoff 1976; Kiparsky 1973; Halle & Marantz 1993). In
particular, the data structure in (2) entails that in order for a (productive) rule to apply
to a word, the system must scan through a list to ensure that it is not one of the exceptions
(w 1 , w 2 , …, w e ).
There is something perverse about (2). For example, to produce walked, one must
scan through the irregular verbs to make sure that walk is not found on the list. But a
moment of reflection suggests that the Elsewhere Condition makes perfect sense. The
alternative to listing the irregulars would have to be listing the regulars. One can imagine assigning each regular verb a flag, which immediately triggers the application of
the “add -d” rule. But that would imply that the morphological status of every word
must be committed to special memory; the irregulars as well, since they are by definition unpredictable. Perhaps even more surprisingly, there is broad behavioral support
for the irregulars-first-regulars-later representation of rules; see Yang 2016: Chapter 3
for review. The psycholinguistic evidence comes from real-time processing of words
Charles Yang
and morphology. When irregulars and regulars are suitably matched for various factors (e.g., stem and surface frequency) that affect the speed of processing, irregulars are
recognized and produced significantly faster than regulars – which is consistent with
the algorithmic interpretation of the Elsewhere Condition as a computational model of
language processing.
From (2), then, we can develop an empirically motivated cost-benefit calculus for the
price of exceptions. Specifically, words that fall under a productive rule must “wait”
for the exceptions to be processed first: the more exceptions there are, the longer the
rule will have to wait. Under very general assumptions about word frequencies, we can
Tolerance Principle
Suppose a rule R is applicable to N items in a learner’s vocabulary, of which e are
exceptions that do not follow R. The sufficient and necessary condition for the
productivity of R is:
e ≤ θ N where θ N :=
ln N
The Tolerance Principle requires two input values, N and e, and returns the productivity
status of a rule. Its application requires a well-defined rule such that N and e can be
measured, by the child learner during language acquisition and by the researcher when
studying linguistic productivity. To learn the structural description of a rule, typically in
the form of “X −→ Y”, one will need to invoke inductive learning models such as those
studied in artificial intelligence, cognitive science, and indeed linguistics (e.g., Chomsky
1955). Almost all inductive models form generalizations over specific learning instances
and try to discover the shared characteristics of individual elements associated with a
shared pattern. For example, suppose two good baseball hitters can be described with
feature bundles [+red cap, +black shirt, +long socks] and [+red cap, +black shirt, +short
socks]. The rule “[+red cap, +black shirt] −→ good hitter” will follow, as the shared
features (cap, shirt) are retained and the conflicting feature (sock) is neutralized. Obviously, the application of inductive learning must encode the structural constraints on
the human language faculty and other cognitive systems implicated in language acquisition (Chomsky 1965). While it is clear that the properties of human language are far
from arbitrary, it remains an open question to what extent they reflect a unique system
of Universal Grammar (e.g., Merge; Berwick & Chomsky 2016) or general principles of
cognition and learning that show continuities with other domains and species; see Yang
2004; Chomsky 2005; Yang et al. 2017 for general discussions.
Table 1 provides some sample values of N and the associate threshold value θ N .
The apparently, and perhaps surprisingly, low threshold has interesting implications
for language acquisition. Most importantly, it suggests that all things being equal, smaller
vocabulary (smaller values of N ) can tolerate relatively more exceptions. That is, productive rules are more detectable when learners have less experience with a language,
especially when they have a small lexicon that only consists of relatively high frequency
words. This may explain children’s remarkably early command of the main ingredients
11 How to wake up irregular (and speechless)
Table 1: The tolerance threshold for rules of varying sizes
of language (Yang 2013), as well as the reason why maturational constraints may aid
rather than hamper language acquisition (Newport 1990); see Yang 2016: Chapter 7 for
extensive discussion.
The Tolerance Principle has proved highly effective. In Yang (2016), it was applied
almost 100 times, making accurate productivity predictions across many languages and
domains using only corpus statistics. Furthermore, experimental studies in collaboration with Kathryn Schuler and Elissa Newport have found near categorical confirmation
for the Tolerance Principle in artificial language studies with young children (Schuler,
Yang & Newport 2016). Some of these robust results are unexpected. This is because the
derivation in (3) makes use of numerical approximations that only hold when N is large.
In the empirical case studies, however, the value of N is often very modest (e.g., 8 or
9 in the artificial language studies) as it refers to the number distinct lexical items in a
morphological category. For the moment, I put these questions aside and return to the
problems in (1): the low threshold of exceptions provides just the right approach to the
productivity problem across languages.
Consider first the acquisition of English past tense. Through an inductive process illustrated earlier, the phonological diversity of the regulars will quickly establish that
any verb can take the “-d” suffix. Its productivity will be determined by the total number
of verbs (N ) and the irregulars (e) in the learner’s vocabulary. The same consideration
applies to the irregular rules. For instance, the seven irregular verbs bring, buy, catch,
fight, seek, teach, and think all follow the stem change “ought”. Such a mixed bag of
phonological shapes will also yield an all-inclusive rule, as shown by computational implementations (Yip & Sussman 1998). But the “ought” rule will fare terribly. It only works
for seven items, with hundreds and thousands of exceptions, far exceeding the tolerance
threshold. As a result, the rule will be relegated to lexicalization. Other irregular patterns can be analyzed similarly: as I show elsewhere (Yang 2016: Chapter 4), they all
wake up nonproductive in the morning, thereby accounting for the near total absence
of over-irregularization errors (Xu & Pinker 1995).
Charles Yang
Following the same logic, we can see that the emergence of the “-d” rule will require a
long period of gestation. Although children can quickly induce its structural description
– using no more than a few dozen verbs (again Yip & Sussman 1998) – their early verbs
will contain many irregulars. Of the top 200 verbs inflected in the past tense (MacWhinney 2000), 76 are irregulars. Because θ 200 is only 37, it follows that children who know
some 200 most frequent verbs cannot recognize the productivity of “-d” despite its “statistical predominance”. During this period of time, even though verbs may be produced
with the “-d” suffix, they are in effect irregular: the suffix has no productivity and does
not extend beyond a fixed list rote-learned from the input. The telltale evidence for productivity comes from the first attested overregularization errors (Marcus et al. 1992). For
individual learners with reasonably complete records of language development, the Tolerance Principle can help us understand why the regular rule becomes productive at
that exact moment it did. For example, “Adam”, the poster child of English past tense
research (Pinker 1999), produced his first over-regularization error at 2;11: “What dat
feeled like?” In the transcript of almost a year prior to that point, not a single irregular
verb past tense was used incorrectly. It must be that by 2;11, Adam had acquired a sufficiently large number of regulars to overwhelm the irregulars. To test this prediction, I
extracted every verb stem in Adam’s speech until 2;11. There are N = 300 verbs in all,
out of which e = 57 are irregulars. This is very close to the predicted θ 300 = 53, and
the small discrepancy may be due to the under-sampling of the regulars, which tend
to be less frequent and thus more likely missing from the corpus. The critical point to
note here is that Adam apparently needed a filibuster-proof majority of regular verbs
to acquire the “-d” rule: this is strongly consistent with the predictions of the Tolerance
Principle as illustrated in Table 1.
The problems of English stress and German plurals in (1) are similar. In the English
case, the assignment of stress to the first syllable may be transiently productive when
the child has a very small vocabulary (Kehoe & Stoel-Gammon 1997; Legate & Yang 2013).
But it will fail to clear the tolerance threshold when the vocabulary reaches a modest
size: even 85% of coverage is not sufficient for larger values of N (e.g., 5000; Table 1). In
the German case, none of the five plural suffixes can tolerate the other four as exceptions, not least the “-s” suffix, which covers the smallest set. In both cases, the learner
will carry out recursive applications of the Tolerance Principle. When no rule emerges
as productive over the totality of a lexical set, the learner will subdivide it along some
linguistic dimension, presumably making use of constraints on language and other cognitive systems, and attempt to discover productive rules within. Such a move, while more
complex, is always more likely to yield productive rules: again, smaller N ’s that result
from subdividing the lexicon tolerate a relatively higher proportion of exceptions than
larger N ’s. For the acquisition of stress, dividing words into nouns and verbs and taking
the syllabic weight into account, as prescribed by all modern metrical theories, lead to
productive rules of stress assignment, an outcome that accords well with both structural
and behavioral findings (Ladefoged & Fromkin 1968; Baker & Smith 1976; Kelly 1992;
Guion et al. 2003). The study by Legate & Yang (2013) also reveals important differences
between theories of stress in their statistical coverage of the English lexicon: while all
11 How to wake up irregular (and speechless)
theories handle a great majority of English words, only the theory of Halle 1998 clears
the tolerance threshold of exceptions. For the acquisition of German plurals, the move is
to subdivide the nouns by grammatical gender as well as phonological conditions, similar to certain theoretical approaches to German morphology (e.g., Wiese 1996). The -s
suffix indeed survives as the default because the other suffixes are productive with more
restrictive domains of nouns.
The emergence of morphological gaps is a logical outcome of the Tolerance Principle,
which constitutes the topic of the present study. When a rule wakes up irregular, the
learner must learn, from positive evidence, the inflected form for each word. Failing to
hear a particular inflected form will render the speaker speechless when that form is
3 Why Am+Not ≠ Amn’t?
3.1 Conditions on Gaps
Many current theories of morphology, including Distributed Morphology (for which
see Halle & Marantz 1993), Optimality Theory (Prince & Smolensky 2004), Dual-Route
Morphology (Pinker 1999; Clahsen 1999), Network Morphology (Brown & Hippisley
2012), Paradigm Function Morphology (Stump 2001) and others, invoke the notion of
competition, which by design results in a default or winning form (at least in the inflectional domain). This architectural feature of the theories is inherently incompatible
with the existence of morphological gaps, which are quite widespread across languages
(Baerman, Corbett & Brown 2010). The Tolerance based approach, while also competition based (through the Elsewhere Condition), does not stipulate a default or productive
rule as a primitive in the theoretical machinery. Rather, the presence or absence of a
productive rule is the outcome of language acquisition, to be determined by children
through the composition of the linguistic data. More specifically, the Tolerance Principle provides the following corollary (Yang 2016: 142):
(4) Conditions on gaps
Consider a morphological
∑ category C with S alternations, each affecting Ni lexical
items (1 ≤ i ≤ S), and i Ni = N . Gaps arise in C only if:
Nj > θN
∀i, 1 ≤ i ≤ S,
∑ is, none of the alternations (Si ) in N are sufficiently numerous to tolerate all the rest
( j,i N j ) as exceptions: no productive alternation will be identified. The speaker must
hear the morphological realization of every word in C; if any is to slip through the cracks,
a defective gap appears. I should note that in the conception and application of the Tolerance Principle, the terms such as “category” and “alternation” are meant to be general
and not restricted to morphology per se. For instance, “category” can be interpreted as
any well-defined structural class with a finite number of elements (phonemes, words,
Charles Yang
morphosyntactic structures, the directionality of a finite number of functional heads,
etc.), and “alternation” can be understood as any outcome of a computational process
defined over such a class. The Tolerance Principle asserts that in order for a productive
pattern to emerge, one of the alternations must be statistically dominant. Elsewhere I
have studied several well-known gaps in English, Polish, Spanish, and Russian (Yang
2016: Chapter 5). Their presence is predictable entirely on numerical ground, requiring
nothing more than tallying up the counts of the lexical items subject to each alternation. In what follows, I provide a Tolerance Principle account of another much-studied
instance of a defective paradigm.
3.2 The Statistics of N’t Gaps
In many dialects of English, n’t is not permitted to contract onto auxiliary verbs such
as am and may, as seen in the unavailability of, for example, “*I amn’t tired” and “*You
mayn’t do that” (e.g., Anderwald 2003a; Bresnan 2001; Broadbent 2009; Frampton 2001;
Hudson 2000; Zwicky & Pullum 1983). Following Zwicky & Pullum (1983), I will assume
that the contracted negative n’t is an inflectional affix. The question is why n’t cannot
attach to all auxiliary verbs residing in the Tense node. From the perspective of the
Tolerance Principle, the emergence of gaps must result from a critical mass of exceptions
to the contraction process.
Let us consider the behavior of the auxiliary hosts for n’t. Zwicky & Pullum (1983:
p508) provide a near comprehensive list, which I revise with some additional information
in Table 2.
Table 2 provides the frequencies of the auxiliary verbs and their negation in both
uncontracted and contracted forms in the 520-million-word Corpus of Contemporary
American English (COCA; Davies 2008). Given the heterogeneity of the textual sources,
a handful tokens of amn’t and mayn’t can be found albeit at very low frequencies. The
n’t-contracted forms of shall and dare – shan’t and daren’t – are also impossible for most
American English speakers but are included here for completeness. Although shan’t is
often perceived as a stereotypically British English feature, it seems to be vanishing
across the pond as well. In a 6.6-million-word corpus of British English (MacWhinney
2000), not a single instance of shan’t is found. And its frequency of usage has been in a
steady decline since 1800, the beginning date of the Google Books Corpus. As of daren’t,
the OED does not provide any citation and it has always been very rare throughout the
period of the Google Books Corpus. These gapped forms are marked by ∅.
The prescriptively maligned ain’t ([eɪnt]), however, is robustly attested for am, are, is,
have, and has in COCA as well as a six-million-word corpus of child-directed American
English (MacWhinney 2000). Since the phonological form of [eɪnt] is unpredictable from
the auxiliary host, it is boldfaced in Table 2 to mark its idiosyncrasy, along with a few
other exceptions to which I return later. Note that the frequency estimates of the ain’t
forms are approximate. First, I only counted strings where ain’t is immediately preceded
by a pronoun – the majority case, but sentences with a lexical subject (e.g., “Kids ain’t
ready”) are not included. Second, because both be and have can take on ain’t, the counts
11 How to wake up irregular (and speechless)
Table 2: The morphophonological alternation of n’t contraction
could [kud]
did [dɪd]
does [dʌz]
had [hæd]
has [hæz]
has [hæz]
have [hæv]
have [hæv]
is [ɪz]
is [ɪz]
might [maɪt]
must [mʌst]
need [nid]
ought [ɔt]
should [ʃud]
was [wʌz]
would [wud]
am [æm]
am [æm]
are [ar]
are [ar]
can [kæn]
dare [dɛər]
do [du]
may [meɪ]
shall [ʃæl]
were [wr̩]
will [wɪl]
n’t contraction
for the auxiliaries are parceled out by extrapolating from the frequencies of the regularly
contracted n’t forms.2 For instance, there are 2,054 instances of “you/they ain’t”: the
“share” for are is based on the count of “aren’t” (50,137) relative to “haven’t” (45,849).
Here I gloss over the fact that there are English dialects in which ain’t is also an alternative form of negative
contraction for do, does, and did (e.g., Labov et al. 1968; Weldon 1994). It would be difficult to estimate their
frequencies but formally, this use of ain’t serves to create additional (unpredictable) exceptions to the
contraction process which, as we discuss below, contributes to the breakdown of productivity and the
emergence of gaps.
Charles Yang
This amounts to 52.2% of 2,054, or 1,073, as recorded in the Table. Finally, the estimate of
ain’t as the contraction of am + n’t cannot follow a similar process because, of course,
amn’t is gapped. I thus allocated roughly 75% of the “I ain’t” counts, which is the share
of “I am not” out of the total of “I am not” and “I have not”, to the contraction of am not.
For these five auxiliaries that can be realized as ain’t, the percentage of the contracted
forms are based on the sum of uncontracted, n’t-contracted, and ain’t-contracted forms.
More precise estimates are certainly possible but as we will see, the exact frequencies
are not especially important for our purposes: it is more pertinent to approximate a
“typical” English speaker’s experience with these forms. Roughly, we would like to know
whether an English speaker will have encountered a specific phonological word at all, by
using some independently motivated frequency threshold (e.g., once per million; Nagy
& Anderson 1984): it is evident that the frequency of ain’t is sufficiently high for this
threshold despite our rough estimates.
A tempting approach to gaps is to appeal to indirect negative evidence (Chomsky
1981; Pinker 1989). A strong version takes the shape of lexical conservatism: do not use a
form unless it is explicitly attested. This recalls Halle’s [-Lexical Insertion] treatment of
gaps in his classic paper (1973) and can be found in recent works as well (e.g., Pertsova
2005; Steriade 1997; Rice 2005; Wolf & McCarthy 2009). A weak version makes use of
frequency information. For instance, if amn’t were possible, language learners would
have surely heard it in the input, especially since am is highly frequent and would have
had plenty of opportunities to undergo n’t contraction. Its conspicuous absence, then,
would provide evidence for its ungrammaticality (e.g., Daland, Sims & Pierrehumbert
2007; Sims 2006; Baerman 2008; Albright 2009).
Traditional acquisition research has always viewed indirect negative evidence with
strong suspicion (Berwick 1985; Osherson, Stob & Weinstein 1986; Pinker 1989). Research
on the amn’t gap (e.g. Hudson 2000) has also questioned its usefulness. However, with
the recent rise of probabilistic approaches to language acquisition especially Bayesian
models of inference, the field has seen a revival of indirect negative evidence. If the
conception of learning is a zero-sum – or more precisely, one-sum — game which assigns
a probabilistic distribution over all linguistic forms, the unattested will necessarily lose
out to the attested, at least in most probabilistic models of language learning. A thorough
assessment of indirect negative evidence within a probabilistic framework is beyond the
scope of the present paper; see Niyogi 2006; Yang 2015; Yang et al. 2017. But a careful
statistical examination of gaps serves to reveal its deficiencies. Note that the question is
not whether indirect negative evidence can account for some missing forms: the absence
of amn’t is indeed unexpected under any reasonable formulation. The real challenge is to
ensure that indirect negative evidence will pick out only the gapped forms but nothing
else, while keeping in mind that morphological inflection is generally not gapped but
fully productive, readily extending to novel items.
Two observations can be made about the frequency statistics in Table 2, which suggest that indirect negative evidence is unlikely to succeed. First the n’t forms of several
auxiliaries such as might and need are in fact quite rare. They appear considerably less
frequently than once per million, which is generally regarded as the minimum threshold
11 How to wake up irregular (and speechless)
to guarantee exposure for most English speakers (Nagy & Anderson 1984). In the sixmillion-word corpus of child-directed American English (MacWhinney 2000), mightn’t
appears only once, needn’t appears only twice, and mustn’t does not appear at all. In
the other words, these n’t forms may be so rare that they are in effect absent for many
children (Hart & Risley 1995). Lexical conservatism thus will not distinguish them from
the truly gapped amn’t, mayn’t, daren’t, and shan’t, the last of which is in fact more frequently attested in COCA. Second, consider a statistical interpretation of indirect negative evidence. The last column of Table 2 provides the percentage of the n’t contraction
out of all negated forms. An auxiliary with an unusually low ratio may mean that it has
performed below expectation and could be a clue for its defectiveness. However, the
statistics in Table 1 suggest otherwise. It is true that amn’t and mayn’t have very low
ratios: this fact alone is not remarkable because these are indeed gaps. But exactly how
low should a ratio be for the learner to regard a contracted form to be defective? On the
one hand, we have mightn’t and oughtn’t at 0.525% and 6.016%, and these are not defective. On the other hand, we have daren’t and shan’t at 7.246% and 8.824%, but these in
fact are defective. There doesn’t appear to be a threshold of frequency or probability that
can unambiguously distinguish gapped from ungapped items.
3.3 N’t Contraction in Language Development and Change
Let’s see how the Tolerance Principle provides an account of the amn’t gaps. The simplest
approach is to consider all the auxiliary verbs and their n’t contractions collectively as a
homogeneous set. Using the once-per-million threshold as a reasonable approximation
of a typical American English speaker’s vocabulary, and taking the size of the Corpus
of Contemporary American English (520 million words) into account, there are 18 auxiliaries with reliably attested n’t forms. The four gapped forms are all below this threshold
and are thus excluded from consideration. It is important to clarify that, unlike various
forms of lexical conservatism and indirect negative evidence discussed earlier, we do
not regard the absence of these forms as evidence for their defectiveness. Rather, the
learner’s task is to deduce, on the basis of the 18 well-attested forms, including am∼ain’t,
that n’t contraction is not a productive pattern in the English auxiliary system.
This is quite easily accomplished. Of the 18 auxiliaries, n’t is realized as follows:
[n̩t]: could, did, does, had, need, should, was, would (8)
[eɪnt] in variation with either [nt] or [n̩t]: have, has, is, are (4)
[nt]: can, were (2)
idiosyncratic vowel change: do, will (2)
[eɪnt]: am (1)
[n̩t] but idiosyncratically deletes [t] in the auxiliary (see Zwicky & Pullum
1983: 508–509 for discussion): must (1)
For any of these alternations to be productive, it must have no more than θ 18 = 6 exceptions. The most promising [n̩t], which applies to 8 auxiliaries and thus has 10 exceptions,
Charles Yang
is a long way off. Even if we are to include the [n̩t]-taking have, has, and is and ignore
the unpredictable variant [eɪnt] form, the rule “nt −→ n̩t” still falls short of productivity.
Thus, the learner will be able to conclude, from the Conditions on Gaps (4), that n’t contraction is not a productive process for English auxiliaries and must be learned lexically.
If amn’t fails to appear in the input, it will be absent. Only after the learner has already
concluded that a category does not have a productive rule can they start to regard the
absence of evidence as evidence of absence.
The preceding analysis, while correctly identifies the n’t gaps, has some inadequacies.
For one thing, based on the 18 contracted forms, the primary evidence for language acquisition, learners would also identify mustn’t and oughtn’t as gapped as they fall below
the minimum frequency of once per million. This is not necessarily a fatal shortcoming:
mustn’t and oughtn’t are still considerably more frequent than amn’t and mayn’t, the two
genuinely gapped forms, and children may acquire them in later stages of acquisition.
But more significantly, as Steve pointed out to me in a personal communication (unrelated to the current celebratory volume), the preceding brute-force approach misses an
important structural generalization. Table 2 is divided into two halves on Steve’s advice.
As he insightfully observes, none of the auxiliaries that ends in an obstruent is gapped;
these are listed in the top portion of the Table. By contrast, gaps are only found in the
auxiliaries that do not end in an obstruent, which are listed in the bottom portion of the
If we carry out a Tolerance analysis along the feature [±obstruent], a much more
elegant and interesting pattern emerges. For the 12 [+obstruent] auxiliaries, only four
have exceptions – has, have, is, and must – just below θ 12 = 4. Thus, English learners
can identify a productive rule:
nt −→ n̩t / [+obstruent] #
This immediately accounts for the fact that speakers generally accept the forms mightn’t
and oughtn’t despite their very low frequencies (well below once per million): these
two auxiliaries, of course, follow the structural description of (6). By contrast, amn’t,
mayn’t, daren’t, and shan’t, some of which appear more frequently than mightn’t and
oughtn’t, are generally rejected because they fail to meet the structural descriptions of
the productive rule in (6).
Consider now the six [-obstruent] auxiliaries in the bottom portion of Table 2. Here
am and are have [eɪnt], can and were add [nt], and do and will have idiosyncratic vowel
changes. Since the Tolerance threshold θ 6 = 3, no distinct pattern will be identified as
productive: lexicalization is required and gaps are predicted for mayn’t, daren’t, shalln’t,
and of course amn’t.
The calculation here is very delicate but it is interesting to push the Tolerance Principle
to the limit. What if the child has not learned ain’t as the n’t-contracted form for am
and are? Although ain’t forms are quite robustly attested in COCA as well as in childdirected English, they are still strongly dialectal and are, at least in the input to some
children, less frequent than the “regular” forms such as aren’t, isn’t, haven’t, and hasn’t.
If so, a child during an early stage of acquisition may in effect have only five [-obstruent]
11 How to wake up irregular (and speechless)
auxiliaries and their contracted forms to learn from: namely, are, can, do, were, and will.
Here the statistically dominant pattern of “nt −→ [nt] / [-obstruent] #
” does reach
productivity: the two idiosyncratic exceptions of do and will fall below the threshold of
θ 5 = 3, and n’t contraction is predicted to be transiently productive!
Bill Labov (personal communication) distinctly recalls being a young amn’t speaker
only to exit that stage at a later time. Indeed, we can find attested examples in American English-learning children’s speech. The three examples in (7) are taken from the
CHILDES database (MacWhinney 2000):
a. I amn’t a dad. (Kate/Kim, 3;6: Sawyer Corpus 3-12-92.cha)
b. I’m doing this puzzle well, amn’t I? (Mark, 3;11: MacWhinney Corpus 67b1.cha)
c. Amn’t I clever? (Mark, 3;11: MacWhinney Corpus 67b2.cha)
The reader is encouraged to listen the audio recordings of the examples in (7) in the
CHILDES database. The first child’s identity is unclear due to discrepancies in transcription. The examples from Mark can be heard as the investigator’s exact revoicing (Brian
MacWhinney, personal communication). Although three examples seem quite rare, it is
worth noting that almost all am’s are contracted onto the pronoun (i.e., I’m not). Of the
one million American English child utterances, there are only 42 full forms of am followed by negation (i.e., I am not), which makes the three amn’t errors not so negligible.
Of course, everyone eventually hears ‘I ain’t’: from pop songs on radio if not from the immediate family and friends. Thus, amn’t will disappear according to the Tolerance-based
analysis, for ain’t introduces an additional exception which leads to the breakdown of
productivity for the [-obstruent] class.
Corroborative evidence for the (transient) productivity of n’t contraction can also be
found in other auxiliaries. To my great surprise, there are numerous instances of willn’t
as the negative contracted form of will and whyn’t for ‘why don’t/didn’t’ in the speech
of many parent-child dyads, apparently all from the New England region. Other than
enriching the empirical data on contraction, willn’t and whyn’t do not tell us much about
the productivity of n’t contraction or its acquisition: if parents use them frequently, and
they do, children will follow. Nevertheless, willn’t can also be found in the spontaneous
speech of children who are not from the New England region:3
a. No we willn’t. (Ross 2;9, Colorado, MacWhinney Corpus 26b2.cha)
b. Oh it willn’t fit in there (Marie 6;6, Ontario, Evans Corpus dyad07.cha)
c. He willn’t be a good boy (Jared 6;7, Ontario, Evans Corpus dyad19.cha)
Perhaps most strikingly is an utterance produced by Sarah, a child from the Harvard
studies (Brown 1973):4
(9) And the reindeer saidn’t.
Brain MacWhinney (personal communication) confirmed that the only time he or his wife ever used willn’t
was when transcribing Ross’s speech.
4 The contraction of n’t onto the main verb as in (9) was attested in the history of English: see Brainerd 1989
for caren’t (‘don’t care’) and Jespersen 1917 for bettern’t, usen’t, and indeed whyn’t.
Charles Yang
Taken together, the examples in (7), (8), and (9) suggest that n’t contraction is at least
transiently productive for some English-learning children.
Ross’s willn’t presents an especially interesting opportunity for studying the productivity of n’t contraction. The CHILDES corpus contains a relatively extensive record
of Ross’s longitudinal language development. We can then study his auxiliaries and
contractions, and subject his individual grammar to the kind of fine-grained analysis
of Adam’s past tense (§2). By the time Ross produced No we willn’t, he had used 9 n’tcontracted auxiliaries:
a. couldn’t, didn’t, haven’t, isn’t, wouldn’t
b. aren’t, can’t, don’t, won’t
If Ross had not started partitioning the auxiliaries by the [±obstruent] feature, the
N = 9 examples in (10) supports the productive use of n’t contraction because the four
examples in (10b) are below the number of tolerable exceptions (θ 9 = 4.2). The 5/4 split
between rule-governed and exceptional items is exactly the stimuli used in the artificial
language study (Schuler, Yang & Newport 2016) where children nearly categorically generalized the rule. If he failed to distinguish the syllabic [n̩] in (10a) and the nonsyllabic
[n] in aren’t and can’t in (10b), it would have been even easier for n’t contraction to
reach productivity. Thus, Ross’s productive use of n’t contraction in (8) is predicted by
the Tolerance Principle.
The naturalistic evidence from child language is admittedly thin, but it suggests that
the emergence of the amn’t and other gaps in the auxiliary system may be due to the
use of ain’t. Again, the gaps would not be the result of mutual exclusivity: there are
doublets such as haven’t∼ain’t etc. so amn’t and ain’t could have coexisted side by
side. Gaps arise/arose because the form of ain’t weakens the numerical advantage of n’t
contraction, pushing it below the Tolerance threshold.
Finally, a little historical detective work bolsters our treatment of the amn’t gap.5 According to Jespersen (1917: 117), “the contracted forms seem to have come into use in
speech, though not yet in writing, about the year 1600.” The change appears to have
originated in non-standard speech before spreading to mainstream usage. Subsequent
scholarship, however, places the date to a somewhat later time (e.g., 1630, Brainerd 1989:
181; see also Warner 1993: 208–209). Pursuing the results from the Tolerance-based analysis, we can make two observations.
First, it is likely that n’t-contraction was at one point productive, which seems especially effective for the [+obstruent] auxiliaries; see also (9) and fn 4. Brainerd’s study
finds that didn’t, hadn’t, shouldn’t, and wouldn’t appeared from 1670s, soon after the n’t
contraction appeared in the English language. These were followed by couldn’t, mightn’t,
needn’t, and mustn’t in the 18th century, and the last to join the group was oughtn’t in
the 19th century, first attested in Dicken’s 1836 The Village Coquette. Thus speakers at
that time must have formed a productive contraction rule for [+obstruent] auxiliaries,
perhaps like the one given in (6). Following this line of reasoning, we make the pre5
I am grateful to Anthony Warner for pointing out the important study of Brainerd 1989.
11 How to wake up irregular (and speechless)
diction, admittedly one that is difficult to test, that if a new [+obstruent] auxiliary is to
appear in the language, it will be immediately eligible for n’t contraction.
Second, and in contrast to the [+obstruent] class that had been expanding the number of n’t contractible auxiliaries, the [-obstruent] class has been steadily losing members. Interestingly, the [-obstruent] auxiliaries were quite systematically available for
n’t contraction by the end of the 17th century (Brainerd 1989). Of special interest are of
course those that were n’t contracted in the past but are presently gapped. According
to Brainerd’s study, the first instance of shan’t appeared in 1631, mayn’t in 1674, daren’t
in 1701: all three are now gapped. The very fact that they fall out of usage points to
the non-productivity of n’t contraction for these [-obstruent] auxiliaries: in general, a
productive rule would have gained rather than lost members.
How, we wonder, did the [-obstruent] class lose its productivity? Much more detailed
historical investigation will be needed but an interesting hypothesis can be offered as
follows. The historical development of n’t contraction may mirror the trajectory of language acquisition by children; that is, ontogeny may recapitulate phylogeny. Our discussion of children’s n’t contraction in modern American English suggests that the use
of ain’t for am not, which children probably acquire later during acquisition, increases
the number of exceptions for the contraction process. It is conceivable that the emergence of ain’t, an unpredictably contracted form of am not, was also the culprit for the
breakdown of productivity.
Historically, an’t/a’nt surfaced as the contracted form of am not between 1673 and
1690. But by the early 1700s, an’t/a’nt began to be used for both am not and are not (Brainerd 1989: 186). Whatever the phonological cause for this convergence, or how/when ain’t
joined the fray, the effect is that am not no longer had a predictable form of contraction.
If our analysis of children’s amn’t and willn’t is correct, then we would find amn’t and
ain’t to be in complementary distribution: If a dialect does not allow ain’t for am not,
amn’t would be possible; otherwise amn’t would be gapped.
The most direct evidence for this suggestion comes from the dialectal distribution of
amn’t, and its correlation with ain’t. The OED notes that amn’t is present in “nonstandard” American English and various northern parts of the UK. There is little to suggest that amn’t is possible in American English at all; all the five occurrences in COCA
come from Scottish and Irish writers.6 It is remarkable, then, that Scotland and Ireland
have “traditionally completely ain’t-free dialects” (Anderwald 2003b: 520): it is precisely
in these regions where amn’t is robustly attested, both in the century-old The English
Dialect Dictionary (Wright 1898) and in recent dialect surveys of English (Anderwald
Before I conclude this section, it is important to clarify the scope of the present analysis. The Tolerance Principle, through Conditions on Gaps (4), can identify defective
morphological category where gaps may emerge. Such categories are defined by the
The corpus of child-directed American-English, surprisingly, contains one instance of amn’t: “I am stirring,
amn’t I?” It was produced by Colin Fraser, on staff in Roger Brown’s Harvard study of language acquisition
(1973). A little Internet research reveals that Fraser, later a Cambridge scholar with a few psychology
textbooks to his credit, is a native of Aberdeen.
7 I would like to thank Gary Thoms for discussion for the distribution of amn’t in Scottish English.
Charles Yang
structural descriptions of rules. It does not predict, at least synchronically, which items
within these categories will be gapped. That issue, in my view, is completely a matter
of usage frequency: if the inflected form of an item in a defective category is used very
rarely or not at all, it will be gapped. Of course, it is also possible that no gaps are found
in a defective morphological category, if all items happen to be inflected sufficiently frequently. In that case, however, we do predict that if a novel item matches the structural
description of a defective category, the speaker will be at a loss to produce an inflected
form. Thus, the emergence of gaps, just as the calibration of productivity, is determined
by the composition of the input data. Finally, the preliminary work on the history of n’t
contraction suggests that the Tolerance Principle can be applied to the study of language
change. It makes concrete predictions about productivity – the rules that could gain new
members, and the rules that could only lose existing members – as long as the relevant
values of N and e from historical data can be reliably estimated. The reader is referred
to Yang 2016 for a case study of the so-called dative sickness in Icelandic morphosyntax.
4 Gaps in I-language
Halle’s classic paper (1973) contains the much criticized proposal that gaps are caused by
the [+Lexical Insertion] feature associated with certain forms. As noted earlier, this kind
of lexical conservatism is difficult to reconcile with the unbounded generativity of word
formation, and similar approaches using indirect negative evidence are also unlikely to
succeed. But in a footnote of that very paper, Halle proposes an alternative approach
which he himself regards as equivalent but has almost never been discussed by other
The proposal just sketched might be modified somewhat as regards the treatment
of words formed by rules that traditionally have been called “nonproductive”. One
might propose that all words formed by non-productive rules are marked by these
rules as [-Lexical Insertion]. The smaller subset of actually occurring words formed
by such rules would then be listed in the filter with the feature [+Lexical Insertion].
… In other words, it is assumed that words generated by a productive process are
all actually occurring and that only exceptionally may a word of this type be ruled
out of the language. On the other hand, words generated by a nonproductive rule
are assumed not to be occurring except under special circumstances. In this fashion
we might capture the difference between productive and nonproductive formations
Hetzron (1975), while arguing against Halle’s [+Lexical Insertion] proposal, makes
essentially the same suggestion. Rules are either productive or lexicalized, and gaps
arise in the unproductive corners of the grammar. His conception of gaps can be strongly
identified with the Elsewhere Condition, a critical component of the present theory:
The speaker must use ready-made material only for “exceptional” forms, while everywhere else he could very well “invoke the word formation component”. Technically, this can be represented by a disjunctive set of rules where idiosyncratic or
11 How to wake up irregular (and speechless)
“exceptional” formations are listed with as much explicitness as necessary, while
the general word formation rules would appear afterward, with the power to apply
“to the rest” (871).
That is, gaps arise when productivity fails. The problem of gaps thus reduces to the
problem of productivity. Some subsequent proposals have adopted a similar approach
(Albright 2009; Baronian 2005; Hudson 2000; Maiden & O’Neill 2010; Pullum & Wilson 1977; Sims 2006), including Steve’s own account (2010): gaps result from conflicting
forces in word formation such that the output form becomes unpredictable and thus
unrealized. The Tolerance Principle provides a precise solution of what makes a rule
productive, and its application to gaps reinforces the general position that gaps and productivity are two sides of the same coin.
The Tolerance Principle is a provable consequence of the Elsewhere Condition and
follows from the general principle of efficient computation: the child prefers faster grammars, a “third factor” in language design par excellence (Chomsky 2005). In fact, a
stronger claim can be made in favor of such an analytical approach. I submit that a
descriptive analysis of languages, however typologically complete or methodologically
sophisticated, cannot in principle provide the right solution for productivity. First, as
noted earlier, the categorical nature of children’s morphological acquisition suggests
that productivity must be demarcated by a discrete threshold (see also Aronoff 1976: 36).
But note that such a threshold is empirically undiscoverable. Productive processes will
lie above the threshold and unproductive processes will lie below, but with arbitrary
“distance” from it in both cases. Thus, the threshold cannot be regressed out of the data.
Second, while linguists now have an ever expanding arsenal of investigative tools to
study productivity, ranging from the Wug test to fMRI to Big Data, the psychological
grammar is developed without supervision in a matter of few years; these new empirical methods presently are at best a description of the speaker’s grammatical knowledge
and not yet learning models that account for how such knowledge is acquired. Finally,
even if we were to discover the threshold of productivity through a statistical analysis –
e.g., a productive rule must hold for at least 85% of eligible words – it would still remain
mysterious why the critical value is exactly what it is, rather than 80% or 90%.
In other words, an I-language approach to productivity is needed, one which builds
exclusively on the inherent constraints on language and cognition that all children have
access to, with deductively established properties that must hold universally across languages. The study of language as a part of human biology, I believe, is an approach
that Steve endorses and pursues (Anderson & Lightfoot 2002), which can be seen in his
writings on morphology and related issues (Anderson 2010a; 2015).
Finally, a personal note. It is no exaggeration to say that I owe my professional career
to Steve. He managed to create a position for me at Yale, which kept me close to my
young family and thus linguistics, and further away from the seductive fortunes in the
tech sector. It was also Steve who taught me, more effectively than anyone, the difference
between linguistic evidence and rhetoric. It has been a privilege to learn from him. To
figure out how to wake up irregular took over 15 years; the answer, I hope, is to his
satisfaction. It may once again win me a spot, this time in the Linguistic Club of Ashville,
North Carolina.
Charles Yang
I would like to thank Steve Anderson, Bill Labov, Brian MacWhinney, Gary Thoms, Anthony Warner, and an anonymous reviewer for helpful discussions and comments.
Albright, Adam. 2009. Lexical and morphological conditioning of paradigm gaps. In Curt
Rice & Sylvia Blaho (eds.), Modeling ungrammaticality in Optimality Theory, 117–164.
London: Equinox.
Anderson, Stephen R. 1969. West Scandinavian vowel systems and the ordering of phonological rules. Cambridge, MA: Massachusetts Institute of Technology PhD thesis.
Anderson, Stephen R. 1973. Remarks on the phonology of English inflection. Language
and Literature 1(4). 33–52.
Anderson, Stephen R. 1992. A-morphous morphology. Cambridge: Cambridge University
Anderson, Stephen R. 2008. Phonologically conditioned allomorphy in the morphology
of Surmiran (Rumantsch). Word Structure 1(2). 109–134.
Anderson, Stephen R. 2010a. An I-language view of morphological ‘exceptionality’: Comments on Corbett’s paper. In Horst J. Simon & Heike Wiese (eds.), Expecting the unexpected: Exceptions in grammar, 127–134. Berlin: Mouton de Gruyter.
Anderson, Stephen R. 2010b. Failing one’s obligations: Defectiveness in Rumantsch
reflexes of DĒBĒRE. In Matthew Baerman, Greville G. Corbett & Dunstan Brown
(eds.), Defective paradigms: Missing forms and what they tell us, 19–34. Oxford: British
Academy & Oxford University Press.
Anderson, Stephen R. 2015. Morphological change. In Claire Bowern & Bethwyn Evans
(eds.), The Routledge handbook of historical linguistics, 264–285. New York: Routledge.
Anderson, Stephen R. & David W. Lightfoot. 2002. The language organ: Linguistics as
cognitive physiology. Cambridge: Cambridge University Press.
Anderwald, Lieselotte. 2003a. Negation in non-standard British English: Gaps, regularizations and asymmetries. London: Routledge.
Anderwald, Lieselotte. 2003b. Non-standard English and typological principles: The case
of negation. In Günter Rohdenburg & Britta Mondorf (eds.), Determinants of grammatical variation in English, 507–530. Berlin/New York: Mouton de Gruyter.
Aronoff, Mark. 1976. Word formation in generative grammar. Cambridge, MA: MIT Press.
Baerman, Matthew. 2008. Historical observations on defectiveness: The first singular
non-past. Russian Linguistics 32(1). 81–97.
Baerman, Matthew, Greville G. Corbett & Dunstan Brown (eds.). 2010. Defective paradigms: Missing forms and what they tell us. Oxford: British Academy & Oxford University Press.
Baker, Robert & Philip Smith. 1976. A psycholinguistic study of English stress assignment
rules. Language and Speech 19(1). 9–27.
Baronian, Luc. 2005. North of phonology. Stanford University PhD thesis.
11 How to wake up irregular (and speechless)
Berko, Jean. 1958. The child’s learning of English morphology. Word 14(2–3). 150–177.
Berwick, Robert. 1985. The acquisition of syntactic knowledge. Cambridge, MA: MIT Press.
Berwick, Robert C & Noam Chomsky. 2016. Why only us: Language and evolution. Cambridge, MA: MIT Press.
Bloch, Bernard. 1947. English verb inflection. Language 23(4). 399–418.
Bowerman, Melissa. 1982. Reorganizational process in lexical and syntactic development.
In Eric Wanner & Lila R. Gleitman (eds.), Language acquisition: The state of the art, 319–
346. Cambridge/New York: Cambridge University Press.
Brainerd, Barron. 1989. The contractions of not: A historical note. Journal of English Linguistics 22(2). 176–196.
Bresnan, Joan. 2001. Explaining morphosyntactic competition. In Mark Baltin & Chris
Collins (eds.), Handbook of contemporary syntactic theory, 1–44. Malden, MA & Oxford:
Blackwell Publishing.
Broadbent, Judith M. 2009. The *amn’t gap: The view from West Yorkshire. Journal of
Linguistics 45(2). 251–284.
Brown, Dunstan & Andrew Hippisley. 2012. Network morphology: A defaults-based theory
of word structure. Vol. 133. Cambridge: Cambridge University Press.
Brown, Roger. 1973. A first language: The early stages. Cambridge, MA: Harvard University Press.
Brown, Roger, Courtney B. Cazden & Ursula Bellugi. 1973. The child’s grammar from I
to III. In Charles A. Ferguson & Daniel I. Slobin (eds.), Studies of child language development, 295–333. New York: Holt, Rinehart & Winston.
Bybee, Joan. 1985. Morphology: A study of the relation between meaning and form. Philadelphia: John Benjamins.
Chomsky, Noam. 1955. The logical structure of linguistic theory. Ms., Harvard University
and MIT. Revised version published by Plenum, New York, 1975.
Chomsky, Noam. 1965. Aspects of the theory of syntax. Cambridge, MA: MIT Press.
Chomsky, Noam. 1981. Lectures in government and binding. Dordrecht: Foris.
Chomsky, Noam. 2005. Three factors in language design. Linguistic Inquiry 36. 1–22.
Chomsky, Noam & Morris Halle. 1968. The sound pattern of English. Cambridge, MA: MIT
Clahsen, Harald. 1999. Lexical entries and rules of language: A multidisciplinary study
of German inflection. Behavioral and Brain Sciences 22. 991–1069.
Clahsen, Harald, Monika Rothweiler, Andreas Woest & Gary Marcus. 1992. Regular and
irregular inflection in the acquisition of German noun plurals. Cognition 45. 225–255.
Cutler, Anne & David M. Carter. 1987. The predominance of strong initial syllables in the
English vocabulary. Computer Speech and Language 2(3–4). 133–142.
Daland, Robert T., Andrea D. Sims & Janet Pierrehumbert. 2007. Much ado about nothing: A social network model of Russian paradigmatic gaps. In Proceedings of the 45th
annual meeting of the Association for Computational Linguistics, 936–943. Somerset,
NJ: Association for Computational Linguistics.
Davies, Mark. 2008. Corpus of Contemporary American English (COCA): 410+ million
words, 1990-present. Available at
Charles Yang
Frampton, John. 2001. The amn’t gap, ineffability, and anomalous aren’t: Against morphosyntactic competition. In Mary Andronis, Christopher Ball, Heidi Elston & Sylvain
Neuvel (eds.), Papers from the 37th meeting of the Chicago Linguistic Society, 399–412.
Chicago: Chicago Linguistic Society.
Guion, Susan, J.J. Clark, Tetsuo Harada & Ratree Wayland. 2003. Factors affecting stress
placement for English nonwords include syllabic structure, lexical class, and stress
patterns of phonologically similar words. Language and Speech 46(4). 403–427.
Halle, Morris. 1973. Prolegomena to a theory of word formation. Linguistic Inquiry 4(1).
Halle, Morris. 1998. The stress of English words 1968-1998. Linguistic Inquiry 29(4). 539–
Halle, Morris & Alec Marantz. 1993. Distributed morphology and the pieces of inflection. In Kenneth Hale & Samuel Jay Keyser (eds.), The view from Building 20: Essays in
linguistics in honor of Sylvain Bromberger, 111–176. Cambridge, MA: MIT Press.
Halle, Morris & Jean-Roger Vergnaud. 1987. An essay on stress. Cambridge, MA: MIT
Hart, Betty & Todd R Risley. 1995. Meaningful differences in the everyday experience of
young American children. Baltimore, MD: Paul H. Brookes Publishing.
Hay, Jennifer & R. Harald Baayen. 2003. Phonotactics, parsing, and productivity. Italian
Journal of Linguistics 1. 99–130.
Hayes, Bruce. 1982. Extrametricality and English stress. Linguistic Inquiry 13(2). 227–276.
Hetzron, Robert. 1975. Where the grammar fails. Language 51(4). 859–872.
Hudson, Richard. 2000. *I amn’t. Language 76(2). 297–323.
Jespersen, Otto. 1917. Negation in English and other languages (Historisk-flologiske Meddelelser 1). Copenhagen: A.F. Høst.
Kehoe, Margaret & Carol Stoel-Gammon. 1997. The acquisition of prosodic structure: An
investigation of current accounts of children’s prosodic development. Language 73.
Kelly, Michael H. 1992. Using sound to solve syntactic problems: The role of phonology
in grammatical category assignments. Psychological Review 99(2). 349–364.
Kiparsky, Paul. 1973. Elsewhere in phonology. In Stephen R. Anderson & Paul Kiparsky
(eds.), A festschrift for Morris Halle, 93–106. New York: Holt, Rinehart & Winston.
Labov, William, Paul Cohen, Clarence Robins & John Lewis. 1968. A study of the nonstandard English of Negro and Puerto Rican speakers in New York City. Volume
1: Phonological and grammatical analysis. Cooperative Research Project No. 3288,
Columbia University, New York.
Ladefoged, Peter & Victoria Fromkin. 1968. Experiments on competence and performance. IEEE Transactions on Audio and Electroacoustics 16(1). 130–136.
Legate, Julie A. & Charles Yang. 2013. Assessing child and adult grammar. In Robert
Berwick & Massimo Piattelli-Palmarini (eds.), Rich languages from poor inputs: In honor
of Carol Chomsky, 168–182. Oxford: Oxford University Press.
MacWhinney, Brian. 2000. The CHILDES project: Tools for analyzing talk. 3rd. Mahwah,
NJ: Lawrence Erlbaum.
11 How to wake up irregular (and speechless)
Maiden, Martin & Paul O’Neill. 2010. On morphomic defectiveness: Evidence from the
Romance languages of the Iberian Peninsula. In Matthew Baerman, Greville G. Corbett
& Dunstan Brown (eds.), Defective paradigms: Missing forms and what they tell us, 103–
124. Oxford: British Academy & Oxford University Press.
Marcus, Gary, Steven Pinker, Michael T. Ullman, Michelle Hollander, John Rosen & Fei
Xu. 1992. Overregularization in language acquisition. Vol. 57 (Monographs of the Society for Research in Child Development 4). Chicago: University of Chicago Press.
McClelland, James L. & Karalyn Patterson. 2002. Rules or connections in past-tense inflections: What does the evidence rule out? Trends in Cognitive Sciences 6(11). 465–472.
Nagy, William E. & Richard C. Anderson. 1984. How many words are there in printed
school English? Reading Research Quarterly 19(3). 304–330.
Newport, Elissa. 1990. Maturational constraints on language learning. Cognitive Science
14(1). 11–28.
Nida, Eugene A. 1949. Morphology: The descriptive analysis of words. 2nd. Ann Arbor:
University of Michigan Press.
Niyogi, Partha. 2006. The computational nature of language learning and evolution. Cambridge, MA: MIT Press.
Osherson, Daniel N., Michael Stob & Scott Weinstein. 1986. Systems that learn: An introduction to learning theory for cognitive and computer scientists. Cambridge, MA: MIT
Pertsova, Katya. 2005. How lexical conservatism can lead to paradigm gaps. In Jeffrey
Heinz, Andrew Martin & Katya Pertsova (eds.), UCLA working papers in linguistics 11:
Papers in phonology 6, 13–30. Los Angeles: UCLA Linguistics Department.
Pinker, Steven. 1989. Learnability and cognition: The acquisition of argument structure.
Cambridge, MA: MIT Press.
Pinker, Steven. 1999. Words and rules: The ingredients of language. New York: Basic Books.
Prince, Alan & Paul Smolensky. 2004. Optimality Theory: Constraint interaction in generative grammar. Cambridge, MA: MIT Press.
Pullum, Geoffrey K. & Deidre Wilson. 1977. Autonomous syntax and the analysis of auxiliaries. Language 53(4). 741–788.
Rice, Curt. 2005. Optimal gaps in optimal paradigms. Catalan Journal of Linguistics 4.
Rumelhart, David E. & James L. McClelland. 1986. On learning the past tenses of English
verbs. In James L. McClelland, David E. Rumelhart & the PDP Research Group (eds.),
Parallel distributed processing: Explorations into the microstructure of cognition. Volume
2: Psychological and biological models, 216–271. Cambridge, MA: MIT Press.
Schuler, Kathryn, Charles Yang & Elissa Newport. 2016. Testing the Tolerance Principle:
Children form productive rules when it is more computationally efficient to do so. In
The 38th Cognitive Society Annual Meeting. Philadelphia, PA.
Schütze, Carson T. 2005. Thinking about what we are asking speakers to do. In Stephan
Kepser & Marga Reis (eds.), Linguistic evidence: Empirical, theoretical, and computational perspectives, 457–485. Berlin: Mouton de Gruyter.
Charles Yang
Sims, Andrea D. 2006. Minding the gap: Inflectional defectiveness in a paradigmatic theory.
Ohio State University PhD thesis.
Steriade, Donca. 1997. Lexical conservatism. In Linguistic Society of Korea (ed.), Linguistics in the morning calm 4: Selected papers from SICOL–1997, 157–179. Seoul: Hanshin.
Stump, Gregory. 2001. Inflectional morphology: A theory of paradigm structure. Cambridge: Cambridge University Press.
Warner, Anthony. 1993. English auxiliaries: Structure and history. Cambridge University
Weldon, Tracey. 1994. Variability in negation in African American Vernacular English.
Language Variation and Change 6(3). 359–397.
Wiese, Richard. 1996. The phonology of German. Oxford: Clarendon Press.
Wolf, Matthew & John J. McCarthy. 2009. Less than zero: Correspondence and the null
output. In Curt Rice & Sylvia Blaho (eds.), Modeling ungrammaticality in Optimality
Theory, 17–66. London: Equinox.
Wright, Joseph. 1898. The English dialect dictionary. London: Henry Frowde.
Xu, Fei & Steven Pinker. 1995. Weird past tense forms. Journal of Child Language 22(3).
Yang, Charles. 2002a. A principle of word storage. Manuscript: Yale University.
Yang, Charles. 2002b. Knowledge and learning in natural language. Oxford: Oxford University Press.
Yang, Charles. 2004. Universal grammar, statistics or both? Trends in Cognitive Sciences
8(10). 451–456.
Yang, Charles. 2013. Ontogeny and phylogeny of language. Proceedings of the National
Academy of Sciences 110(16). 6324–6327.
Yang, Charles. 2015. Negative knowledge from positive evidence. Language 91(4). 938–
Yang, Charles. 2016. The price of linguistic productivity: How children learn to break rules
of language. Cambridge, MA: MIT Press.
Yang, Charles, Stephen Crain, Robert C. Berwick, Noam Chomsky & Johan J. Bolhuis.
2017. The growth of language: Universal Grammar, experience, and principles of efficient computation. Neuroscience and Biobehavioral Review.
Yip, Kenneth & Gerald J. Sussman. 1998. Sparse representations for fast, one-shot learning.
Tech. rep. 1633. MIT Artificial Intelligence Laboratory.
Zwicky, Arnold M. & Geoffrey K. Pullum. 1983. Cliticization vs. inflection: English N’T.
Language 59(3). 502–513.
Part III
Syntax and Morphosyntax
Chapter 12
Special clitics and the right periphery in
Judith Aissen
University of California, Santa Cruz
This paper documents the distribution of the definite enclitic =e in Tsotsil (Mayan), a clitic
which occurs on the right periphery of utterances. On the basis of this distribution, it is
argued (contra some restrictive theories of clitic placement) that =e cannot reach its surface
position in the syntax, but must be positioned by the phonology. The property of =e which
determines its placement is its obligatory association with the prosodic peak of the intonational phrase, a peak which is located at the right edge of that phrase. The relation of =e
to several other elements which likewise occur at or near the right periphery of the intonational phrase in Tsotsil is considered, and a possible historical scenario which can account
for the properties of =e is suggested.
1 Introduction
This paper has two goals. The first is to document more fully than has been done previously the distribution of the definite enclitic =e in Tsotsil (Mayan), a clitic which is
restricted to the right periphery of utterances, (§2-§3).1 The second is to suggest that =e
is a special clitic in the sense of Anderson (2005) (following Zwicky 1977): “a linguistic element whose position with respect to the other elements of the phrase or clause follows
a distinct set of principles, separate from those of the independently motivated syntax
of free elements in the language” (31–32). The property of =e which makes it “special” is
the extent to which it may be separated from the phrase in which it is licensed (§4).2 In
the analysis proposed here, this separation results from the requirement that =e function
as the prosodic peak of the intonational phrase in which it occurs (§5), a requirement
which can place it at a significant remove from its syntactically-motivated position. The
requirement of prosodic prominence is unusual for a clitic. Anderson (2005) emphasizes
the fact that clitics cannot be defined by the absence of “accent”, as a clitic can bear an
The distribution of this enclitic is noted in Aissen (1992: 61) but without much supporting data or discussion.
It is also discussed in Skopeteas (2010) as part of a broader treatment of terminal clitics in Mayan languages.
This property is emphasized in Skopeteas (2010).
Judith Aissen. 2017. Special clitics and the right periphery in Tsotsil. In Claire Bowern,
Laurence Horn & Raffaella Zanuttini (eds.), On looking into words (and beyond), 235–
262. Berlin: Language Science Press. DOI:10.5281/zenodo.495449
Judith Aissen
accent if it happens to fall in an accented position within a larger prosodic constituent.
But cases in which a clitic is required to occupy such a position – and will reorder in
order to reach it – have not, to my knowledge, been documented. §6 speculates on how
=e might have come to be associated with the phonological properties that force it to its
surface position.
The fact that =e can occur outside the syntactic domain in which it is syntactically
licensed poses a challenge to theories which hold that clitics reach their surface positions through syntactic operations, e.g., Bošković (2000) and Bermúdez-Otero & Payne
(2011). For them, even clitics which are pronounced in prosodically determined positions
nonetheless reach those positions in the syntax, with the role of phonology limited to
filtering the outputs of a possibly overgenerating syntax. §4 suggests that this view is difficult to maintain in the case of =e. It thus adds to a body of work which has argued that
phonology can determine word order, especially in the case of weak elements (Halpern
1995; Chung 2003; Agbayani & Golston 2010; Agbayani, Golston & Ishii 2015; Bennett,
Elfner & McCloskey 2015).
2 The definite enclitic in Tsotsil
2.1 The basics
All dialects of Tsotsil have at least one enclitic which is associated with definite determiners, as well as with several other elements. The dialects differ with respect to how many
such clitics they have, how many determiners they have, and what other elements the
clitics associate with. Under discussion here is the dialect of Zinacantec Tsotsil (Z Tsotsil). Z Tsotsil has one such clitic, =e.3 Among other elements, =e is associated with both
of the definite determiners, li (proximate) and ti (remote) (this association is indicated
in examples by an overbar).4
a. I-bat la ti vinik=e.
cp-go cl det man-def
‘The man went (they say).’ (Laughlin 1977: 28)
Tsotsil is spoken in Chiapas, Mexico by some 400,000 people. Claims made here about Zinacantec Tsotsil
are based on a large body of text material and work with five native speakers over a number of years. Texts
include naturally occurring speech, texts originally written in Tsotsil, and texts translated from Spanish
to Tsotsil (the New Testament, cited as nt). Grammatical examples are almost all taken from texts; unpublished sources are cited as author. Examples cited as ungrammatical have been checked with several
speakers and their impossibility is consistent with the patterns seen in the text material.
4 Like other Mayan languages, Tsotsil is verb-initial, usually V(O)S. It is also a head-marking language with
ergative alignment. Affixes glossed erg, abs, gen express φ features of arguments on agreeing nouns and
verbs. Absolutive 3rd singular has no exponent and is not indicated in examples. Orthographic symbols
have the expected values except for x = [ʃ], j = [x], ch = [tʃ], and ’ = [ʔ] (except in symbols for ejectives, p’,
t’, ts’, ch’, k’).
12 Special clitics and the right periphery in Tsotsil
b. Buy li j-ve’el=e?
where det gen.1-meal-def
‘Where is my meal?’ (Laughlin 1977: 57)
The deictic distinctions made by determiner+enclitic are fairly subtle and both determiners can be translated by English ‘the’. More salient distinctions are made by incorporating deictic adverbs into the dp. As these examples suggest, =e occurs in a “final” position
and I sometimes refer to it as a terminal clitic. This distinguishes it both from second
position clitics (e.g., the reportative clitic la in (1a)) and from terminal elements which
are not clitics (e.g., those discussed in §5.2).
2.1.1 Licensing
There is a dependency between the definite determiners and =e: the determiners almost
always co-occur with =e. Written texts rarely omit it, and speakers judge sentences without it to be “incomplete”. In spoken language, =e is sometimes omitted, perhaps due to
performance factors, to register, to individual speaker style, or to some other factor. The
claims made here hold for relatively careful speech and for written texts. Other elements
which license =e include a set of deictics which function as demonstratives and adverbs,
as well as certain subordinators. The lexical elements which license =e in Z Tsotsil are
shown in Table 1. The determiners li and ti figure in many of the temporal adverbs and
Table 1: =e licensors in Zinacantec Tsotsil
Definite determiner
li (prox)
ti (distal)
Spatial demonstrative/adverb
li’ ‘(this) here’
le’ ‘(that) there’
taj ‘(that) over there’
Temporal adverb
lavi ‘today’
ti (complementizer)
ti mi ‘if’
(ti) k’alal ‘when’
(ti) yo’ ‘place where’
subordinators listed in Table 1: in the third category, lavi ‘today’ is derived from li avi; in
the fourth, the complementizer ti may be the determiner, serving to nominalize a clause;
mi is the polar question particle, but always occurs with ti when it introduces the protasis to a conditional; k’alal ‘when, the time when’ frequently occurs in collocation with
Judith Aissen
ti, as does yo’ (‘place where’). I assume then that =e realizes the feature [+def] in this
dialect.5 Examples (2a,b) show =e licensed by elements other than determiners:
a. Och-an ech’el li’ ta ch’en=e!
enter-imp dir here in cave-def
‘Enter the cave here!’ (Laughlin 1977: 71)
b. K’alal i-k’ot
ta s-ch’en=e…
when cp-arrive p erg.3-cave-def
‘When he arrived at his cave…’ (Laughlin 1977: 72)
Aside from the qualification noted in fn. 5, elements which are not [+def] do not
license =e in Tsotsil. This includes lexical categories (nouns, verbs, adjectives), related
semi-functional categories like auxiliaries, and functional categories like the indefinite
article, prepositions, negation, focus markers, coordinators, etc. Thus, =e does not occur
in the position marked by the asterisk in any of the following examples as none of them
contains an appropriate licensor.
a. S-nup
la ta be jun tseb un *.
erg.3-meet cl on path indf girl par
‘He met a girl on the path.’ (Laughlin 1977: 306)
b. I-k’opoj la tal
ta vinajel *.
cp-speak cl coming p heaven
‘He spoke on arriving in heaven.’ (NT: Mark 1,11)
c. Ta xa x-’och
k’ok’ ok’ob
icp cl asp-enter fire tomorrow.
‘The war will start tomorrow.’ (Laughlin 1977: 119)
2.1.2 Terminal position: 1st approximation
Examples (1)-(2) suggest that =e occurs at the right edge of the phrase headed by its
licensor. We will need to revise this, but it is true that =e in dp’s, for example, must
follow all post-head material in the phrase, including modifiers (4a,b) and possessors
(4c). There are no other possible positions for =e in these examples – in particular, it
=e sometimes occurs without an overt licensor, but still associated with a definite interpretation. Nominal
cases include 1st and 2nd person pronouns (in certain syntactic positions), proper names (occasionally),
and headless relatives with definite interpretations (frequently). These are all clearly definite, so association with a [+def] head seems unproblematic. Certain semantically dependent clauses can also end in =e
without an overt licensor being present (e.g., a determiner or subordinator). These usually present background (given) information and correspond, for example, to English when or since clauses. Whether =e in
these cases should be viewed as the realization of a [+def] feature or some other related feature is unclear.
The clausal cases are not directly relevant to present concerns since =e is never separated in these from the
domain in which it is licensed (the entire clause). Hence I leave them aside.
12 Special clitics and the right periphery in Tsotsil
absolutely cannot attach to the head noun nor to the first prosodic word in the phrase
(these positions are marked with asterisks).
a. [ti moletik * vo’ne tey ta Ats’am=e]DP
det elders
long.ago there p Salinas-def
‘the elders of long ago from (there in) Salinas’(author)
b. [ti anima * j-muk’tot=e]DP
det late
‘my late grandfather’(author)
c. [li j-me’
* [li vo’on=e]DP ]DP
det gen.1-mother det pro.1sg-def
‘my mother’(author)
(4a-c) come from texts in which the dp is a topic. These occur “external” to the clause and
are thus isolated from the effects of other elements which (as we will see below) interact
with the position of =e.
2.1.3 Coalescence
An important property of =e is coalescence. In (4c), the larger dp contains two licensors,
each of which should be matched by =e. One (the first li) is the head of the larger dp (the
possessum), the other (the second li) is the head of the embedded dp (the possessor). The
right edge of the two dp’s coincide and only a single clitic is possible at this edge. This
is a general property of terminal clitic systems in Mayan; even when multiply licensed,
only a single such clitic occurs (within the relevant domain) (Skopeteas 2010).
2.1.4 Clitic vs. affix
Though it is generally accepted that “clitic” is a cover term for a diverse set of elements
and not a formal grammatical category, the term is still used descriptively. To motivate
the use of the term “clitic” to refer to Tsotsil =e, I survey some of the criteria that have
been used in the past to distinguish clitics from (ordinary) affixes (Zwicky & Pullum
1983). All of these align =e more closely with “clitics” than with inflectional affixes. [1]
it imposes no selectional restrictions on the host, but may attach to members of any
lexical category that falls in the appropriate right-edge position. In addition to nouns,
these include verbs, as in (5c), adjectives, particles (see §5.2), and even second position
clitics like the reportative clitic la in (5a); [2] there are no arbitrary gaps in the possible
X=e combinations; [3] the form of the host is not sensitive to the presence of the clitic
(the clitic triggers no allomorphy and does not participate in lexical phonology); [4] there
Judith Aissen
are no semantic idiosyncracies associated with =e; and [5] =e attaches outside all other
suffixes, e.g., noun plurals, (5b), and agreement suffixes, (5c).
a. a ti vo’ne la=e …
top det long.ago cl-def
‘as for long ago (they say)’
b. ti jeneral-etik=e
det general-pl-def
‘the generals’
c. li tak’in ta j-ta-tikotik=e
det money icp erg.1-find-1pl.excl-def
‘the money that we could find’
At the same time, =e is prosodically more like an affix than other clitics in the language. Tsotsil has various “simple” clitics, i.e., syntactic words which are prosodically
weak. Like other words in the language, all of these have an onset, e.g., the interrogative polarity particle mi, the definite determiners ti, li, negation mu, second position
modal and aspectual clitics (xa, to, me, la). In contrast though, =e, like many inflectional
affixes, lacks an onset. Further, except for the second position clitics, the simple clitics
all precede their complements, while =e follows everything in its phrase.
If “clitic” is not a formal grammatical category, then the properties of =e must follow
from its analysis as a word or affix. There are a number of possible analyses that could
be considered. We could analyze it as a prosodically deficient word which heads its own
phrase within the dp, as shown in (6).
12 Special clitics and the right periphery in Tsotsil
Here, =e heads a DefP which is selected by D and which itself takes a NP complement.
We could account for the phrase- final position of =e by assuming that =e requires that
its specifier be filled, and that the NP complement raises to its left to satisfy this requirement (this would follow proposals of Cinque 2005 and Simpson 2005, who account for
the phrase-final position of demonstratives in various languages via leftward movement
of NP within DP).6 Another possibility would be to analyze =e as inflectional morphology which spells out a definiteness feature associated with the noun phrase on the rightmost terminal of that phrase, much as Miller (1991) analyzes the French deictic clitics
-ci and -là. A third possibility is to analyze =e as a phrasal affix, analogous to the treatment that Anderson (2005) proposes for the English genitive marker ’s and somewhat
tentatively for definitive accent in Tongan. In this approach, =e would be introduced
post-syntactically by the phrasal morphology as spell-out of the feature [+def] on DP
and its surface position would be determined by a constraint operating within an OT
constraint system. Any of these approaches will have to confront the issues discussed
in the next section; how well each would fare is not a question I address here. Going forward, I will assume the analysis sketched in (6), according to which =e is a prosodically
deficient word which is introduced in the syntax.
Under any of these analyses, =e is licensed within the phrase headed by its licensor,
usually dp, and I take this phrase to be the “syntactic domain” of the clitic. The puzzle
that gives rise to this paper is the fact that =e does not in fact always close the phrase
in which it is licensed but often occurs considerably further to the right. I argue below
that this is because =e can occur only at the right edge of an intonational phrase (ιP),
an edge which is often located further to the right than the right edge of the phrase in
which =e is licensed. The evidence for this is presented in §3; in §5, I consider why =e is
constrained in this way.
3 Prosodic constraints on =e
Although =e frequently appears at the right edge of the phrase in which it is licensed, the
larger descriptive generalization about its position is not syntactic, but prosodic (Aissen
1992; Skopeteas 2010):
=e occurs at the right edge of the ιP which contains its licensor.
Descriptions of Z Tsotsil characterize prosodic prominence at two levels – the word
and the phrase. At the word level, stress falls on the initial syllable of the root; at the
phrase level, it falls on the final syllable of the ιP (Laughlin 1975,23; Haviland 1981,14)
(stress being predictable, it is not marked in the orthography). I assume then that the
final syllable of the ιP is its prosodic peak.7 A detailed phonetic study of intonational
Note that this movement would violate the anti-locality constraints proposed in Pesetsky & Torrego (2001)
and Abels (2003) which preclude movement of the complement of a head (a phase head in Abels’ account)
to the specifier of that head.
7 The association of prosodic prominence with the final syllable of the ιP is reported for other Tsotsil dialects
(Cowan 1969: 4; Delgaty & Sánchez 1978: 11) as well as for the sister language Tseltal (Shklovsky 2011; Polian
2013); see Bennett (2016: §6.1) for an overview of lexical and phrasal stress in Mayan.
Judith Aissen
phrasing in Tsotsil does not yet exist, but some preliminary observations are possible.
The final syllable is associated with characteristic boundary tones. The most common
pattern involves a rise in pitch on the vowel of the final syllable, with the larger context
determining whether that rise is sustained throughout the syllable or followed by a fall
(relevant factors include whether the ιP is final in the utterance or not (as in the case
of topics, for example, §3.2)). The final syllable of the ιP is sometimes followed by a
significant pause and when it is, the vowel of that syllable is often lengthened.
Some of these properties are evident in Figure 1, taken from a naturally-produced
narrative by a Z Tsotsil speaker; this example occurs utterance-finally and shows a final
Pitch (Hz)
Figure 1: Pitch track and waveform for (8).
ta anil.
cp-abs.1-come-1pl.excl in hurry
‘We came in a hurry.’ (author)
A key observation is that because =e aligns with the right edge of the ιP, then, whatever else it is, it is the final syllable of the ιP. It thus carries the boundary tone, and is
often followed by significant pause and lengthened. This is illustrated in Figure 2, which
is based on (9), from the same narrative as Figure 1; this phrase is also utterance-final.
… te
ta s-na
li Maryan Papyan=e.
there p gen.3-house det Mariano Papyan-def
‘…there in the house of Mariano Papyan.’ (author)
The analysis proposed in §5 hinges on the obligatory association between =e and the
prosodic peak of ιP.
12 Special clitics and the right periphery in Tsotsil
Pitch (Hz)
Figure 2: Pitch track and waveform for (9).
As in other languages, utterances consisting of a simple clause are parsed as a single
ιP. There are also two structures in Z Tsotsil which are associated with obligatory ιP
breaks, resulting in utterances with multiple ιP’s, and therefore multiple positions for
=e under (7): an external topic is parsed as an ιP separate from that of the following
comment clause and an extraposed cp is parsed as a ιP separate from that the preceding
matrix clause.8 Other complements, as well as relative clauses, are usually not extraposed
and they are prosodically integrated into the ιP of the matrix clause. In this section we
provide support for (7), starting with simple clauses (§3.1), then considering structures
with multiple ιP’s (§3.2-§3.3), and finally syntactically complex structures which map to
a single ιP (§3.4). §3.5 suggests an algorithm for mapping syntactic structure to prosodic
3.1 Simple clauses
In utterances consisting of a single clause, regardless of where =e is licensed, it appears
at the right edge of the ιP corresponding to the clause. When the licensing phrase itself
is clause-final, as in (1)-(2), that phrase has the appearance of being closed by =e. But if a
clause contains several phrases which are headed by licensors, no phrase which occurs
medially can end in =e. Adding it in in the positions of the asterisks in (10) and (11) is
Some adverbial clauses are obligatorily parsed as separate ιP’s and some only optionally. These are not
discussed here, but see Aissen (1992: 59).
Judith Aissen
(10) S-jipan la ta=ora
[ti ok’il *] [ti t’ul] un=e.
erg.3-tie cl right.away det coyote det rabbit par-def
‘The rabbit tied Coyote up right away.’ (Laughlin 1977: 160)
la tal [li aniyo *] ta yut=vo’
[li choy] un=e.
cp-erg.3-find cl dir det ring
p inside.water det fish par-def
‘The fish found the ring in the water.’ (Laughlin 1977: 354)
One might think that =e is simply omitted when the licensing phrase does not occur
clause-finally. But examples like (12)-(14) show otherwise. Here (and generally), the
clause-medial dp does license =e but the clitic is delayed to the end of the clause.
xchi’uk [li Kumpa
Lol]dp ta museo-e.
cp-abs.1-work-1pl.excl with det Compadre Lol p museum-def
We worked with Compadre Lol at the museum. (Laughlin 1980: 25)
Ch-’och xa [li k’ok’]dp [ok’ob]adv [ta Nibak]pp =e.
tomorrow p Ixtapa-def
icp-enter cl det fire
‘The war will begin tomorrow in Ixtapa.’ (Laughlin 1977: 119)
Ta=x-[y]-ak’-ik [ti kantela]dp [noxtok]adv =e.
icp-erg.3-give-pl det candle
‘They too were offering the candles.’(author)
There are two properties to note in these examples. First, =e must be licensed by the
determiner since there is no other licensor present; and second, the intervening pp’s and
adverbs are not part of the dp headed by the licensor. In (12)-(14), they modify the entire
sentence (or the predicate), not the head noun. In (14), the adverb noxtok ‘too, also’
is associated with additive focus on the subject ‘they’ (= shamans in the town under
discussion) not the object (‘the candles’) – the preceding discourse describes shamans
from a neighboring town offering candles; the current utterance asserts that the ones
in this town too were offering candles. =e attaches then outside its syntactic domain,
assuming that domain to be the dp headed by its licensor.
Going back to (10)-(11), the right conclusion, I think, is that both determiners require
=e, but that that requirement is satisfied by the single, clause-final enclitic (see also
Skopeteas 2010). These cases too then involve coalescence, but in a configuration different from the one illustrated by (4c). In (4c), the right edges of the two dp’s which
12 Special clitics and the right periphery in Tsotsil
license =e coincide, but here they do not. (10)-(11) actually provide another kind of evidence that =e does not always close its syntactic domain: the particle un which occurs
in both examples (and in many subsequent ones) is not part of the preceding dp, yet
whenever it occurs, it separates =e from its licensing phrase, (see §5.2 on un).
Examples like (15) and (16) provide further evidence that =e can occur outside its syntactic domain: they show that when the phrase that licenses =e is preposed, =e still
surfaces in post-verbal position, at the right edge of the clause. (15) is from a narrative in
which a mother gives advice to her son, (16) from one about the Kennedy assassination.
[Ta sba me l-av-ajnil]pp
on top cl det-gen.2-wife icp-abs.2-climb-def
‘It’s on top of your wife that you should climb [not onto the rafters].’ (Laughlin
1977: 56)
ta bala=e.
(16) Ja’ nox [li viniketik]dp i-laj-ik
cp-end-pl p bullet-def
foc only det men
‘[The women weren’t hit by the bullets], it was only the men that were wounded
by bullets.’ (Laughlin 1980: 15)
In (15), a pp has been fronted into focus position, as sketched in (17) (the larger context makes clear that we are dealing with contrastive focus in both (15-16)). Note that a
fronted focus does not occasion an ιP break (Aissen 1992).
(17) = structure of (15)
you climb
your wife
The licensor for =e in (17) is the head of the circled dp, which is embedded quite deeply
within the fronted pp, but the enclitic does not close that dp. Instead it surfaces clausefinally. (16) is a cleft construction where the focus occurs preverbally. Again =e is licensed
Judith Aissen
by the head of that dp but occurs clause-finally (the verb phrase which follows the focus
does not modify the focus and is presumably not embedded in it).
With respect then to simple, monoclausal structures, examples (12)-(16) show (in various ways) that in Z Tsotsil, =e does not in general close the phrase headed by its licensor.
A closer approximation is that it closes the clause containing the licensor (though we
will see shortly that this is not the whole story either). This holds whether the licensor is a determiner or some other element, e.g., a deictic adverb. (18)-(19) show that
an =e licensed by a deictic adverb also occurs clause-finally, again separated from the
phrase containing the licensor by intervening material (in (19), the adverb functions as
the clausal predicate).
[lavi] [ta k’in]-e. today p fiesta
‘Let’s arrest him today at the festival.’ (NT: Matthew 26:5)
(19) Muk’ li’ s-malal=e.
neg here gen-husband-def
‘Their husbands weren’t around here [they had gone to the lowlands].’ (Laughlin
1977: 101)
3.2 Topics
As in many other languages, external topics in Tsotsil are parsed as separate ιP’s (by
“external topic”, I mean one which is attached outside the sentence, often entering into
an anaphoric relation with a pronoun inside the sentence) (Aissen 1992). Topics are
usually definite in Tsotsil and therefore are almost always closed by =e (the ιP break is
indicated by “‖”):
(20) Ti moletik vo’ne tey ta Ats’am=e, ‖ i-s-tsob
la s-ba-ik
det elders long.ago there p Salinas-def cp-erg.3-gather cl gen.3-rr-pl
ta snuts-el
li biyaetik=e.
p chase-nomzl det Villistas-def
‘The elders of long ago (from) there in Salinas gathered to chase the Villistas.’
Ti anima j-muk’tot=e
‖ x-’ok’ xa la sutel
det late gen.1-grandfather-def asp-cry cl cl returning here
‘My late grandfather returned crying.’ (author)
12 Special clitics and the right periphery in Tsotsil
3.3 Complex clauses with CP complements
CP complements obligatorily extrapose in Tsotsil. While normal order in transitive
clauses is vos, when o is a cp complement, it occurs utterance-finally (Aissen 1992).
ti s-me’
un=e ‖
cp-erg.3-see det gen.3-mother par-def
muk’=bu ta s-sa’
ti s-krem un=e .
comp neg
icp erg.3-seek gen.3-wife det gen.3-son par-def
‘His mother saw that her son was never going to find a wife.’ (Laughlin 1977: 55)
Extraposition is associated with an obligatory ιP break and, as expected, the matrix and
cp complements form separate domains for clitic placement: the =e licensed by the first
determiner closes the first ιP and the one licensed by the second closes the second ιP.
Extraposition of cp complements also occurs in ditransitive clauses. While the theme
precedes the goal when both are nominal, the theme follows the goal when it is a cp:
(23) Ikalbe li kumpa Lol un=e ‖ ti
yu’un chicham xa un=e.
I.told det compadre Bob par-def comp because I.was.dying cl par-def
‘I told Compadre Bob that I was feeling awful.’ (Laughlin 1980: 30)
Again, extraposition forces an ιP break between the matrix clause and its extraposed
complement. And as above, the two clauses form separate domains for clitic placement.
3.4 Prosodically integrated subordinate clauses
While cp complements extrapose, there are other embedded clauses which do not and
thus remain in their base position. These include ip complements (selected by verbs of
perception and some other higher predicates) as well as relative clauses. Prosodically
these do not form separate ιP’s, but are integrated into the ιP of the matrix clause (see
An 2007 on languages in which restrictive relatives do not form separate ιP’s).
Consider the ip complement in (24). It remains in its internal position and is followed
by the matrix subject:
Mi ja’uk o=bu y-a’i
[lok’ ti y-ajnil
*]ip ti vinik un=e.
neg even ever erg.3-feel leave def gen.3-wife
det man par-def
‘The man didn’t even feel his wife slipping out.’ (Laughlin 1977: 49)
There is no extraposition here and the entire utterance is pronounced as a single ιP. If
=e closed the (smallest) clause in which was licensed, we would expect one to surface
in the position of the asterisk. But =e is not possible there. Instead, it appears that the
enclitic licensed within the complement is delayed until the end of the entire utterance,
Judith Aissen
where it coalesces with the one licensed by the subject. Consistent with (7), the enclitic
licensed within the complement clause is pronounced at the right edge of the ιP which
contains its licensor.
Relative clauses (rc) also generally do not extrapose. Relative clauses with external
heads do not occur utterance-internally (if necessary, the sentence is restructured so
that they occur utterance-finally or sentence-initially as part of the topic), but headless
relatives (or better, “light-headed” relatives involving a determiner + cp) can.9 In (25)
and (26), the rc is sandwiched between the matrix verb and the matrix subject.
(25) Y-il-oj
[ti [bu k’ot ti j’ik’al *]RC ] ti vinik un=e.
erg.3-see-prf det where arrive det Spook
det man par-def
‘The man saw (the place) where the Spook landed.’ (Laughlin 1977: 63)
la [taj [k’alal ch-lok’ tal taj
cp-erg.3-feel cl det when icp-leave dir det
chon *]RC ] taj ants
det woman par-def
‘That woman felt (the moment) when that snake left.’ (Laughlin 1977: 371)
Like ip complements, rc’s do not constitute separate ιP’s, but are parsed together with
the matrix clause. Examples (25)-(26) show that an =e licensed in such a relative clause
is realized not at the edge of the relative clause (marked here by an asterisk), but again
at the right edge of the entire utterance where it coalesces with the clitic licensed by the
matrix subject.
3.5 Summary
The position in which =e is pronounced in Z Tsotsil does not coincide with the edge of
the phrase in which is licensed, nor even with the edge of the (minimal) clause in which
it is licensed. Rather, it coincides with the right edge of ιP containing its licensor.
While it is not necessary for our purposes to provide an algorithm for mapping syntactic structure to prosodic structure (what is important is that ιP breaks fall in certain
positions, not why they fall there), there is a simple principle which determines this mapping if we assume that external topics and extraposed clauses are both adjoined at the
root of the sentence (Aissen 1992). Assuming that an element X which adjoins to Y is not
dominated by Y, then neither topics nor extraposed clauses are dominated by any node.
Hence, like simple clauses, the nodes which define these constituents are “undominated”.
In this respect they are like root nodes and, following Frank, Hagstrom & Vijay-Shanker
I take the rc to be a cp since it contains a fronted wh expression in (25) and a complementizer in (26).
12 Special clitics and the right periphery in Tsotsil
(2002), I will refer to them as such. With this understanding, the mapping from syntax to ιP can be characterized as a match between a certain syntactic constituent (one
dominated by a root node) and a corresponding prosodic constituent (an ιP) (on Match
constraints, see Selkirk 2009 and Elfner 2012; on the relevance of the root to defining ιP,
see Downing 1970; Nespor & Vogel 1986; Selkirk 2009). The formulation in (27) is based
on Bennett, Elfner & McCloskey (2015) and Elfner (2012),
(27) Match Root
If a root node R in a syntactic representation S dominates all and only the set of
terminal elements {a,b,c, …, n}, then there must be in the phonological representation P corresponding to S an ιP which dominates all and only the phonological
exponents of a,b,c, …, n.
When Match Root is satisfied in Tsotsil (as it appears always to be), simple clauses and
some complex structures are parsed as single ιP’s; extraposed CP’s and topics are parsed
into their own ιP’s.
In §5, I develop an account of clitic placement in Tsotsil in which the syntax positions
=e at the edge of the phrase in which it is licensed, per (6), and the phonology accounts for
its subsequent dislocation to the right edge of ιP. This attributes a more significant role
to the phonology than some theories of clitic placement permit. Hence before turning
to the phonological account, I consider the prospects for accounts in which phonology
plays at most a filtering role in the placement of =e in Tsotsil.
4 A syntactic account?
While recognizing that the positioning of some clitics is sensitive to prosodic constituency, some recent theories of clitic placement propose that the role of phonology is
limited to filtering outputs from the syntax. Consider, for example, Bošković (2000)’s
account of second-position clitics in Serbo-Croatian. Bošković argues that these clitics
attach to the first prosodic word within an ιP. This is a prosodic generalization, but in
his account, the prosody does not directly determine the position of second-position
clitics. Rather, clitics reach their surface positions through syntactic mechanisms. Since
syntactic mechanisms sometimes place clitics in other than “second” position, PF filters
out derivations in which the clitics do not suffix to the initial prosodic constituent in the
ιP. Bermúdez-Otero & Payne (2011) propose that all cases of prosodic conditioning of
clitic placement can be handled in the same way, i.e., clitics are positioned by a possibly
over-generating syntax, with ill-formed configurations filtered out at PF.
The problem posed by =e is clear. If its syntactic domain is the phrase headed by its
licensor (typically, dp), then the syntax should place =e somewhere within that domain.
However, we have seen that =e can occur outside the phrase in which it is licensed,
indeed outside the clause in which it is licensed. In fact, it must occur outside that phrase
(or clause) when it is not ιP-final. The only option for an account of clitic placement
in which phonology does no more than filter outputs from the syntax is to extend the
Judith Aissen
syntactic domain of =e (or the [+def] feature which it realizes) beyond the phrase in
which it is licensed.
Conceived syntactically, the dependency between the position in which =e is licensed
and the position in which it is pronounced can span a significant amount of syntactic
structure – it crosses clause-boundaries including ones which define relative clauses.
There are various ways that apparent long-distance dependencies are handled, depending both on the nature of the dependency and on the particular syntactic model – longdistance movement (Transformational Grammar), a sequence of local movements (Minimalism), feature percolation (GPSG/HPSG) and others (Alexiadou, Kiss & Müller 2012).
It is beyond the scope of this article to develop a syntactic analysis of =e placement, but
we can point out two properties of the phenomenon that any such analysis must account
for. One is that the top of the dependency is limited to root (undominated) nodes: =e can
spell out only at the right edge of an undominated node, and not at the right edge of any
other node. If movement or percolation are involved, they must therefore be to the root,
whether that node corresponds to a simple clause, a topic, or an extraposed complement.
The other is that the bottom of the dependency can be located anywhere within the structure dominated by the root. In particular, it can be located within a constituent which
is otherwise an island for extraction, for example within a pp, as in (15/17) (see Aissen
1996 for evidence that pp’s are islands for extraction), or a relative clause, (25)-(26) (see
Aissen 1992).
It is instructive to consider a particular analysis which would position =e in its low,
syntactically-licensed position and account for its appearance at the right edge of ιP’s
through late, prosodically-conditioned linearization. Bermúdez-Otero & Payne (2011)
mention this as a possible analysis for cases in which a clitic attaches to a prosodically
defined domain, like the second position clitics in Chamorro (Chung 2003). They point to
Linear Syntax (Kathol 2004), a theory of linearization embedded in HPSG, as a possible
framework for implementation. Linear Syntax imposes precedence relations on sisters
but, in order to handle discontinuities, permits those relations to be “passed up” the tree
and then “shuffled” with relations among higher elements. In this way, elements from
an embedded domain may be separated from one another by elements that belong to
higher syntactic domains. In the case at hand, =e, linearized, for example at the right
edge of the phrase in which it is licensed, could be separated from that phrase at higher
levels, extending its syntactic domain to a higher constituent.
The question for this account is just what constraints it imposes on the upward “percolation” of precedence relations. In a language which does not in general permit scrambling, which nodes pass precedence relations upwards and which do not? The most
obvious challenge is posed by the fact that an =e licensed somewhere within a relative
clause or a pp cannot surface within those phrases if they are not utterance-final, but
must surface in the matrix. In the shuffling account of examples like (15/17) and (25)-(26),
the precedence relation between =e and the rest of the licensing phrase (its specifier,
under (6)) would be obligatorily passed up through the relative clause or pp and then
shuffled with precedence relations among elements in the matrix clause. Since pp’s and
relative clauses are otherwise impermeable in Tsotsil, one must wonder why Shuffling,
but not other syntactic operations, can access elements within them.
12 Special clitics and the right periphery in Tsotsil
On the other hand, it is a prosodic fact, independent of anything about =e, that pp’s and
relative clauses in Tsotsil do not form separate ιP’s. Hence the fact that an =e licensed
within them surfaces outside them when they are not utterance-final follows from the
prosodic generalization in (7). In short, if the relation between =e and the phrase in which
it is licensed is conceived as a syntactic dependency, its properties are unexpected. But
if the relation is instead phonological and holds within an ιP at a point when syntactic
structure is no longer relevant, the distribution of =e and its relation to the licensing
phrase begin to make sense.
5 A prosodic account
5.1 Association with prosodic prominence
I outline here an account of =e in Z Tsotsil. This account shares with Anderson’s 2005
approach to clitic placement the assumption that the surface position of =e is determined
post-syntactically through an optimization that evaluates alternative positions of the
clitic against a set of ranked constraints (Prince & Smolensky 1993/2004). It differs from
Anderson in that =e is not itself subject to a constraint which aligns it with the edge
of a constituent. Rather the position of =e is motivated by an inherent lexical property,
namely its association with the prosodic prominence that characterizes the right edge
of ιP’s in the language. In this, I closely follow Henderson (2012)’s account of certain
“status” suffixes in K’iche’ (also Mayan),10 which surface only at the right edge of ιP.
These suffixes attach only to verbs and surface only when the verb occurs ιP-finally, (28a).
Otherwise, the suffix is suppressed, (28b) (accent marks here represent the prosodic peak
of the utterance):
a. X-in-tij-ó.
‘I ate it.’
le súb’.
b. X-in-tij
cp-erg.1sg-eat det tamalito
‘I ate the tamalito.’ (Henderson 2012: 775–776)
Henderson notes that status suffixes are simply omitted from phrase-medial verbs, rather
than being displaced to ιP-final position (see 28b) and attributes this to the fact that the
suffix is an affix (not a clitic) and attaches only to verbs. He raises the issue of what would
happen if the element in question were a clitic. The distribution of Tsotsil =e instantiates
exactly this case: =e is not tied to any particular word class and thus faithful realization
carries it away from the position in which it is licensed.
The lexical entry for =e is shown in (29), where the asterisk indicates association with
the prosodic peak of ιP:
These suffixes mark the transitivity status of the predicate and make other distinctions related to mood
and dependency.
Judith Aissen
I also adopt Henderson’s constraint set, it being as well-suited to Tsotsil =e as it is to
the K’iche’ status suffixes. The constraints fall into three groups. The first two concern
the location of prosodic prominence in the ιP and are independent of the distribution
of =e. An alignment constraint (McCarthy & Prince 1993) locates the peak of prosodic
prominence at the right edge of the ιP, (30). Culminativity (31) limits such peaks to one
per ιP (Hayes 1995).
Alignι : A peak of prominence lies at the right edge of the ιP.
Culm(inativity): Every prosodic domain has exactly one peak of prominence.
The second two are faithfulness constraints on the morphology-to-phonology correspondence (Prince & Smolensky 1993/2004; McCarthy & Prince 1995). RealizeMorph (32),
a general constraint, calls for faithful parsing of morphemes in the phonology (Kurisu
2001). IdentProm (33) is the key constraint here: it requires that the lexical association
of =e with prosodic prominence be preserved in the output (Henderson 2012).11
(32) RealizeM(orph): Every morpheme in the input has a phonological exponent in
the output.
(33) IdentProm: if morpheme M has prominence P in the input, then M’, the phonological correspondent of M, has prominence P in the output.
Tableau (34) shows the effect of these constraints on the evaluation of an input, that of
(12), in which the syntactically determined position for =e does not correspond to the
right edge of an ιP. The input in Tableau (34) is a morphophonological representation
in which syntactic terminals have been spelled-out and in which the hierarchical structure of syntax has been replaced by precedence relations and prosodic structure. =e is
a morphophonological element. Its position is syntactically determined per (6) and its
association with the prosodic peak is indicated in the input by the asterisk, a morphological diacritic. Candidates for the output are fully linearized phonological representations,
parsed into prosodic constituents. Prosodic prominence in the ιP is marked by an acute
The optimal candidate is [b], which violates none of the constraints shown. However,
it does violate one which is not shown, Linearity, which penalizes outputs which diverge from the precedence relations of the input (McCarthy & Prince 1995).12 Linearity
must be lower ranked than any of the four constraints shown in Tableau (34).
I have slightly reworded IdentProm from Henderson to emphasize the distinction between M in the input
and its correspondent M’ in the output.
The high-ranked constraint Match Root (27) prevents =e from moving “too far”, by requiring that it be
realized within the same ιP as its licensor.
12 Special clitics and the right periphery in Tsotsil
Tableau for (12)
[… li Kumpa Lol=e* ta museo]ι
… det compadre L-def p museum
a. […li kumpa lol=é ta museo]ι
b. + […li kumpa lol ta museo=é]ι
[…li kumpa lol=e ta museó]ι
[…li kumpa lol-é ta museó]ι
[…li kumpa lol ta museo]ι
Lin(earity): The precedence structure of the input is consistent with that of the
output and vice versa.
When the input has two enclitics, they coalesce in the output.
S-jipan la ta=ora
[ti ok’il [ti t’ul] un=e.
erg.3-tie cl right.away det coyote det rabbit par-def
‘The rabbit tied Coyote up right away.’ (Laughlin 1977: 160)
Taking the input to (36) to be […ti ok’il=e* ti t’ul=e*], we can see that the optimal output,
[b] in (38), violates none of the four constraints (30)-(33): the prosodic peak is aligned
with the right edge of the ιP, there is only a single prosodic peak, the prosodic prominence associated with =e in the input is preserved in the output, and every morpheme
in the input has a phonological exponent in the output. The association of input morphemes to phonological exponents, however, is many-to-one, as indicated by the subscripts on =e in input and output. Hence the optimal candidate, [b] (=36), violates the
Anti-Coalescence constraint, Uniformity (McCarthy & Prince 1995), as well as Linearity. Like Linearity, Uniformity is ranked below the other constraints shown.
(37) Unif(ormity): No element in the output has multiple correspondents in the input.
Tableau for (36)
[…ti ok’il=e*1 ti t’ul=e*2 ]ι
…det coyote-def det rabbit-def
a. […ti ok’il=é1 ti t’ul=é2 ]ι
b. + […ti ok’il ti t’ul=é1,2 ]ι
[…ti ok’il ti t’ul=é2 ]ι
d. […ti ok’il=e1 ti t’ul=é2 ]ι
[…ti ok’il=é1,2 ti t’ul]ι
[…ti ok’il ti t’ul]ι
g. […ti ok’il ti t’ul=é1 =é2 ]ι
Candidates not shown include variations on [g] in which one =e or the other does not
realize the prosodic prominence of the ιP, i.e., […ti t’ul=e1 -é2 ] and […ti t’ul-é1 =e2 ]. Both
violate IdentProm and the second one violates Alignι as well.
Judith Aissen
Some additional facts, not yet presented, show that RealizeMorph must be indexed
for particular morphemes and that the one which indexes =e is ranked below Alignι ,
Culminativity and IdentProm. The definite enclitic =e is not the only morpheme in
Zinacantec Tsotsil which is lexically associated with the prosodic peak of ιP. The other
is an epistemic particle, a’a, which Laughlin (1975) classifies as an “exclamation” and
translates indeed!, surely! certainly! of course!. a’a does not require licensing, though
statistically, it tends to occur in utterances with 1st and/or 2nd person arguments and is
likely cognate with a reduplicated form of the terminal clitic a’ ‘proximate’ in Yucatec.
Relevant here is that a’a occurs in the same position as =e, i.e., at the right edge of ιP,
with its second syllable functioning as the prosodic peak of ιP.
a. Ta’ajebal
li j-ve’el-tik
almost.cooked det exclam
‘Our meal certainly is about cooked.’ (Laughlin 1977: 285)
b. Ta j-ti’
lavi a’a.
icp erg.1-eat today exclam
‘Of course I’ll eat it today.’ (Laughlin 1977: 283)
c. Ik’-o
le’ a’a!
take-imp dem exclam
‘Take her!’ (Laughlin 1977: 126)
d. A li Pineda=e mas mas ts’akal
top det Pineda-def more more afterwards exclam
‘Pineda was later, of course.’ (Laughlin 1977: 116)
=e and a’a compete with one another, with priority given to realization of a’a. Thus =e
must be omitted when a’a occurs. (39a-c) contain various elements (underlined) that
otherwise require =e (see Table 1). Here though, a’a entirely precludes realization of =e.
As an epistemic operator, I assume that a’a occupies a position in the syntax; its exact
location cannot be determined since it is pronounced only at the right edge of ιP. Assuming that e* and a’a* can both be present in the input, one or the other must “disappear”.
Which is preserved is determined by the ranking of morpheme-specific RealizeM constraints. In Zinacantec Tsotsil, Realize(a’a*) ≫ Realize(=e*). The overall ranking of the
constraints under discussion then is shown in Figure 3.
5.2 Notes on the right periphery
I close this section by discussing the relation between the terminal elements =e and a’a,
and two other elements which “pile up” at the right periphery. The ordering of the four
is shown in (40):
12 Special clitics and the right periphery in Tsotsil
Figure 3: Constraint ranking
The particle un occurs in many of the examples cited above. No meaning (propositional
or otherwise) has yet been identified for it. Some speakers have the intuition that it
contributes some nuance of meaning to the sentence; others say that the sentence just
“sounds better” with it. un has a distribution similar to that of =e and a’a: like them, un
occurs at the right periphery of root sentences and of topics, and it can separate a matrix
clause from its extraposed complement. Also like them, it occurs in no other positions.
Unlike =e, it is not lexically licensed.
Aissen (1992) analyzed un as an enclitic which aligns with the right edge of ιP. While
it is true that un always occurs very near the right edge of ιP, it does not occur rightmost when any of the other elements in (40) is present. While it is not yet clear what
is responsible for its appearance and position, I assume that it is not lexically associated
with the prosodic peak in ιP and that its position is therefore not determined by IdentProm. For one thing, as observed in Skopeteas (2010), it does not coalesce with =e (nor
with a’a). One possibility is that it is present already at Spell-Out at the right edge of ιP.
It would then be present in the input to evaluations like those in (34) and (38), and the
constraint ranking in Figure 3 would position e and a’a to its right. Another possibility
is that un is introduced by the phonology for eurhythmic reasons, e.g., to improve the
prosodic structure of the utterance, perhaps at lower levels of the prosodic hierarchy. I
leave further development of these ideas for a later time.
The other element in (40) is che’e, which occurs only in the absolute final position.
che’e is a discourse particle which Laughlin (1975) translates as ‘then’ (roughly Spanish
xa li vo’on=e
cp-abs.1-go cl det pro.1sg-def then
‘Me, I went, then.’ (Laughlin 1977: 131)
che’e can co-occur with =e and when it does, the high boundary tone appears to be
realized on the last syllable of che’e, not on =e. It seems then to be a counterexample to
the descriptive generalization that =e is always the prosodic peak of the ιP in which it
Judith Aissen
A plausible scenario is that che’e is incorporated into the ιP which ends in =e after the
point at which the constraints discussed above have had an effect. Fleshing this out a
little, che’e might be syntactically adjoined to the root and mapped into its own ιP (like
topics and extraposed clauses). This ιP being however subminimal (two syllables, one
word), che’e is incorporated into the preceding ιP (on the tendency to avoid short ιP’s
or sequences of ιP’s of different length, see Nespor & Vogel 1986 and Dehé 2009). The
result here is to push =e back from the edge, and for the boundary tone to fall on the final
syllable of che’e. An account along these lines assumes that the constraints in Figure 3
apply within the domain of ιP’s that result from the initial prosodic parsing and do not
reapply at a later stage when prosodic restructuring of multiple ιP’s occurs. If they did,
=e would be reordered again, to the right of che’e. How such an account with its implied
serial optimization fits into the larger theory of the syntax-phonology interface remains
to be seen.
6 An historical scenario
Definite markers which close the phrase in which they are licensed are not uncommon
(Dryer 2013).13 It is plausible then that the definite enclitic =e in Z Tsotsil might, at an
earlier stage, have been the final element in the noun phrase, a position in which it would
not necessarily have functioned as the prosodic peak of ιP. Here I offer a suggestion for
how =e might have come to be associated with that peak, an association which now
sometimes forces it out of its licensing phrase.
The basic idea is simple: the syntax usually determines an utterance-final position for
the phrase which licenses =e. Hence even without intervention from the phonology, =e
would have found itself in most cases at the right edge of the utterance. As such, it would
become statistically associated with the prosodic peak of the ιP and this could have been
reanalyzed as a lexical property.
There are several reasons why the syntax usually puts the phrase which licenses =e
in utterance-final position. A number of them come down to the fact that certain grammatical relations in Tsotsil are almost always instantiated by definite noun phrases and
the syntax determines a position for these relations at the right edge of the utterance
anyway. These include especially subjects, possessors, and topics. The usual ordering of
these elements is shown in (42). Starting with topics, as we have already seen, the topic
precedes its associated clause and always constitutes its own ιP. As the final element in
the topic then, =e automatically falls at the right edge of ιP.
◦ Topic X
◦ V-O-S
◦ Possessum - Possessor
Basic word order in Tsotsil is usually described as VOS, with the subject in clause-final
position. Transitive subjects (as well as active intransitive ones) are almost always defi13
Languages with such markers include Wolof (Niger-Congo, Torrence 2013), Basque (Laka 1996), Angami
(Tibeto-Burman, Giridhar 1980, cited in Dryer 2013), and Gaahmg (Nilo-Saharan, Stirtz 2012).
12 Special clitics and the right periphery in Tsotsil
nite, so generally license =e. Unless the subject is followed by some other element (e.g.,
an adverb, a PP, an element in a matrix clause), =e again finds itself at the right edge of
ιP. Finally, Tsotsil being a head-initial language, the possessor follows its possessum, as
in (43).
ta [s-na
[li Xun=e]].
cp-abs.1-go p gen.3-house det Juan-def
‘I went to Juan’s house.’
Possessors too are almost always definite, and often end up as the final phrase in an
utterance. Here too, =e’s position at the right edge of ιP is determined by the syntax. In
all these cases then, =e is the last syllable in ιP, the position associated with the prosodic
Of course, the phrase which licenses =e does not always occur utterance-finally – if
it did, there would be no motivation for this paper. But in a fragment of written text
containing 156 instances of =e, there were only three in which that phrase did not occur
utterance-finally. In these cases, =e was separated from its licensing phrase, as in (12)-(17)
above. Thus, if it is true that the position of =e was originally determined syntactically,
it would nonetheless have had a statistical association with the phonological properties
that characterize the prosodic peak of ιP and reanalysis of this association as a lexical
property would have resulted in the situation we see today.
7 Conclusion
This paper has attempted to lay out the case for Z Tsotsil =e as a special clitic – one
whose surface position is not always a position it could have reached syntactically. If
this is correct, the phonology does something here other than select the prosodically
optimal position for =e from among the syntactically possible ones. It must achieve
the effect of moving =e within a prosodically-defined domain. In the analysis proposed
here, =e is not subject to an alignment constraint; rather, it ends up at the right edge of
ιP because it must function as the prosodic peak of ιP, and that peak is located at the
right edge of ιP. Complying with this requirement sometimes involves reordering the
enclitic over a fairly large distance. Since the reordering occurs in the phonology, it is
not subject to syntactic locality. It is, though, subject to prosodic locality, as =e always
remains within the ιP that contains its licensor (fn. 12).
Tsotsil =e thus appears to be different from the the second position clitics discussed in
Bošković (2000) and Bermúdez-Otero & Payne (2011), clitics which can reach their surface positions by syntactic means. The difference might be understood in terms of the
property which determines their surface position. The position of the second-position
clitics of Chamorro and Serbo-Croatian is determined by a prosodic alignment condition.
But prosodic constituency is introduced in the interface between syntax and phonology
and is therefore present before the phonology proper. The placement of second-position
clitics can therefore be determined prior to the phonology and without any involvement
Judith Aissen
of the phonology. On the other hand, if the analysis of Z Tsotsil =e suggested here is on
the right track, its position cannot be determined until the phonology proper, since it is
only in the phonology that the location of prosodic prominence within the ιP is fixed at
the right edge. In this light, the special clitic status of =e arises because the condition
which makes it “special” – which forces it out of its licensing phrase – references a purely
phonological property and not a prosodic edge.
completive aspect
definite terminal clitic
existential predicate
exclamatory particle
incompletive aspect
status suffix
Earlier versions of this paper were presented at UCSC, UC Berkeley, and at Exploring the
Interfaces III, McGill University. I would like to thank all those who have given me feedback on this material, especially a number of anonymous reviewers – their comments
were extremely helpful. I owe special thanks to Ryan Bennett, both for his comments on
an earlier version and for sustained discussion of various issues relevant to this material.
Ryan also prepared the pitch tracks included here. Needless to say I alone am responsible
for the content of the paper.
I am delighted to contribute a paper on clitics to a festschrift which honors Steve
Anderson. Steve’s work in this area, as in others, has been foundational. But on a more
personal note, it was at his suggestion in 1972 that I went to Mexico for the first time
to work on Tsotsil. Steve was thus directly responsible for the direction my career took
and I am glad to be able to thank him for that here.
Abels, Klaus. 2003. Successive cyclicity, anti-locality, and adposition stranding. University
of Connecticut PhD thesis.
Agbayani, Brian & Chris Golston. 2010. Phonological movement in Classical Greek. Language 86(1). 133–167.
Agbayani, Brian, Chris Golston & Toru Ishii. 2015. Syntactic and prosodic scrambling in
Japanese. Natural Language & Linguistic Theory 33. 47–77.
Aissen, Judith. 1992. Topic and focus in Mayan. Language 63(1). 43–80.
12 Special clitics and the right periphery in Tsotsil
Aissen, Judith. 1996. Pied piping, abstract agreement, and functional projections in
Tzotzil. Natural Language & Linguistic Theory 14. 447–491.
Alexiadou, Artemis, Tibor Kiss & Gereon Müller. 2012. Introduction. In Artemis Alexiadou, Tibor Kiss & Gereon Müller (eds.), Local modelling of non-local dependencies in
syntax, 1–48. Berlin: Walter de Gruyter.
An, Duk-Ho. 2007. Clauses in noncanonical positions at the syntax-phonology interface.
Syntax 10. 38–79.
Anderson, Stephen R. 2005. Aspects of the theory of clitics. Oxford: Oxford University
Bennett, Ryan. 2016. Mayan phonology. Language and Linguistics Compass 10(10). 469–
Bennett, Ryan, Emily Elfner & Jim McCloskey. 2015. Lightest to the right: An apparently
anomalous displacement in Irish. Linguistic Inquiry 47(2). 169–234.
Bermúdez-Otero, Ricardo & John Payne. 2011. There are no special clitics. In Alexandra
Galani, Glyn Hicks & George Tsoulas (eds.), Morphology and its interfaces, 57–96. Amsterdam: John Benjamins.
Bošković, Željko. 2000. Second position cliticization: Syntax and/or phonology? In Clitic
phenomena in European languages, 71–119. Amsterdam: John Benjamins.
Chung, Sandra. 2003. The syntax and prosody of weak pronouns in Chamorro. Linguistic
Inquiry 34(4). 547–599.
Cinque, Guglielmo. 2005. Deriving Greenberg’s universal 20 and its exceptions. Linguistic Inquiry 36(3). 315–332.
Cowan, Marion. 1969. Tzotzil grammar. Norman: Summer Institute of Linguistics of the
University of Oklahoma.
Dehé, Nicole. 2009. Clausal parentheticals, intonational phrasing, and prosodic theory.
Journal of Linguistics 45(3). 569–615.
Delgaty, Alfa Hurley & Agustín Ruíz Sánchez. 1978. Diccionario tzotzil de San Andrés con
variaciones dialectales. Mexico, D.F.: Instituto Lingüístico de Verano.
Downing, Bruce. 1970. Syntactic structure and phonological phrasing in English. University
of Texas at Austin PhD thesis.
Dryer, Matthew S. 2013. Definite articles. In Matthew S. Dryer & Martin Haspelmath
(eds.), The world atlas of language structures online. Leipzig: Max Planck Institute for
Evolutionary Anthropology.
Elfner, Emily. 2012. Syntax-prosody interactions in Irish. University of Massachusetts,
Amherst PhD thesis.
Frank, Robert, Paul Hagstrom & K. Vijay-Shanker. 2002. Roots, constituents and ccommand. In Artemis Alexiadou (ed.), Theoretical approaches to universals, 109–137.
Amsterdam: John Benjamins.
Giridhar, Puttushetra Puttuswamy. 1980. Angami grammar. Manasagangotri, Mysore:
Central Institute of Indian Languages.
Halpern, Aaron. 1995. On the placement and morphology of clitics. Stanford: Center for
the Study of Language (CSLI).
Judith Aissen
Haviland, John. 1981. Sk’op sotz’leb: el tzotzil de San Lorenzo Zinacantán. Mexico, D.F.:
Hayes, Bruce. 1995. Metrical stress theory: Principles and case studies. Chicago: University
of Chicago Press.
Henderson, Robert. 2012. Morphological alternations at the intonational phrase edge:
The case of K’ichee. Natural Language & Linguistic Theory 30. 741–787.
Kathol, Andreas. 2004. Linear syntax. Oxford University Press.
Kurisu, Kazutaka. 2001. The phonology of morpheme realization. University of California,
Santa Cruz PhD thesis.
Laka, Itziar. 1996. A brief grammar of Euskara, the Basque language. Leioa-Donostia: University of the Basque Country.
Laughlin, Robert. 1975. The great Tzotzil dictionary of San Lorenzo Zinacantán. Washington, D.C.: Smithsonian Institution Press.
Laughlin, Robert. 1977. Of cabbages and kings. Washington, D.C.: Smithsonian Institution
Laughlin, Robert. 1980. Of shoes and ships and sealing wax. Washington, D.C.: Smithsonian Institution Press.
McCarthy, John & Alan Prince. 1993. Generalized alignment. In Geert Booij & Jaap van
Marle (eds.), The yearbook of morphology 1993, 79–153. Dordrecht: Kluwer Academic
McCarthy, John J. & Alan Prince. 1995. Faithfulness and reduplicative identity. In
Jill Beckman, Laura Walsh Dickey & Suzanne Urbancyzk (eds.), University of Massachusetts occasional papers 18: Papers in Optimality Theory, 249–384. Amherst, MA:
Graduate Linguistic Student Association, UMass.
Miller, Philip. 1991. Clitics and constituents in phrase structure grammar. University of
Utrecht PhD dissertation.
Nespor, Marina & Irene Vogel. 1986. Prosodic phonology. Dordrecht: Foris.
Pesetsky, David & Esther Torrego. 2001. T-to-C movement: Causes and consequences. In
Michael Kenstowicz (ed.), Ken Hale: A life in language, 355–426. Cambridge, MA: MIT
Polian, Gilles. 2013. Gramática del tseltal de Oxchuc. México, D.F: CIESAS.
Prince, Alan & Paul Smolensky. 1993/2004. Optimality Theory: Constraint interaction in
generative grammar. Oxford: Blackwell Publishing.
Selkirk, Elisabeth. 2009. On clause and intonational phrase in Japanese: The syntactic
grounding of prosodic constituent structure. Gengo Kenkyu 136. 35–73.
Shklovsky, Kirill. 2011. Petalcingo Tseltal intonational prosody. In Kirill Shklovsky, Pedro Mateo Pedro & Jessica Coon (eds.), Proceedings of Formal Approaches to Mayan
Linguistics I, 209–220. Cambridge, MA: MITWPL.
Simpson, Andrew. 2005. Classifiers and DP structure in Southeast Asian languages. In
R. Kayne & G. Cinque (eds.), The Oxford handbook of comparative syntax, 806–838.
Oxford: Oxford University Press.
12 Special clitics and the right periphery in Tsotsil
Skopeteas, Stavros. 2010. Syntax-phonology interface and clitic placement in Mayan languages. In Vicenç Torrens, Linda Escobar, Anna Gavarró & Juncal Gutiérrez (eds.),
Movement and clitics, 307–331. Newcastle: Cambridge Scholars Publishing.
Stirtz, Timothy. 2012. A grammar of Gaahmg. University of Leiden PhD thesis.
Torrence, Harold. 2013. The clause structure of Wolof: Insights into the left periphery. Amsterdam: John Benjamins.
Zwicky, Arnold. 1977. Hierarchies of person. In Papers from the 13th regional meeting of
the Chicago Linguistic Society, 712–733. Chicago: Chicago Linguistic Society.
Zwicky, Arnold M. & Geoffrey K. Pullum. 1983. Cliticization vs. inflection: English N’T.
Language 59(3). 502–513.
Chapter 13
Another way around causatives in
Sandra Chung
University of California, Santa Cruz
In Anderson’s (1992) theory of a-morphous morphology, the traditional observation that
inflection occurs “outside of” derivation follows from the assumption that only lexically
complete stems can instantiate morphosyntactic representations. Anderson discusses an
apparent counterexample to the traditional observation that involves causative verbs and
number agreement in the Austronesian language Chamorro. Anderson defuses the apparent
counterexample by proposing, following Durie 1986, that Chamorro number agreement is a
derivational, rather than inflectional, process. I show that there is a different way of finessing
the issue that preserves the intuition that Chamorro number agreement is inflectional. This
alternative takes the causative ‘prefix’ to be a prosodically deficient verb, in the overall spirit
of Anderson 2005.
1 Introduction
In Anderson’s theory of a-morphous morphology, the traditional observation that inflection occurs “outside of” derivation follows from the assumption that only lexically complete stems can instantiate morphosyntactic representations. Anderson (1992: 127–128)
discusses an apparent counterexample to the traditional observation from Chamorro, an
Austronesian language of the Mariana Islands. Chamorro has causative verbs which,
according to previous accounts, are formed by attaching the prefix na’- to a verb or adjective (see e.g. Baker 1985; Gibson 1980, Safford 1904: 108, and Topping & Dungca 1973:
247–249). The point of interest is that na’- can attach to a verb or adjective that already
shows number agreement. Assuming that na’- is derivational but number agreement is
inflectional, the fact that na’- can occur “outside of” number agreement is problematic.
Anderson defuses the apparent counterexample by proposing, following Durie (1986:
364–365), that Chamorro number agreement is a derivational, rather than inflectional,
Here I explore a different way of finessing the issue, one that preserves the intuition
that Chamorro number agreement is inflectional. The key to this alternative is to take
Sandra Chung. 2017. Another way around causatives in Chamorro. In Claire Bowern,
Laurence Horn & Raffaella Zanuttini (eds.), On looking into words (and beyond), 263–
288. Berlin: Language Science Press. DOI:10.5281/zenodo.495450
Sandra Chung
the causative “prefix” to be a prosodically deficient verb, in the spirit of Anderson’s 2005
approach to clitics as phrasal affixes. Chamorro has a small class of prosodically deficient verbs that are instances of Zwicky’s 1977 bound words. These forms have the morphosyntax of verbs, but cannot serve as phonological words on their own. Instead, they
must remedy their prosodic deficiency by undergoing stray adjunction to the phonological word to their immediate right, which is always the first phonological word of their
I show that much of the distinctive profile of Chamorro causatives is accounted for if
the causative na’ is treated as a prosodically deficient verb that selects a vP complement.
Moreover, once this route is taken, Chamorro causatives no longer pose a threat to the
“outside-inside” order of inflection and derivation, even if number agreement is inflectional. This is because the causative na’ that can appear “outside of” number agreement
is not, in fact, derivational morphology, but rather the prosodically deficient content of
a syntactic verb.
§2 of this paper gives a mini-introduction to the morphosyntax of Chamorro clauses.
§3 presents the basics of causatives and their interaction with the language’s two types
of subject-verb agreement. §4 looks closely at Durie’s 1986 evidence that Chamorro
number agreement is derivational and concludes that it is not decisive. Then, §5 gives an
overview of Chamorro’s prosodically deficient verbs. §6 proposes that the causative na’
is a prosodically deficient verb and explores some positive consequences of this proposal.
§7 handles some challenges, and §8 concludes.
2 Chamorro Morphosyntax in Brief
Chamorro is a head-initial language that allows predicates of all major category types
and a range of null arguments. When the predicate is a verb or adjective, the default word
order of the clause is Predicate Subject Object Other, but the order of arguments and
adjuncts after the predicate is flexible (see Chung 1998 and the references cited there).1
a. Ha baba si Antonio i petta.
p.agr open unm Antonio the door.
‘Antonio opened the door.’
b. Kumåti i neni sa’
ni sanye’yi’.
n.agr.cry the baby because n.agr.afraid obl spider
‘The child cried because she’s afraid of the spider.’ (CD, entry for sanye’yi’)
DPs are case-marked with a proclitic that occurs to their immediate left. There are
three morphological cases: unmarked, local, and oblique. Subjects, direct objects, possessors, predicate nominals, the objects of most overt prepositions, and DPs that occupy
Most of the Chamorro examples cited here are from the CD database, which consists of some 30,000 sentences constructed by Chamorros in the CNMI as illustrative examples for the revised Chamorro-English
dictionary. Other examples are from published sources listed in the references; unattributed examples are
from my fieldwork.
13 Another way around causatives in Chamorro
topic or focus position occur in the unmarked case, which is overtly realized only when
the DP is a proper name. Otherwise, DPs that denote locations or goals occur in the local
case; most other types of DPs occur in the oblique case.
(2) Ma rikuknisa si Esthera ni finatton-ña gi hunta.
p.agr recognize unm Esther obl arrival-poss loc meeting
‘They acknowledged Esther for her coming to the meeting.’ (CD, entry for
Predicates that are verbs or adjectives show subject-verb agreement via forms that
also indicate mood (realis vs. irrealis) and are sensitive to transitivity. There are two
types of subject-verb agreement. Person-and-number agreement (glossed p.agr in the
examples) is realized via forms that could be analyzed as proclitics or prefixes, but are
written as separate words in the Chamorro orthography; see the paradigm in (3).2 In
the realis mood, this type of agreement is found only on transitive verbs; in the irrealis
mood, it is found on all verbs and adjectives.
Person-and-Number Agreement
(bai) hu / bai
(bai) in
u (intr) / uma (tr)
Number agreement (glossed n.agr) is realized via a prefix or infix; see the paradigm
in (4). This type of agreement is found only on intransitive verbs and adjectives.3
(4) Number Agreement
-um- / —
Chamorro has various standard and nonstandard orthographies (see Chung 1998: Appendix A). The orthography used here, which was officially adopted by the CNMI legislature in 2010, differs in small ways
from the transcription used in Chung 1998, and more substantially from earlier spelling systems, including
the official orthography on Guam.
The choice between the two realizations of realis singular number agreement is determined lexically. Generally, -um- is used for event predicates, as well as for state predicates in the inchoative aspect; otherwise,
the agreement is generally unrealized for state predicates. But there are exceptions. The realizations of
plural number agreement have a final /n/ that undergoes the alternation known as nasal substitution.
Sandra Chung
Notice that dual is aligned with plural for the purposes of person-and-number agreement, but with singular for the purposes of number agreement. This will become important later.
Both types of agreement are the default realizations of subject-verb agreement for
predicates of the relevant type, and fully productive; e.g. they can be added to recently
borrowed words (as in 5a), even when the borrowings are creative or innovative (as in
a. Man-meeting
ham gi Lunis.
n.agr-have.meeting we loc Monday
‘We had a meeting on Monday.’ (CNMI Senate session SJ 17–22: 20)
b. Bai hu “love-mark” i kurason-mu.
p.agr love-mark the heart-poss
‘I will “love-mark” your heart.’ (EM 60)
Finally, the two types of agreement have overlapping distributions. Transitive verbs
show only person-and-number agreement (see 1a, 2, 5b and 6a); intransitive predicates
in the realis mood show only number agreement (see 1b, 5a, and 6b); but intransitive
predicates in the irrealis mood show both. Note that when the two types of agreement
co-occur, person-and-number agreement occurs “outside of” – i.e. to the left of – number
agreement (see 6c).
a. Hu afuetsas gui’ para u
atan yu’.
p.agr compel her fut p.agr me
‘I compelled her to (lit. that she would) look at me.’ (CD, entry for afuetsas)
b. Durånti-n i tinaitai, bula mang-åti.
during-l the prayer many n.agr-cry.
‘During the prayer, many cried.’ (CD, entry for durånti)
c. Ti para u
fang-åti i famalåo’an.
not fut p.agr n.agr-cry the women
‘The women are not going to cry.’
With this much in place, let us now zero in on causatives.
3 Causatives
Previous accounts describe Chamorro causatives as formed by adding the prefix na’- to
a verb or adjective (see e.g. Baker 1985; Gibson 1980; Safford 1904; Topping & Dungca
1973). This process creates a derived transitive verb with an additional argument, which
denotes the causer. The causer argument is realized as the subject of the causative; the
subject of the original predicate (henceforth the inner predicate) is realized as the direct
object of the causative; and the direct object of the inner predicate, if any, is realized as
13 Another way around causatives in Chamorro
an oblique (see Gibson 1980). To illustrate, the causatives na’baba ‘make open’, na’kåti
‘make cry’, and na’ma’å’ñao ‘make afraid, frighten’ are derived, respectively, from the
transitive verb baba ‘open’ (cf. 1a), the intransitive verb kåti ‘cry’ (cf. the first clause of
1b), and the adjective ma’å’ñao ‘afraid’ (cf. the second clause of 1b).
a. In
na’-baba si Antonio ni petta.
p.agr caus-open unm Antonio obl door
‘We made Antonio open the door.’
b. Ha na’-kåti si Gene i lahi-ña anai ha
p.agr caus-cry unm Gene the son-poss when p.agr scold
‘Gene made his son cry when he scolded him.’ (CD, entry for kåti)
c. Un na’-ma’å’ñao yu’ ni taklalo’-mu.
p.agr caus-afraid me obl great.anger-poss
‘You made me afraid with your great anger.’
Gibson’s 1980 investigation of the syntax of Chamorro causatives established three
points that will be in the spotlight here. First, causatives have the morphosyntax of the
language’s transitive verbs (Gibson 1980: 86–91). Like other transitive verbs, they can
occur in the passive.4
a. Ma-na’-gimin
i patgun åmut
ti dinanchi.
n.agr.pass-caus-drink the child medicine comp not n.agr.right
‘The child was made to drink medicine that was not right.’ (CD, entry for
b. Kulan nina’-ma’å’ñao
i biha
nu esti na klåsi-n tinanum.
sort.of n.agr.pass.caus-afraid the old.lady obl this l type-l plant
‘The old lady was kind of made afraid by this type of plant.’ (MAK 2)
They can also occur in the antipassive.5
(9) Mu-nana’-gupu
papaloti si Juanito gi kantu-n tåsi.
n.agr-ap.caus-fly.prog kite
unm Juanito loc edge-l ocean
‘Juanito is flying a kite (lit. making a kite fly) by the seashore.’ (CD, entry for
And they can be used to create reciprocals – derived intransitive verbs, formed with
the stressed prefix á-, which are Chamorro’s primary means of expressing reciprocal
Passive verbs are formed with the infix -in- or the prefix ma-. The choice between -in- and ma- is determined
primarily by the number of the passive agent: -in- appears when the agent is singular, ma- when the agent
is dual/plural or implicit (see Chung 1998: 38, note 8).
Antipassive verbs are usually formed with the prefix man-/fan-. However, some transitive verbs have
suppletive antipassive forms (e.g. the antipassive of kånnu’ ‘eat’ is chotchu); others have antipassive forms
identical to their transitive forms (e.g. gimin ‘drink’). The antipassive of a causative is formed by shifting
primary stress to the causative prefix.
Sandra Chung
(10) Kao um-á-na’-patcha
hamyu ni feggun?
q n.agr-recip-caus-touch obl stove
‘Did you two make each other touch the stove?’
Second, causatives can be derived from verbs that are morphologically complex (see
Gibson 1980: 114–121). The causatives in the examples in (11) are derived from verbs –
surrounded by brackets – that are passive (11a–11b), antipassive (11c), and reciprocal (11d).
a. In
na’-[ma-baba] as Antonio.
p.agr caus-pass-open obl Antonio
‘We made it be opened by Antonio.’
b. Bai na’-[sinaolak] hao nu i ma’estra.
p.agr caus-pass.spank you obl the teacher
‘I will let you be spanked by the teacher.’ (CD, entry for sinaolak)
c. I bakulu-hu
ha-na’-[fang-gånna] yu’.
the shooter.marble-poss p.agr-caus-ap-win me
‘My shooter marble made me win.’ (CD, entry for båkulu)
d. Ma na’-[á-dispatta]
i dos tåotao ni
p.agr caus-recip-separate the two person comp n.agr.fight
‘They separated (lit. caused to separate from each other) the two people who
were fighting.’ (CD, entry for na’ádispatta)
As these observations might lead one to expect, causatives derived from morphologically complex verbs can themselves occur in the passive, antipassive, or reciprocal.6 The
verbs in boldface in (12) are the passive of a causative derived from a passive verb (in
12a) and the passive of a causative derived from an antipassive verb (in 12b).
a. … yan maseha håyi malago’-ña
i Lahi-ña para u
and ever
who wh.want-poss the son-poss fut p.agr
nina’-[ma-tungu’] Gui’.
pass.caus-pass-know he
‘…and whoever his Son wants to cause Him (lit. that He be caused) to be
known by.’ (NT 124)
b. Nina’-[fañ-otsut]
anai ma-nå’i
måolik na kunseha.
n.agr.pass.caus-ap-repent when n.agr.pass-give good l advice
‘She repented (lit. was caused to repent) when she was given good advice.’
(CD, entry for na’fañotsut)
Although it is possible in principle for causatives formed from a verb in any voice to occur in any voice,
the naturally occurring data suggest that some combinations are more frequent than others. When the
causative is active transitive or passive, the inner predicate can be active (transitive or intransitive), passive,
antipassive, or reciprocal. When the causative is antipassive or reciprocal, the inner predicate is most often
active (transitive or intransitive).
13 Another way around causatives in Chamorro
A causative can even be derived from the passive of a causative, as (13) shows.
(13) Si Josephine ha
i atgoya gi gui’eng-ña.
unm Josephine p.agr caus-pass-caus-go.away the nose.ring loc nose-poss
‘Josephine had her nose ring removed (lit. caused the nose ring to be caused to go
away).’ (CD, entry for atgoya)
Third, the inner predicate – the verb or adjective from which a causative is derived
– does not show person-number agreement. But, surprisingly, the inner predicate does
show number agreement (see Gibson 1980: 112–114). Inner predicates that are intransitive
agree with the DP that would have been their subject via irrealis number agreement,
which is unrealized in the singular/dual, but spelled out as the prefix fan- in the plural.
This number agreement is not realized on the inner predicates in (11–13), because the
DPs that would have been their subjects are singular/dual (e.g. the null pronoun ‘it’ in
(11a), hao ‘you (sg.)’ in (11b), yu’ ‘me’ in (11c)), but it is overt on the inner predicates
in (14), because the DPs that would have been their subjects are plural. (Note that the
inner predicates in (14) are clearly not agreeing with the subject of the causative, which
is singular.)
a. Hu na’-[fang-gupu] i petbus.
p.agr caus-n.agr-fly the dust
‘I made the (particles of) dust fly around.’ (CD, entry for na’gupu)
b. Ha na’-[fan-luhan] ham.
p.agr caus-n.agr-afraid us
‘[The wind] scared us (lit. made us afraid).’ (CD, entry for diripenti)
c. Ha na’-[fan-ma-kotti] i guåtdia, ya ha
p.agr caus-n.agr-pass-try the guard and p.agr caus-n.agr-pass-kill
‘He had the guards brought to trial, and had them killed.’ (NT 235)
d. I abisu ha
i taotao na …
the alarm p.agr caus-n.agr-ap-know.prog the person that
‘The alarm is letting the people know that…[the typhoon is close]. (CD, entry
for abisu)
Baker (1985) used the relative order of the plural fan- with respect to the causative
and passive affixes to argue for the Mirror Principle. As he observed, “clear examples of
agreement morphemes that can appear intermixed with GF-rule morphemes seem quite
unusual” (Baker 1985: 386). What matters here is that the plural fan- in the examples
in (14) occurs “inside of” – i.e. to the right of – the causative na’-. Assuming that fanis inflectional but na’- is derivational, this ordering appears to counterexemplify the
traditional claim that inflection always occurs “outside of” derivation.
Sandra Chung
4 Number agreement revisited
A natural question to raise at this point is whether Chamorro number agreement might
be derivational as well.
4.1 Is it derivational?
As Anderson (1992: 127–128) observes, this question is answered in the affirmative by
Durie (1986), who contends that across languages, verbal number – whether realized by
stem suppletion or productive affixation – is “selectional concord” (i.e. derivational) as
opposed to “agreement”. Durie’s evidence for this claim comes from various languages,
including Chamorro. In the suppletion cases he examines (in e.g. Huichol), verbal number is sensitive to semantic roles like patient or affected participant, not to syntactic
relations like subject. Chamorro number agreement does not conform to this pattern,
but instead cross-references the (surface) subject regardless of semantic role; this is one
way that it behaves like a paradigmatic case of agreement. Still, Durie argues that number agreement in Chamorro is “inherent verbal Number morphology” (Durie 1986: 364)
whereas person-and-number agreement is inflectional, on the basis of the following:
• Number agreement distinguishes plural from nonplural (i.e. plural from singular/dual), but the number feature on nouns and pronouns distinguishes singular
from nonsingular (i.e. singular from dual/plural), so “[t]here is no [±plural] feature
for the verb to agree with” (Durie 1986: 364).
• Number agreement can have an overt pronoun as antecedent, whereas personand-number agreement cannot.
• Number agreement appears in infinitives, imperatives, and attributive modifiers,
whereas person-and-number agreement does not.
• Number agreement is preserved in lexical derivations, such as causatives (see
above), whereas person-and-number agreement is not.
These may look like good reasons for classifying number agreement as derivational
– a move that would make it unsurprising in the extreme that the plural fan- can occur
“inside of” the causative na’-. But further examination suggests a more equivocal picture.
4.2 A second look
Consider, to begin with, the claim that Chamorro nouns and pronouns have a different number feature than what is registered by number agreement. The specific claim
is that nouns and pronouns employ the feature [±singular] – they distinguish singular
from dual/plural – whereas number agreement employs the feature [±plural] – it distinguishes singular/dual from plural (see the paradigm in (4)). Assuming that inflectional
morphology is the spell-out of syntactic features, the disconnect between these features
13 Another way around causatives in Chamorro
might seem to pose an insuperable problem for the view that number agreement is inflectional (but see below).
Overt pronouns in Chamorro do indeed employ the feature [±singular] – they distinguish singular from dual/plural, as observed explicitly by e.g. Safford (1903: 308). The
second person independent pronouns hågu and hamyu, for instance, differ in that hågu
refers to just one addressee, while hamyu refers to two or more addressees. The other
overt pronouns are similar. It is less obvious how number is handled in nouns, because
most Chamorro nouns do not show obligatory number inflection. Just a handful of nouns,
listed in (15), are inflected obligatorily, and somewhat irregularly, for number.
‘man, son’
The noun che’lu has separate forms for singular, dual, and plural. The other nouns
have forms which are usually termed “singular” and “plural” (e.g. Safford 1903: 302–304,
Topping & Dungca 1973: 325), but actually distinguish singular/dual from plural. That
is, they employ the feature [±plural]. The examples in (16) reveal that when these nouns
refer to just two individuals, they are realized in the singular/dual form, not the plural
a. Um-iskuekuela
i dos påtgun sanlagu. the two child continental.US
‘The two children are attending school in the continental U.S.’ (CD, entry for
b. Dos na palåo’an u
two l woman p.agr n.agr.ap-grind.prog grain
‘Two women will be grinding grain.’ (NT 48)
The claim that the nouns in (15b) align dual with singular is supported by naturally
occurring data.7 There are 30 instances in the CD database, and 23 instances in the first
150 pages of the Chamorro New Testament (NT), of these nouns occurring in explicitly
dual DPs – DPs whose noun is preceded by the numeral dos ‘two’. In 51 out of the
combined 53 instances, the noun occurs in the singular/dual form.
It is now clear that Chamorro pronouns employ the feature [±singular], but obligatorily inflected nouns employ the feature [±plural] or – in the case of che’lu – both features.
Native speakers’ judgements trend in the same direction, but are more forgiving. For instance, when asked
which of the following two forms she would use to refer to two children, one speaker commented that i
dos påtgun ‘the two children’ (with the singular/dual form of the noun) was better for her, but that i dos
famagu’un (with the plural form of the noun) “will be understood in most circumstances”.
Sandra Chung
This makes it reasonable to suppose that Chamorro DPs are specified for [±singular] and
[±plural], even though in the vast majority of cases, these features have no DP-internal
realization. But then the way that number is handled by the agreement system is compatible with the idea that both types of agreement are inflectional. Person-and-number
agreement simply registers one of the number features (namely, [±singular]), while number agreement registers the other ([±plural])
I now turn to Durie’s other evidence that number agreement is derivational. It consists
of the following:
- Number agreement can have an overt pronoun as antecedent, but person-and-number agreement cannot. (The only pronouns that can antecede person-and-number agreement are null pronouns; see also Chung 1998: 30–31.) Durie takes these facts, which
are illustrated in (17), to show that person-and-number agreement is “anaphoric”, but
number agreement is not.
a. Yayas
n.agr.tired s/he
‘S/he is tired.’
b. Ha fåhan (*gui’) i lepblu.
p.agr buy s/he the book
‘S/he bought the book.’
Now, the contrast in (17) could ultimately reflect a difference between derivation and
inflection. But it is equally likely that it flows from some linguistic notion of “efficiency”
or “brevity” (cf. Grice) plus the featural content of the two types of agreement. Personand-number agreement encodes exactly the same features as Chamorro pronouns –
namely, person features and [±singular] – so a ban that prevents this type of agreement from being anteceded by an overt pronoun contributes to the goal of minimizing
redundancy. A comparable ban on number agreement would have no rationale, because
number agreement encodes a different feature – [±plural].
- Number agreement appears in infinitives, imperatives, and attributive modifiers, but
person-and-number agreement does not. Consider the imperative in (18).
(*En) Fan-man-hokka’ sa’
pineddung mångga gi egga’an.
p.agr n.agr-ap-pick because n.agr.many fallen.l
mango loc morning
‘Go and do some picking, because there were many fallen mangos in the
morning.’ (CD, entry for poddung)
To the extent that this observation is valid,8 it could bear on the contrast between
derivation and inflection, but other explanations are possible. Suppose, for instance, that
number agreement realizes a feature of small v, whereas person-and-number agreement
In conjoined imperatives, the leftmost imperative verb does not show person-and-number agreement, but
verbs in subsequent conjuncts generally show irrealis person-and-number agreement as well as number
agreement (if applicable). The embedded “clause” in restructuring constructions can either be inflected like
an infinitive or show realis person-and-number agreement; see 6.2.
13 Another way around causatives in Chamorro
realizes features of T. Then number agreement would be expected to appear in infinitives
and imperatives, because these constructions are at least vPs; there might be no similar
expectations for person-and-number agreement. I will adopt a version of this approach
below. As for attributive modifiers, it should be noted that Chamorro allows relative
clauses to precede or follow the head NP; it also allows relative clauses whose head NP
is null (see Borja, Chung & Wagers 2015). The attributive modifiers that show number
agreement can straightforwardly be analyzed as predicates of one or another of these
relative clause types.
- Finally, number agreement is claimed to be preserved in lexical derivations, such
as causatives and what Durie calls “nominal derivatives”. Causatives are, of course, the
focus of investigation here. The “nominal derivatives” are not, in fact, derived nouns
but rather relative clauses whose head NP is null. Two of Durie’s examples are given
below, with the spelling normalized. In these constructions, the word that shows number
agreement is the verb of the relative clause, which happens to be intransitive.
a. i humånao
the n.agr.go
‘the (one) who went’ (translated by Durie as ‘the goer’)
b. i man-hånao
i n.agr-go
‘the (ones) who went’ (translated by Durie as ‘the goers (> 2)’)
Notice that when the verb of the relative clause is transitive, it can show person-andnumber agreement; see the relative clauses in brackets below.9 This too is expected if
these constructions are relative clauses.
a. Abånsa [i un chochogui].
advance the p.agr do.prog
‘Go forward with the (thing) which you are doing.’ (CD, entry for abånsa)
b. Hu angokku na para un cho’gui [i hu faisin hao].
p.agr trust
comp fut p.agr do
the p.agr ask you
‘I trust that you will do the (thing) which I ask you.’ (CD, entry for angokku)
In the end, the evidence cited by Durie provides no firm basis for classifying number
agreement as derivational or inflectional. But then we are back to the original conundrum: why can the plural fan- occur “inside of” the causative na’-? I propose to answer
this question by analyzing the causative na’ not as a derivational prefix, but as a prosodically deficient verb.
5 Prosodically deficient verbs
The proposal to analyze the causative na’ as a prosodically deficient verb assimilates it to
a very small class of frequently used Chamorro verbs. This class contains the intransitive
The verb of the relative clause can also show wh-agreement, but that is irrelevant here.
Sandra Chung
verb malak/falak ‘go to, head to, depart for’ and the transitive verb fa’ ‘pretend’.10 Both
verbs are clearly the content of lexical categories; they are not derivational prefixes. Like
other verbs, they serve as the predicates of clauses, show subject-verb agreement, are
inflected for mood and aspect, and so on. More significantly for our purposes, they select
a functional projection as their complement.
Malak/falak ‘go to’ selects a DP that is linked to its goal argument. This DP, which is
bracketed in (21), can include determiners (see 21a) and modifiers (21a–21b); it can also
consist of a place name (21c) or an interrogative pronoun (21d). This range of expansions
reveals that syntactic incorporation, however analyzed, is not involved.
a. Man-malak [i Pala na kasinu] ham. the Pala l casino we
‘We went to the Pala casino.’ (CD, entry for kasinu)
b. Ti ya-hu malak
[ottru tånu’].
not like-poss other land
‘I don’t like to go to other places.’ (CD, entry for gåstu)
c. Yanggin gaigi
hao Saipan ya para un falak
[Tinian], siempri
if you Saipan and fut p.agr Tinian indeed
humånao hao luchan.
n.agr.go you south
‘If you are on Saipan and traveling to Tinian, you will have to go south.’ (CD,
entry for luchan)
d. Malak
[månu] hao nigap? where? you yesterday
‘Where did you go yesterday?’ (CD, entry for malak)
Fa’ ‘pretend’ selects a finite realis TP complement. This embedded TP can have a
predicate of any major category type, and when the predicate is a verb or adjective, it
shows subject-verb agreement, as expected.
a. In
[in tingu’ i ti un tungu’].
p.agr pretend p.agr know the not p.agr know
‘We (excl.) pretend we know what you don’t know.’ (from a conference
I represent these verbs without dashes in order to highlight the fact that they are not prefixes. Note that
malak/falak is an m/f verb; its initial consonant is realized as /m/ in the realis mood or when preceded by
plural number agreement, but as /f/ otherwise. Fa’ is, confusingly, homophonous with a prefix fa’- that
creates derived verbs meaning ‘make (into, with)’. This prefix attaches productively to nouns (e.g. fa’hånum
‘liquefy’, from hånum ‘water, liquid’; fa’denni’ ‘prepare with hot sauce’, from donni’ ‘hot pepper’), and less
productively to adjectives (e.g. fa’baba ‘deceive’, from båba ‘bad’; fa’tinas ‘make’, from tunas ‘straight’).
The verb fa’ ‘pretend’ and the derivational prefix fa’- are treated as the same affix by Topping & Dungca
(1973: 176–77).
13 Another way around causatives in Chamorro
b. Ma tutuhun ma fa’
[man-kubåtdi siha].
p.agr begin p.agr pretend n.agr-cowardly they
‘They began to pretend that they were afraid.’ (NT 343)
c. Ha fa’
p.agr pretend n.agr.extremely-strong he
‘He pretended to be extremely strong.’
d. Ha fa’
[i anghit gui’] si Juan sa’
p.agr pretend the angel he unm Juan because the
John is acting like an angel (lit. pretending he is the angel) because his
girlfriend is here’ (CD, entry for ånghit)
The distinctive property of these verbs is that they are prosodically deficient; they are
not phonological words and cannot bear primary stress. They remedy this deficiency by
undergoing stray adjunction to the phonological word to their immediate right, which
(in Chamorro) is always the first phonological word of their complement.11 In (21c), for
instance, stray adjunction attaches falak to the phonological word Tinian (as shown in
(23a)); in (22a), it attaches fa’ to the phonological word in tingu’, which itself consists
of an agreement proclitic adjoined to the phonological word tingu’ ‘know’ (as shown in
Morphophonological processes which affect verbs, but whose domain is the phonological word, cannot affect a prosodically deficient verb directly. Instead, they must target
the phonological word that immediately dominates it. In Chamorro, for instance, the
progressive aspect is realized via reduplication of the primarily stressed CV of the predicate. When malak/falak or fa’ occurs in the progressive, the CV that is reduplicated is
the primarily stressed CV of the phonological word that immediately dominates them
In prosodic theory, stray adjunction is the operation that incorporates elements that are not parsed as
prosodic units at a given level of prosodic structure into an adjacent prosodic unit at that level; see e.g.
Anderson 2005: 13. The text assumes that in the cases under discussion, stray adjunction literally produces
an adjunction structure. As Nick Kalivoda observes, another possibility is that a prosodically deficient verb
simply becomes a daughter of the phonological word to its immediate right.
Sandra Chung
(which is also the primarily stressed CV of the phonological word to which they are
adjoined). See (24).
a. Siempri [malak
i tetenda] yu’.
indeed the store.prog I
‘I will definitely be going to the store.’
b. Ha [fa’
ha’ gui’.
p.agr pretend n.agr.stupid.prog emp he
‘He is just pretending that he is stupid.’ (CD, entry for mudoru)
The same holds true for other processes that are sensitive to prosodic structure. Among
the overt pronouns of Chamorro are a set of weak pronouns which are second position
clitics (e.g. yu’ ‘I’, hao ‘you (sg.)’). These weak pronouns occur right after the first phonological phrase of the intonational phrase corresponding to their clause (see Chung 2003).
Because most Chamorro clauses have predicates that are verbs or adjectives, and most
verbs or adjectives are phonological words that project a phonological phrase of their
own, a weak pronoun is usually positioned right after them (see e.g. 21c). But when the
verb is prosodically deficient, a weak pronoun is – as expected – positioned right after
the phonological word (and phonological phrase) dominating it. The relevant phonological word is enclosed in brackets below.
a. Tåtnai [malak
Luta] yu’.
never Luta I
‘I’ve never been to Rota.’ (CD, entry for tåtnai)
b. [Ha fa’
gui’ na tåotao.
p.agr pretend n.agr.very-good he l person
‘He pretended to be a very good person.’ (CD, entry for fa’)
6 Causative na’ as a prosodically deficient verb
The preceding should be enough to suggest why it would be helpful to reanalyze the
causative na’- as a prosodically deficient verb. Then the exuberance of its interplay with
voice, agreement, and the like can be attributed to the fact that it combines morphosyntactically with the material on its left, but merely prosodically with the material on its
6.1 Proposal
I propose to flesh out the details of this reanalysis as follows. Suppose that instead of a
causative prefix na’-, Chamorro has a prosodically deficient verb na’ ‘make, let, cause’,
which selects a small clause complement – specifically, a vP complement. In Chamorro,
small v selects a complement that is VP or AP, so the verb na’ will occur in syntactic
structures of the type shown in (26) (with specifiers omitted for convenience).
13 Another way around causatives in Chamorro
The V or A that heads the embedded VP or AP in (26) corresponds to what was referred to earlier as the inner predicate. Because the inner predicate has small v in its
functional layer, and small v is responsible for voice, the inner predicate can be a verb
that is passive, antipassive, or reciprocal. At the same time, when the verb na’ occurs as
the predicate of a finite clause, it will project its own small v (not represented in (26)), so
it can independently be passive, antipassive, or reciprocal. This will account for much of
the exuberant interplay that causatives exhibit.
What about subject-verb agreement? In Chamorro, person-and-number agreement is
always realized to the left of number agreement. This makes it reasonable to suppose that
the two types of agreement spell out features of different functional heads, where the
head whose features are spelled out by person-and-number agreement is the higher of
the two. Now, word order aside, finite clauses in Chamorro have a familiar architecture
in which the functional layer of the clause contains (at least) T and small v (see Chung
1998; 2004). Let us assume, then, that T is specified for finiteness, mood, aspect, and the
person and number of the DP in its specifier (= the subject). The relevant number feature
here is, of course, [±singular]. These features of T are spelled out by person-and-number
agreement when the predicate is transitive or the mood is irrealis; see (3). Let us make
the further, more interesting assumption that small v is specified for the number of the
DP in its specifier via the feature [±plural]. This feature of v is spelled out by number
agreement when the predicate is intransitive; see (4).12 In the finite clauses of interest
here, T has a vP complement, the DP in vP’s specifier raises to the specifier of T, and
number agreement spells out some features of T (finiteness and mood) as well as the
number feature of small v. The mechanisms responsible for the multiple exponence of
finiteness and mood are irrelevant here. What matters is that in structures in which vP
is the complement of the verb na’, number agreement is spelled out with “irrealis” forms:
as the prefix fan- when the DP in small v’s specifier is [+plural], and with no realization
Let us now turn to the prosody. The verb na’ is prosodically deficient, so in the
prosodic structure corresponding to (26) it will undergo stray adjunction to the phonological word to its immediate right, which is always the first phonological word of its
complement. Assuming – crucially – that the word order of the small clause complement
has already been determined, this phonological word will be the content of a verb or adjective. The verb or adjective may be morphologically complex and may begin with the
A reviewer asks how transitivity is folded into the picture. I assume that T’s features are spelled out as
person-and-number agreement when T shares features with transitive small v – a small v that assigns
abstract Case. Small v’s number feature is spelled out as number agreement when small v does not assign
abstract Case. See 7.2.
Sandra Chung
plural fan-. In other words, stray adjunction will lead to one of the outcomes schematized
in (27).
Overall, this proposal gives a remarkably successful account of the morphosyntactic profile of Chamorro causatives presented in §3. Causatives have the morphosyntax
of transitive verbs (see (8–10)) because na’ is, in fact, a transitive verb. The prosodic
deficiency of this verb makes it appear to be a prefix – and therefore derivational morphology – but the appearance is illusory. Like other prosodically deficient verbs, the verb
na’ selects a complement that is a functional projection – namely, vP – and undergoes
stray adjunction to the phonological word to its immediate right, which is always the
first phonological word of its complement. For independent reasons, this word is always
the content of a verb or adjective (see (7)). The verb or adjective projects a vP in the
syntax, so it is inflected for number agreement (see (14)) and can be morphologically
complex (see (11–12)). Moreover, the claim that na’ is a verb, as opposed to derivational
morphology, makes it unsurprising that it can be the content of both the main verb and
an embedded verb in recursive structures like (13).
Note, finally, that the proposal is consistent with the way that na’ interacts with
morphophonological processes whose domain is the phonological word or phonological phrase. When na’ occurs in the progressive aspect, the CV that is reduplicated is the
primarily stressed CV of the phonological word that immediately dominates it (see Gibson 1980: 79–81).13 (For consistency, I continue to use the parsing and glossing conventions adopted earlier for causatives, even though na’ is now analyzed as a prosodically
deficient verb.)
a. Esta
ni malangu-ña.
already n.agr.pass.caus-speak.incoherently.prog obl sickness-poss
‘Her sickness is making her speak incoherently.’ (CD, entry for chátkuentus)
b. Hu ripåra na un [na’-malilisia]
mampus i palabråk-ku.
p.agr notice comp p.agr caus-malicious.prog too.much the word-poss
‘I noticed that you really are making my words malicious.’ (CD, entry for
Further, weak pronouns are positioned not immediately after na’, but right after the
phonological word (and phonological phrase) that dominates it.
The progressive aspect in these examples must be interpreted as affecting the causative na’; it cannot be
interpreted as affecting the inner predicate of the causative. See especially (28b).
13 Another way around causatives in Chamorro
(29) [Man-na’-hanao] ham åbiu
para i man-disgrasiåo.
n.agr-ap.caus-go we support for the n.agr-in.accident
‘We sent help for those involved in that accident.’ (CD, entry for åbiu)
This is what we expect from a prosodically deficient verb; see §4.
6.2 Consequences
If this new approach turns out to be correct, Chamorro causatives no longer provide a
counterexample to the traditional observation that inflection is “outside of” derivation.
Instead, the causative na’ is a prosodically deficient verb, and its relative order with
respect to morphology which it happens to be prosodically attached to, but which belongs morphosyntactically with a different predicate, is immaterial. The result stands
even if Chamorro number agreement is taken to be inflectional, as in 6.1. This is why we
embarked on the investigation in the first place.
Further, and interestingly, the small clause complement of the verb na’ turns out to
fill a gap in the paradigm of Chamorro complementation. As might be expected, the
language has various types of clausal complements, including finite clauses, infinitive
clauses, and the embedded “clause” of restructuring constructions. Finite clauses and
infinitive clauses are clearly TPs. They differ in that finite clauses are specified for mood
and can have an overt subject, whereas infinitive clauses are mood-invariant and cannot
have an overt subject (see Chung 1998: 64–68). Embedded “clauses” of restructuring
constructions are similar to infinitive clauses in these respects, but smaller (see Chung
2004). Given the claim in 6.1 that person-and-number agreement realizes features of T,
these embedded “clauses” are best analyzed as defective TPs, as proposed by Bhatt 2005
for Hindi-Urdu – TPs whose head is parasitic on the T of the clause under which they
are embedded.
The three types of clausal complements just described show number agreement and
some person-and-number agreement. Finite clauses make full use of the agreement paradigms in (3–4). Infinitive clauses show realis number agreement when their predicate
is intransitive and the invariant infix -um- when it is transitive. Embedded “clauses” of
restructuring constructions show realis number agreement when their predicate is intransitive and either realis person-and-number agreement, or the infix -um-, when it is
If na’ truly is a verb, then its small clause complement differs from the other types
of clausal complements just mentioned along all of these dimensions. The small clause
complement of the verb na’ is merely a vP – even smaller than the embedded “clause” of
restructuring constructions – but it can have an overt subject. And, because it is merely
a vP, it shows (irrealis) number agreement but no person-and-number agreement at all.14
Interestingly, Chamorro has at least one other verb that can select a vP complement: the imperative verb
cha’- ‘don’t, shouldn’t, better not’. As expected, the vP complement of cha’- (a) does not show person-andnumber agreement, but (b) when intransitive, does show irrealis number agreement. Less expectedly, the
specifier of this vP is always controlled PRO, and the verb or adjective from which vP is projected must be
inflected for progressive aspect. Thanks to Pranav Anand for questions that uncovered this.
Sandra Chung
None of this language-particular fine detail is theoretically necessary or even expected,
of course. But it is reassuring that the vP complement posited by our alternative approach to causatives can be integrated smoothly into the overall picture of complementation in Chamorro.
7 Other aspects of the morphosyntax of causatives
Other aspects of the profile of causatives present more of a challenge to the proposal
just outlined. I discuss two such aspects below, with the aim of showing that they can be
handled relatively straightforwardly once the right infrastructure is installed. One set of
facts involves wh-agreement; the other involves morphological case.
7.1 Wh-agreement
When a constituent undergoes wh-movement in Chamorro, the verb or adjective on
which it depends shows a special morphological agreement called wh-agreement (see
Chung 1998 and the references cited there). This special agreement, which supersedes
the normal forms of normal subject-verb agreement, signals the grammatical relation of
the wh-trace – whether it is a subject, direct object, or oblique. For instance, when the
wh-trace is a direct object, wh-agreement is realized by the infix -in- and nominalization
of the verb. Nominalization is indicated, among other things, by the fact that the subject
is cross-referenced by (suffixal) possessor agreement (glossed poss) rather than subjectverb agreement. Compare the sentence in (30a), in which the verb shows person-andnumber agreement, with the constituent question in (30b), in which the verb shows
a. Hu kåkannu’ i gollai.
p.agr eat.prog the vegetable
‘I’m eating vegetables.’ (CD, entry for nos)
b. Håfa kinannono’-mu?
‘What have you been eating?’ (from a tape-recorded narrative)
In earlier work I analyzed wh-agreement as the result of feature sharing in abstract
Case between a wh-trace and the T that most immediately commands it. The shared
Case feature is then spelled out on the verb or adjective that projects T’s complement.
I will adopt this analysis here, noting that in minimalist syntax, abstract Case is often
reconfigured in terms of the syntactic head that licenses the relevant DP via Agree.
Let us now turn to causative sentences and consider the DP described at the beginning
of §3 as the subject of the inner predicate. The proposal we are exploring treats this DP
as the subject of the small clause complement of na’ – in other words, as the specifier of
the embedded vP in the schematic diagram below. (This specifier is represented to the
13 Another way around causatives in Chamorro
right for convenience; see Chung 1998 on the derivation of predicate-first word order in
The small clause subject is in an ECM configuration, so it is licensed by Agree with
the small v that immediately commands na’ (= v1 in the diagram in (31)) in essentially
the same way as if it were the direct object of na’. This licensing is confirmed by whagreement: when the small clause subject undergoes wh-movement, wh-agreement signals that the wh-trace is a direct object (see Gibson 1980: 82, 164).15
a. Håyi na pilotu nina’-baståm-mu?
who? l pilot wh.caus-quit-poss
‘Which pilot did you fire (lit. make quit)?’
b. Ha na’-moderåtu si Lillian i [nina’-maipen-ña] hånum.
p.agr caus-moderate unm Lillian the wh.caus-hot-poss water
‘Lillian moderated (the temperature of) the water that she was making hot.’
(CD, entry for moderåtu)
Next, consider structures in which the inner predicate is transitive and so the small
clause complement of na’ contains a direct object. This embedded direct object is licensed
by Agree with the small v that immediately commands the inner predicate (= v2 in (31)).
Therefore, when it undergoes wh-movement, wh-agreement signals that the wh-trace is
a direct object (see Gibson 1980: 197).
Håfa nina’-li’e’-ña
si Maria nu hågu?
what? wh.caus-see-poss unm Maria obl you
‘What did Maria show you (lit. cause you to see)?’
Not only does wh-agreement register the same Case feature for both types of whtraces, but in both constructions the verb on which the agreement is realized is the
In (32b), the construction of interest is a prenominal relative clause (in brackets), and what has undergone
wh-movement is – depending on one’s assumptions – either a null relative operator or else the head NP
hånum ‘water’.
Sandra Chung
higher verb, namely, the causative na’. It is this verb that is infixed with -in- and undergoes nominalization, as can be seen from the fact that its subject (the causer) is the
DP cross-referenced by possessor agreement. It may seem surprising that wh-agreement
is realized on the higher verb, given that the wh-traces in (32) and (33) are arguments
of the inner predicate. But the pattern follows from the syntactic structure proposed for
causatives in §6.1, plus the independently motivated assumption that wh-agreement involves feature sharing between the wh-trace and T. Because small clauses do not contain
T, a wh-trace in the small clause complement of a causative must share its abstract Case
feature with the matrix T. As usual, the shared Case feature is spelled out on the verb or
adjective that projects T’s complement, which in this case is the causative na’.
It may seem even more surprising that the possessor agreement that ought to be realized on the nominalized verb is spelled out on what is apparently the inner predicate. I
contend that what lies behind this unusual spell-out is the prosodic deficiency of na’. In
Chamorro, affixes must attach to phonological words. This point emerges most clearly
for suffixes, perhaps because suffixation invariably causes primary stress to shift to the
penultimate syllable of the suffixed word. Since na’ is not a phonological word, but
rather prosodically deficient, the suffix that realizes possessor agreement must attach
instead to the phonological word immediately dominating it – the phonological word
formed by stray adjunction of na’ to the inner predicate. This, I claim, is responsible for
the unusual location of possessor agreement in (32–33).16
One might wonder how the same facts would be handled by a more traditional analysis of Chamorro causatives that treats na’- as a derivational prefix. Such an analysis
could deal straightforwardly with the spell-out facts just described, because it takes
the combination of na’- plus the inner predicate to be a complex word (and therefore
a phonological word). It would, however, have more trouble with the evidence provided
by wh-agreement that both the subject and direct object of the inner predicate are licensed by (different instances of) small v. This is because the more traditional analysis
assumes that there is just one verb, and therefore just one small v, in the structure.
It should be noted that Chamorro has no double object verbs – no verbs whose small v
licenses more than one DP as a direct object. Verbs of transfer, for instance, have just one
DP that activates the object form of wh-agreement when it undergoes wh-movement –
namely, the DP that realizes the theme (not the DP that realizes the goal; see Gibson 1980:
161–163). What this means is that a more traditional analysis of Chamorro causatives will
have to stipulate that the derived causative verb, exceptionally, has two arguments that
activate this form of wh-agreement. But no such stipulation is needed in the small clause
analysis of this construction, as we have just seen.
A reviewer asks if -in- infixation might target the phonological word containing the relevant verb. It might
indeed. However, what matters here is that infixation does not target the phonological word consisting
only of the inner predicate (which, recall, is distinct from the phonological word consisting of the inner
predicate plus na’). This can be seen from the ill-formedness of *na’-lini’e’-ña as opposed to nina’-li’e’-ña
‘she caused to see’. More generally, it is hard to locate Chamorro evidence that prefixes and infixes must
attach specifically to phonological words (as opposed to just any phonological material).
13 Another way around causatives in Chamorro
7.2 Morphological case
I mentioned in §2 that Chamorro has three morphological cases – unmarked, oblique,
and local – and that subjects and direct objects occur in the unmarked case. We must
now confront the fact that a causative sentence has just two DPs in the unmarked case:
the subject of na’ and the subject of the inner predicate. The direct object of the inner
predicate, if there is one, occurs in the oblique case. See the examples below.
a. Hu na’-ayao
si Isidro ni kareta.
p.agr caus-borrow unm Isidro obl car
‘I let Isidro borrow the car.’ (CD, entry for ayao)
b. Maila’ ya bai hu na’-li’i’ hao ni cha’ka gi kodu-mu.
come and p.agr caus-see you obl rat
loc arm.muscle-poss
‘Come and let me show you (lit. I will make you see) the rat in your arm
muscle.’ (CD, entry for chå’ka)
This pattern raises a question. Given the wh-agreement evidence that the subject and
direct object of the inner predicate are licensed in the same way (by a small v), why do
these DPs occur in different morphological cases?
In minimalist syntax, one way of resolving disconnects between morphological case
and morphological agreement is to take case to reflect some mechanism other than licensing by a syntactic head. The mechanism usually invoked is case competition (also
known as dependent case assignment; see Marantz 1991 and many others since). The
leading idea behind case competition is that if two DPs are in the same local domain,
independent of each other, and not already case-marked, the presence of one DP will
cause the other DP to be assigned case.
Baker (2015) develops a theory of structural case in which various aspects of case competition are parameterized, including the local domain in which the DPs occur and the
specifics of the c-command relation holding between them. In his theory, dependent
case can be assigned in two local domains, VP and TP, which are the spell-out domains
of phases. Significantly for our purposes, the evidence that VP is a local domain comes,
in part, from causative sentences in Chamorro. Baker (2015: 137–139) assumes that Chamorro causatives are morphologically complex verbs, and therefore causative sentences
have a single VP that contains the complex verb’s direct object (= the subject of the inner
predicate) and can contain another DP (= the direct object of the inner predicate). The
dependent case assignment that he proposes for Chamorro is essentially as follows.
Baker (2015) on dependent case assignment in Chamorro
a. Suppose DP1 has not been marked for case. If DP1 is c-commanded by DP2
and both are in the VP domain, assign DP1 oblique case;
b. Otherwise, assign DP1 unmarked case.
Sandra Chung
As he observes, this case assignment handles the distribution of oblique versus unmarked case in causative sentences as well as other clause types (e.g. clauses constructed
from verbs of transfer).17
Obviously, this proposal does not mesh well with the analysis of Chamorro causatives
being explored here. The small clause structure I proposed in §6.1 for causatives locates
the subject and the direct object of the inner predicate in different VP domains (see (31));
this will prevent dependent case assignment from occurring. However, Baker’s theory
of case allows structural case to be assigned under Agree or through case competition,
and this suggests other options.
Baker takes unmarked case to be the default case in Chamorro. Suppose we take
the opposite position and declare oblique case to be the default case. Then the task
becomes to assign unmarked case to the various types of Chamorro DPs that exhibit
it.18 Among the DPs that occur in the unmarked case are subjects, possessors, and DPs
in topic/focus position. These DPs are the specifiers of the functional heads T, D, and
C, which license them via Agree (see Chung 1998). Moreover, each licensing relation
gives rise to some type of morphological agreement: person-and-number agreement (for
subjects), possessor-noun agreement (for possessors), or operator-C agreement (for DPs
in topic/focus position). All this suggests that unmarked case is assigned to these DPs
under Agree.
Direct objects also occur in the unmarked case, where the “direct object” of a causative
sentence is the inner predicate’s subject but not the inner predicate’s object. Since direct
objects – including the inner predicate’s object – are licensed by transitive small v via
Agree, the obvious move is to try to get their case to follow from a more limited version
of that relation. I claim that unmarked case is assigned to these DPs under Agree, but
only when transitive small v is selected by T. The italicized extra requirement may look
stipulative. But there is evidence from several areas of Chamorro grammar that feature
sharing occurs between small v and the T that selects it. Number agreement spells out
not only the number feature of small v but the finiteness and mood features of the T
that selects it (see §6.1). Further, the morphological operations responsible for personanimacy effects in Chamorro require that this feature sharing extend to person and other
features of the DPs licensed by these heads (see Chung 1998; 2014). This feature sharing
can be achieved in multiple ways which, frankly, are not of particular interest. What
is relevant is that case assignment to direct objects can now be understood as follows:
unmarked case is assigned by transitive small v under Agree, but only when it shares
features with T.
This achieves the desired outcome. In causative sentences, the subject of the inner
predicate will be assigned unmarked case, because it is licensed in an ECM configuration
by a small v (= v1 in (31)) that shares features with T. But the direct object of the inner
The local case does not enter into the picture, because it is not a structural case.
The unmarked case is also used for predicate nominals and the objects of most overt prepositions. The
oblique case is also used for various DPs treated in Chung 1998 as objects of null prepositions: passive
agents, instruments, and DPs that realize the complements of antipassive verbs, other intransitive predicates, and nominalized verbs. It is unclear whether the proposals for case assignment in the text can, or
should, be extended to these other uses.
13 Another way around causatives in Chamorro
predicate will not be assigned unmarked case, because it is licensed by the embedded
small v (= v2 in (31)), which does not enter into a feature sharing relation with T. This
DP will instead be assigned oblique case by default.
The Agree-based case assignment that I have just proposed is summarized below.
Agree-based case assignment in Chamorro
a. Assign unmarked case to DP if it is licensed by T, D, C, or by a transitive
small v that shares features with T;
b. Otherwise, if DP has not been marked for case, assign oblique case.
This case assignment handles the distribution of unmarked versus oblique case in causative sentences as well as clauses constructed from verbs of transfer and various other
clause types. In other words, it has the same empirical coverage as Baker’s dependent
case assignment for Chamorro, but does not require causatives to be analyzed as complex verbs.19 A more sustained comparison of the two approaches to Chamorro case
assignment is better left for another time. My goal here is merely to show that it is possible to give a coherent description of morphological case in causative sentences within
the small clause analysis I propose.
8 Conclusion
Chamorro has many types of inflectional material that could perfectly well be analyzed
as affixes or clitics; for instance, the material that realizes person-and-number agreement (and – conceivably – even the material that realizes number agreement). I hope to
have shown here that the same freedom of analysis, when extended to material that is
apparently derivational, can have thought-provoking theoretical consequences.
number agreement
possessor agreement
Mark Baker (personal communication) observes that a dependent case account of Chamorro morphological
case can be maintained if the VP embedded under na’ is what he calls a ’soft phase’. For reasons of space,
the details are not spelled out here.
Sandra Chung
Sources for the Examples:
CD database: The database for the revised Chamorro-English dictionary.
Saipan, CNMI and University of California, Santa Cruz
Borja, Joaquin F., Manuel F. Borja, and Sandra Chung. 2006. Estreyas Marianas: Chamorro. Saipan, CNMI: Estreyas Marianas Publications.
MAK Marciano, Dolores. n.d. Månnge’ na Alåguan Kalamåsa. National Dissemination and Assessment Center, CSULA
2007. Nuebo Testamento [The Chamorro New Testament]. Diocese of Chalan
Kanoa, CNMI
This paper is for Steve, who for many years has been my teacher, mentor, intellectual
role model, and friend.
Anderson, Stephen R. 1992. A-morphous morphology. Cambridge: Cambridge University
Anderson, Stephen R. 2005. Aspects of the theory of clitics. Oxford: Oxford University
Baker, Mark C. 1985. The mirror principle and morphosyntactic explanation. Linguistic
Inquiry 16(3). 373–415.
Baker, Mark C. 2015. Case: Its principles and parameters. Cambridge: Cambridge University Press.
Bhatt, Rajesh. 2005. Long distance agreement in Hindi-Urdu. Natural Language and Linguistic Theory 23. 757–807.
Borja, Manuel F., Sandra Chung & Matthew Wagers. 2015. Constituent order and parser
control processes in Chamorro. In Amber Camp, Yuko Otsuka, Claire Stabile & Nozomi
Tanaka (eds.), Proceedings of AFLA 21, vol. 25, 15–32. Asia-Pacific Linguistics.
Chung, Sandra. 1998. The design of agreement: Evidence from Chamorro. Chicago: University of Chicago Press.
Chung, Sandra. 2003. The syntax and prosody of weak pronouns in Chamorro. Linguistic
Inquiry 34(4). 547–599.
Chung, Sandra. 2004. Restructuring and verb-initial order in Chamorro. Syntax 7. 199–
Chung, Sandra. 2014. On reaching agreement late. In Andrea Beltrama, Tasos Chatzikonstantinou, Jackson L. Lee, Mike Pham, & Diane Rak (eds.), Cls 48, vol. 1, 169–190.
Chicago: Chicago Linguistic Society.
Durie, Mark. 1986. The grammaticization of number as a verbal category. In Proceedings
of the twelfth annual meeting of the Berkeley Linguistics Society, 355–368. Berkeley
Linguistics Society.
13 Another way around causatives in Chamorro
Gibson, Jeanne D. 1980. Clause union in Chamorro and in universal grammar. University
of California, San Diego PhD thesis.
Marantz, Alec. 1991. Case and licensing. In German Westphal, Benjamin Ao & Hee-Rahk
Chae (eds.), Proceedings of the 8th Eastern States Conference on Linguistics (ESCOL 8),
234–253. Ithaca: CLC Publications.
Safford, William Edwin. 1903. The Chamorro language of Guam. American Anthropologist
5(2). 289–311.
Safford, William Edwin. 1904. The Chamorro language of Guam – III. American Anthropologist 6. 95–117.
Topping, Donald M. & Bernadita C. Dungca. 1973. Chamorro reference grammar.
reprinted in 1980. Honolulu: University of Hawaii Press.
Zwicky, Arnold. 1977. On clitics. Bloomington: Indiana University Linguistics Club.
Chapter 14
Preliminaries to the investigation of
clitic sequencing in Greek and
Mark Hale
Concordia University, Montréal
Beginning with Wackernagel (1892), scholars have dedicated a great deal of attention to
the question of the placement of enclitic elements within their clause, particularly those
which (tend to) appear in so-called “Second Position”. Anderson (2005) summarizes the
honorand’s long-standing interest in and contribution to the critical “interface” issues which
arise for linguistic theory from these elements. As has been well known, again since at least
the time of Wackernagel’s writings on the matter, several archaic Indo-European languages
enjoy particularly rich inventories of relevant elements. Greek, in particular, has a set of
toneless (“enclitic”) and tonic (“postpositive”) so-called “second position” lexemes. Since
only one entity can technically occupy “second position” in any given string, an obvious
empirical issue arises regarding clauses which contains multiple “second position” elements.
This matter has received significantly less careful attention than has ”second position clisis”
generally in the past 125 years. In this paper I present a detailed analysis of what happens
when multiple elements seem to have a demand on “second position” in the Attic Greek
clause (focussing on the Plato and Euripides corpora), and demonstrate that in developing
an account for why the ordering is the observed one, a richer understanding of the actual
mechanisms behind our (ultimately epiphenomenal) Wackernagel’s Law can be developed.
1 Introduction
In Anderson (2005) the honorand of this volume presented a detailed analysis of clitic
placement, including the positioning of so-called “second position” (henceforth, 2P) clitics. I have neither the competence nor the space to fully engage with his insightful
and intriguing proposals in this venue. However, given that his analysis was developed
against the backdrop of a consideration of empirical data from a broad set of languages,
we can all recognize that it is necessary for such cross-linguistic approaches to abstract
away from a certain number of seemingly low-level technical matters which arise in individual linguistic systems so as to not impede the development of a general theory. In
Mark Hale. 2017. Preliminaries to the investigation of clitic sequencing in Greek
and Indo-Iranian. In Claire Bowern, Laurence Horn & Raffaella Zanuttini (eds.),
On looking into words (and beyond), 289–310. Berlin: Language Science Press.
Mark Hale
some cases, one is leaving to one side matters which are both well described and well
understood in the specialist literature on the language in question. However, in the case
of what is perhaps the most famous data on 2P clitics – the Greek and Indo-Iranian data
made use of by Wackernagel (1892) in the grounding document for so-called “Wackernagel’s Law” – this is not the case. Surprisingly, the empirical data for these languages
is relatively poorly understood, in my view, even in the specialist literature.
It should be clear that building a general theory of clitic behavior in human linguistic
systems is only going to be as successful as the quality of the empirical data on individual
languages used as input to the theory construction process allows. The more poorly such
data is described, or understood, the potentially weaker the resulting general theory. In
this paper I focus on one aspect of clitic behavior in systems of the Greek/Indo-Iranian
type:1 the sequencing of 2P clitics. I argue that the weakness in our capacity to insightfully account for observed sequencing of 2P clitics highlights the shortcoming of our
understanding of the phenomenon in these languages generally, and point toward some
ways we might, in my opinion, approach these issues so as to improve our understanding. This would allow the Greek and Indo-Iranian data to play their rightful role in the
evaluation of more broadly-based theories such as that of Anderson (2005).
2 Wackernagel’s so-called Law
The literature on so-called “second-position” clitics goes back at least to Bartholomae
(1886). Wackernagel’s Law universally recognizes a tendency for certain types of prosodically deficient element to occupy “second position” in the clause in Ancient Greek and
Indo-Iranian (at least). A very clear statement of Wackernagel’s own version of his “law”
can be seen at the beginning of his famous Vorlesungen über Syntax (Wackernagel 1920:
Z. B. im ältesten Griechisch, in sehr hohem Masse bei Homer, auch noch bei Herodot, ist das Gesetz lebendig, dass schwach betonte Wörtchen, welches immer ihre
syntaktische Beziehung sei, unmittelbar hinter das erste Wort des Satzes gestellt
In spite of the 130 years which has passed since Bartholomae first posited the tendency,
and the heaps of follow-up scholarly literature, work which addresses the specific issue
we will concern ourselves with today – the sequencing of 2P clitics – in any serious
way is quite sparse. Early work was impeded by the descriptive goals of traditional IndoEuropean studies: after all, if you only need to catalogue observed sequences of 2P clitics,
the task simply involves gathering and reporting on the data provided by one’s corpus.
I think it is widely recognized that these two systems are very similar to one another, in an Indo-European
context, both in terms of the richness and diversity of their clitic inventory and in the syntax of these
“For example, in the earliest Greek the law is active – to a very large extent in the case of Homer, also
still in the case of Herodotus – which holds that weakly stressed ‘little words’, whatever their syntactic
relationship might be, are placed directly after the first word of the clause.”
14 Preliminaries to the investigation of clitic sequencing in Greek and Indo-Iranian
However, when we try not to describe the superficial properties of a string, but to view
that string as the epiphenomenal output of a computational device (the grammar), the
system needs to be imbued with powers which will enable it to bring the linear order of
the output string into existence, rather than just describe that order after the fact (the
way the researcher does).
3 The “clitic cluster”
The various modern Indo-Europeanist approaches to Wackernagel phenomena in Greek
and/or Indo-Iranian reference different methods for dealing with the sequencing problem. One approach, commonly seen in work that takes a “prosody-centric” view of
things, posits a “second position” clitic “cluster” (or “chain”) into which the clitics are
placed, and assumes some kind of (usually stipulative) ordering algorithm which arranges the clitics within this cluster. This is my understanding of the general approach
(though, of course, they differ in detail) of Keydana (2011), Lowe (2014) and Goldstein
(2016). A rough graphical representation of such approaches is presented in Figure 1.
2P Clitic Cluster
Figure 1: The “Clitic Cluster”
Typically, the domain within which 2P is defined is taken in these approaches to be
prosodically defined, and the entities which undergo 2P placement (the clitics) are taken
to be a prosodic class, but I have never seen an empirically-grounded proposal claiming
that the observed ordering within the clitic cluster is determined by prosodic considerations alone. I imagine that this is because it is pretty hard to see any trace of the
domination of such a factor in the observed clitic sequences in these languages.
Mark Hale
But in general these authors have not yet ventured a systematic hypothesis about the
sequencing, so it is hard to determine in any detail the nature of the envisioned system.
For example, Keydana (2011: 108, fn. 3) says “[a]nother issue not to be addressed in this
paper is the internal structure of clitic clusters.” Goldstein (2016: 88) says that because
the matter is a “difficult issue”, he will “leave it for future research, and for the moment
assume templatic ordering…”3 Lowe (2014: 28) writes that “there are regular (though
not inviolable) orders when more than one element from any one of those categories
occurs…It is beyond the scope of this paper to account for those patterns.”
Interestingly, each of the authors does seem to feel that the observed ordering is related to “classes” into which the clitics fall. Thirty years ago (Hale 1987: 73), I argued
that there were three distinct classes of 2P clitics, taking distinct 2Ps for distinct reasons. Although these are not the labels I would use if I were creating them today, and
the mechanisms described in 1986 for how the categorization of the clitic relates to its
placement are woefully antiquated, the classes were: (1) “emphatic” clitics, (2) “sentence
connective” clitics, and (3) “sentential” (usually pronominal) clitics. The first I took to involve word-level attachment and the second clause-level. The third were constrained to
appearing in a very low position in what we would now call the CP domain, or, indeed,
at the top of the IP domain.
I wrote then that “[t]he position of these elements relative to one another follows
naturally from this account of their origins: the regular sequence is emphatic + sentence
connective + sentential…” It is hard to see this statement as anything more than either
wishful thinking or blissful ignorance, both of which I possessed in spades back in those
days. From my crankier contemporary perspective, it is pretty easy to see that no explicit
characterization of the membership in these classes was provided (though I gave isolated
examples of each), nor was any mechanism even hinted at for what might trigger any
specific ordering in cases of multiple instantiations of one of these categories in a single
clause. In these matters, unfortunately, I have been largely followed by more recent
I think it is clear enough, however, what we would all like to see. Overt stipulation is
in essence an admission of explanatory failure, and all principles of the scientific pursuit
demand of us that we attempt to minimize the role of stipulation in our models. So, our
hope must be that the clitics fall into non-arbitrarily-defined classes, and that these clitic
classes occupy well-motivated positions (relative to the functions the clitics instantiate)
in the linguistic representation. Since it is safe to assume that no two clitics do exactly the
same work and co-occur in a single clause, if order falls out in some way from function,
we should always be able to generate a predicted ordering for any pair of clitics. It is the
“in some way” that I want to explore today. As a step in that direction, though, I must
dwell a little longer on previous work.
No explicit template is provided which accounts for the cited data.
14 Preliminaries to the investigation of clitic sequencing in Greek and Indo-Iranian
4 The Hock template
Around the same time as I was writing the account discussed above, Hans Hock was
working out the details of an overtly “templatic” approach to clitic sequencing in Vedic
Sanskrit (e.g., Hock 1996, with earlier literature). The Hock system was not a model of
internal consistency. In this system, the 2P clitics fall into three classes, P, Ṕ, and D:
atonic non-“deictic” (i.e., non-pronominal) clitics
tonic non-“deictic” clitics (sometimes called “postpositives”)
atonic “deictics” (i.e., pronominal clitics)
Clearly, the classification is based on both prosodic (tonic vs. atonic) and functional considerations (deictic vs. non-deictic). The template is basically stipulative, but in its most
common instantiation, shown in (2), it is supposed to be partially motivated: non-deictics
precede deictics (thus P, Ṕ before D, D´) and atonic elements from each class precede the
corresponding tonic elements (thus P before Ṕ, D before D´).
(2) X´
The template is often described as “phonological” or “prosodic” (including by Hock himself), but that, of course, is not accurate. It is important to note that no motivation is
cited for either of the two ordering principles: no reason for why non-deictics should
come earlier in the clause than the deictics is given, nor for why the atonic element of
each category should precede the tonic one.4
The graphical representation of the template in (2), especially when coupled with the
claim that the template is “prosodic”, invites the inference that the “goal” of the system is
have an alternating strong vs. weak tonicity sequence. But since P, Ṕ and D can all double,
and since any of those elements can be absent in any given clause, no such alternating
pattern is observed in the vast majority of utterances in the actual corpus. The prosodic
motivation for the template is thus highly abstract (not that that determines whether it
is accurate or not).
My actual point in discussing the Hock system, however, is to point out that the mechanisms involved are quite distinct from that which we saw in our discussion of “clitic
cluster” approaches to clitic sequencing. We can visualize the system as seen in Figure 2
There is no “second position” in this conception of things, but a variety of ordered
positions into which clitics are placed based on their properties (prosodic and functional).
We can see the differences between the two general models if we imagine a clause with
only a single 2P clitic in it. In the “clitic cluster” approach that clitic will be “in 2P”, plain
and simple. The ordering algorithm will have no work to do. In an approach such as that
found in the Hock template, we can (and must) still ask the question: where is the clitic?
In fact, D´ elements may appear in the X´ position and, for Hock’s version of the Rigvedic template, also
in the Ṕ position, and thus come earlier in the clause than D elements, indicating that there isn’t much
content to this latter principle in any event.
Mark Hale
Figure 2: The “Hock template”
It could be in P (with Ṕ and D empty), or in Ṕ (with P and D empty), or in D (with P and
Ṕ empty).
There is, of course, an inherent advantage, given that we are seeking to develop a
non-stipulative account of ordering, if we assume that specific clitics map consistently
to specific positions in the string, and that whether that position ends up being “second”
or “third” or “fourth” is simply the epiphenomenal by-product of the mapping algorithm.
A model which, by contrast, dumps the clitics unordered into a 2P cluster and then must
impose an order on them has, in some sense, missed a chance to impose that ordering earlier in the initial mapping. Particularly if that mapping is prosody-driven, only a general
slackening of constraints on the relationship between position (at the point of insertion)
and interpretation – constraints which seem to hold of the linguistic representation generally – can allow a set of unordered elements, placed by a prosodic algorithm, to be
realigned in an interpretationally-relevant manner (i.e., on the basis of the “functional
class” of the clitic).
Hale (1996) represents an attempt to link the “positions” in the Hock template to specific structural elements within the clause (a similar orientation is found in other syntaxcentric but prosody-sensitive approaches). I will not dwell on that reinterpretation of
Hock here, but instead turn to one more modern model of 2P placement.
5 Wackernagel’s optimality-theoretic approach
Wackernagel himself pretty clearly conceived of the 2P phenomenon as a kind of a compromise between a clitic element representing the kind of information which should
come early in the clause (generally because of its “linking” function to the preceding
context) and that same element being atonic, and thus not particularly suitable for initial position. This traditional conception is very closely followed by modern Optimality
Theoretic approaches to 2P placement, as reflected, for example, in Anderson (2005).
The basic idea of such approaches, as with Wackernagel’s, is that there is a drive for
2P clitics to be initial (captured by an AlignLeft constraint, which incurs one violation
14 Preliminaries to the investigation of clitic sequencing in Greek and Indo-Iranian
for each step away from the left edge a clitic is) and a prohibition against the clitic appearing in initial position (captured by a NonInitial constraint). Obviously, both of
these constraints cannot be satisfied in the same string. As always in OT, the conflicting
demands are resolved via a ranking of the constraints. As long as NonInitial outranks
AlignLeft, the clitic will appear as close as possible to string-initial position, but not in
initial position – hence the 2P effect.
We can see this with a tableau fragment concerning the positioning of Greek gár
‘because, since’ relative to two tonic lexical items, which I designate simply as X´ and Ý.
Greek gár is a standard example of a “Wackernagel’s Law” element in the language.5
X´, Ý, gár
X´ Ý gár
X´ gár Ý
gár X´ Ý
gár Ý X´
Ý gár X´
Ý X´ gár
Wackernagel believed that the drive towards initial position was of different strength
for different enclitic objects. As long as we make the AlignLeft constraints clitic (or
clitic class) specific, a strict ranking version of OT will require us to decide for each clitic
just how strong the demand is that it be aligned left. For example, if we take the Greek
focus-marking enclitic mén, we are compelled by the model to ask: is it more important
for mén to be on the left, or for gár to be on the left? (In Wackernagel’s terms, is the
“drive for initial position” stronger for mén or for gár?)
The tableau in (4) below shows how such a system generates – again by necessity – a
sequencing of clitics. We should always be attentive, I think, when our modern models
seem to converge on the ideas of important scholars like Wackernagel, who could not
have envisioned the workings of OT, and thus independently had a similar conception
of a certain class of linguistic phenomena. If Wackernagel and Anderson agree, it would
be wise to not dissent too quickly! But is OT the happy confirmation of Wackernagel’s
less formal and more intuitive analysis?
The model in some sense combines properties of the two earlier approaches I sketched:
there is a sense of a “fixed stipulative order” (as seen in the Hock Template) in the relative rankings of the AlignLeft(x) constraints, but there is no sense of fixed stipulative
positions in the resulting representation (as that model implies), thus making it more like
the “clitic cluster” analysis. It has the advantage, over the “clitic cluster” analysis, of not
adding a new object to our model of the grammar (the “clitic cluster”), nor requiring a
distinct computational process which is responsible for explicitly ordering the elements
within the cluster (the “ordering algorithm” mentioned above).
The reader will notice that this tableau fragment does not select between two “winning” outputs: X´ gár Ý
and Ý gár X´. Obviously other constraints will determine the ultimate optimal output form with respect to
the relative ordering of the tonic elements to one another.
Mark Hale
X´, Ý, gár, mén
X´ Ý gár mén
X´ Ý mén gár
X´ mén Ý gár
mén X´ Ý gár
X´ gár Ý mén
X´ gár mén Ý
X´ mén gár Ý
mén X´ gár Ý
gár X´ Ý mén
gár X´ mén Ý
gár mén X´ Ý
mén gár X´ Ý
gár Ý X´ mén
gár Ý mén X´
gár mén Ý X´
mén gár Ý X´
Ý gár X´ mén
Ý gár mén X´
Ý mén gár X´
mén Ý gár X´
Ý X´ gár mén
Ý X´ mén gár
Ý mén X´ gár
mén Ý X´ gár
I have both conceptual and empirical concerns about the model. At the conceptual
level, stipulation is built into the model pretty deeply: the so-called “factorial typology” argument says that we could just as easily have AlignLeft(gár) outrank AlignLeft(mén), and indeed that every ordering of every available AlignLeft(x) constraint
should be, in principle, observed. This does not seem to me to be very consistent with
the data I have seen from archaic Indo-European languages (which often involves etymologically unconnected, but functionally similar enclitics showing the same ordering
Empirically, one of the great challenges to this model holds, in my view, for all of
the other models we have treated to this point as well. The domain over which the
AlignLeft(x) constraint must be assessed is, in the model, a pure stipulation (i.e., there
are no principles regulating what a given enclitic might be “aligned” to). The same problem, in my view, plagues the so-called “prosody-centric” approaches which have become
popular: the clitic is said to move into the “clitic cluster” in second position of some domain, but none of the approaches I have seen (Keydana 2011; Lowe 2014; Goldstein 2016)
present a non-stipulative characterization of how the appropriate domain is established.
14 Preliminaries to the investigation of clitic sequencing in Greek and Indo-Iranian
A clitic is said to take 2P within (some) Intonational Phrase, but the question of which
one (when there are multiple IPs in the string) is left unclear. It also seems like the “Intonational Phrase” portion of the specification (and, in non-OT approaches, the 2P part
as well) is purely stipulative: could one place a clause-conjoining clitic in 2P of (some)
Utterance Group or (some) Phonological Word? How about in 3P in the Intonational
Phrase? The models proposed are so inexplicit it is difficult to determine precisely what
their allowable elements and operations are – the models are thus both highly stipulative
and poorly constrained.
As mentioned above, my concerns about the Hock Template have been spelt out in
considerable detail in Hale (1996), where I attempt to reduce the empirically valid aspects
of the template to the interaction of (1) normal syntactic placement of the relevant entities and (2) Halpern’s (1995) “prosodic inversion”. I won’t go through these details here,
except to note that in this conception of things enclitic elements can be in a variety of
structural positions (in the syntax) and may (or may not) undergo “prosodic movement”
in the phonology (depending on whether they are “properly hosted” without such movement).
It is obvious that a system which leverages both syntax and phonology to account for
observed clitic placement is in some sense more complex than either a purely syntactic
or a purely phonological one. But the evidence that the process is not purely syntactic is,
as far as I can tell, universally accepted. The 2P clitics often interrupt manifest syntactic
constituents (in a manner the syntax does not allow) and are quite regularly placed in
positions in the string from which appropriate scope relations could not possibly be
established. “Prosodic” rearrangement allows the syntax to be mundane rather than
strikingly bizarre (with all of the implications such bizarreness would have, if allowed,
for our theories of grammar).
That the syntax is involved is, in my view, absolutely required if we are to solve the
“domain” problem. Sanskrit ca and Attic te are 2P clitics, but they appear in clause-second
position only when their domain is the clause. They appear in “DP-second” position
when their domain is the DP, and in VP-second position when their domain is the VP.
Their domain simply cannot be defined with respect to prosody alone.
6 Sample application: Enclitic subordination
Both of these observations can be seen to be at work in a set of examples involving
enclitic subordinators. Before examining this data, which will also display the mechanisms I believe are at work in clitic placement (and thus what we might use to explain
clitic sequencing), I will remind the reader that Hale (1987) showed that there was a
process manifest in the language of the Rigveda, in Avestan, and in Greek, whereby a
single constituent could be fronted to the left of a WH-element. Obviously, this analysis
extends (though the matter was not overtly discussed – but rather implicitly assumed –
in that earlier work) to other subordinators present in C (‘if’, ‘when’, ‘because’). I will
call that fronting process “topicalization” (and the position into which the fronting takes
place Top), with no particular commitment to the discourse functions involved.
Mark Hale
Familiar examples of this fronting looks like this:
a. [áśmānam cid]
rock-Asg Rel-NPlM smashed-IIIPl words-ISgN
‘who smashed even rock with (mere) words…’
b. [idhmáṃ]
kindling-ASg Rel-NSg would bear-IIISg
exerting himself-NSg
‘who, exerting himself, would bear the kindling to you…’
RV 4.16.6c
RV 4.12.2a
Armed with the following assumptions, then, let us see what some structures look like,
and how they might impact the development of a theory of clitic sequencing:
• enclitic elements are placed in “expected” syntactic position
• “prosodic inversion” is triggered when they are not properly hosted on their left
Wackernagel proposed long ago that there were traces in the Rigveda of a reflex of IE
*kʷe which (like OLat. absque me esset ‘if it were without me’, some uses of Gothic nih,
and, although not known to Wackernagel, Hittite takku) is subordinating in function,
generally rendered ‘if, when.’ As might be expected, the verb in such clauses, as in
subordinate clauses in Vedic generally, is accented. This has given rise to some anxiety
that the true subordination marker is the verbal accent, and that ca is simply (weakly)
coordinating. Typical examples are:
a. níuptāś
ca babhrávo vā́cam
scattered-down ca brown-NPl voice-ASg they made
eṣāṃ niṣkr̥táṃ
I go=PTCL theircl appointed place-ASg paramour-NSg=like
‘And as soon as, scattered down, the brown (dice) have raised their voice, I
just go to their appointed place, like a girl with a lover.’ (SJ/JB)
b. tuváṃ ca soma
no váśo
you-NSg ca Soma-VSg uscl you should wish
jīvā́tuṃ ná marāmahe
to live Neg we will die
‘And if you will wish us to live, Soma, we will not die.’ (SJ/JB)
RV 1.91.6ab
I have provided the translation of Jamison & Brereton (2014) because it reflects directly
the unease that some Vedicists feel about “subordinating” ca: they have translated ‘and
as soon as (=when)’ and ‘and if’, leaving it unclear whether they believe that ca is coordinating (‘and’) and the verbal accent subordinating (‘as soon as’, ‘if’), or whether perhaps
they believe that this ca actually means ‘and as soon as/if’.
14 Preliminaries to the investigation of clitic sequencing in Greek and Indo-Iranian
As one can imagine, determining whether or not such clauses are weakly connected
to the preceding discourse – i.e., whether the ‘and’ should be in the English translation
– is no easy task. And from this example, and other widely accepted ones (such as RV
8.21.6ab below), in which ca occupies second position, it doesn’t seem likely that our
approach to clitics is going to be much help in this task.
áchā ca tvainā́
PV ca youcl =this-ISg homage-ISg we address
ví dīdhayaḥ
Q-marker for a moment evencl PV you will think
‘When we address you with this homage, will you hesitate even for a moment?’
RV 8.21.6ab
Hettrich (1988: 252) notes overtly on subordinating ca that ca stands “wie nach Wackernagels Enklisengesetz[fn deleted] zu erwarten, überwiegend an zweiter Stelle im Satz.”6
It is somewhat striking to see the überwiegend in this statement, because coordinating
ca is extremely regular in its “second position” behavior, being postponed only in cases
in which there is a “phonological word” at the start of the conjoined domain. Are there
actually cases of “late ca” in subordinating function? The following examples seem to
answer this question “yes”.
a. asyá ślóko
divīý ate
his call-NSg heaven-LSg=speeds earth-LSg
ná yaṃsad
steed-NSg like will control bringing-wondrous-apparitions-NSg
ná hetáyo
yánti cemā́
wild beasts-GPl like charges-NPl they go ca=these-NPl Brhaspati-GSg
abhí dyū́n
having-snake-wiles-APl to heavens-APl
‘The discriminating one [=Br̥haspati?], like a steed, bringing wondrous
apparitions, will control it when these (words) of Brhaspati, like the charges
of wild beasts, go to the snake-wiles-possessing heavens.’
RV 1.190.4
b. ubháyaṃ
śr̥ṇávac ca na
twofold-ASg will hear ca uscl
arvā́g idáṃ
Indra-NSg nearby this-ASg speech-ASg
fully-focussed-ISg benefactor-NSg soma-drinking-DSg
‘…as would be expected according to Wackernagel’s Law, overwhelmingly in the second position in the
Mark Hale
ā́ gamat
thinking-ISg most-powerful-NSg PV will come
‘When Indra nearby will hear this twofold speech of ours, the most powerful
benefactor will come here to the soma-drinking by reason of our fully
focussed insight.’
RV 8.61.1
To understand this data (and there are one or two more examples), we need to ask the following question: how does conjunctive ca end up in “second position” when it conjoins
clauses, and how would we expect a “subordinating ca” to behave given the assumptions
outlined above?
The behavior of coordinating ca is fairly straightforward. No matter what kind of
syntactic (or prosodic!) entity ca is coordinating, it appears in this configuration:
[X Y Z] can be a clause, as in the case under discussion, or an DP or VP or PP or whatever.
Obviously, ca sits at the left edge of the XP-domain and thus does not have a proper
prosodic host to its left. It therefore must undergo the “prosodic flip”, and, assuming X
is a “phonological word”, the resulting operation will give rise to this string:
ca X ca Y Z
It should be clear that one thing that ca can bring into a coordination relationship with
what precedes it is a clause introduced by a topicalized phrase. Here’s an example.
yád agna
when Agni-VSg this-NSg assembly-NSg will become
devī ́
devéṣu yajatā́
godly-NSg gods-LPl sacrifical-NSg sacrificial one
ca yád vibhájāsi
treasures-APl andcl when you will share out
bhāgáṃ no átra vásumantaṃ
share-ASg uscl then rich-in-goods-ASg pursue
‘When, o Agni, this assembly will become godly among the gods, a sacrificial
one, o sacrificial one, and when you will share out treasures, o you of
independent will, then pursue a share for us rich in goods.’ (SJ/JB)
RV 10.11.8
14 Preliminaries to the investigation of clitic sequencing in Greek and Indo-Iranian
The c-pada of this verse arises from a syntactic structure of the form:
vibhájāsi svadhāvo
So where do we expect subordinating ca to be syntactically? Well, it is isofunctional
with yád in the clause above. So the structure of the subordinate clause in (6a) at the end
of syntactic computation would presumably be:
[níuptās babhrávas] vā́cam ákrata
Since ca is alone up there in the left periphery, it must undergo the “prosodic flip” in the
phonology to be properly hosted.
What would happen if we were to have a Top element in such a clause? In the case
of coordination, it of course is the entire clause, including its initial Topic-phrase, which
gets coordinated to a preceding (or following) clause, and ca thus dominates TopP. But
topicalized material appears to the left of the subordinator (relative pronouns, yádi ‘if’,
etc.), so if ca is a subordinator, we predict a structure such as (contrast coordinating ca
in (12)):
Mark Hale
If the proper “hosting domain” for ca is the CP, then it is unhosted on its left, and we
predict the phonology to restructure this to:
Topic [ca X ca Y Z]
Note that, if there are such examples, since coordinating ca won’t act this way, but subordinating ca should, we would have clear and unambiguous evidence for the syntactic
separation of ca into two distinct types of grammatical element. And of course there are
such examples. We saw them above in (8ab), whose structures are:
a. [Top mr̥gā́ṇāṃ ná hetáyo] [CP ca [yánti ca imā́ bŕ̥haspáter áhimāyām̐ abhí dyū́n]]
b. [Top ubháyaṃ] [CP ca [śr̥ṇávac ca na índro arvā́g idáṃ vácaḥ]]
The construction as a whole is rare, but confirmation for this analysis is provided by
the reflex of subordinating ca in later Vedic texts (and, rarely, already in the Rigveda) –
the subordinating particle céd (etymologically from ca + íd). The “normal” position for
this particle is, of course, in second position.
a. ná vā́
Neg PTCL Lady of the Wilderness-NSg slays
cén nā́bhigáchati
another-NSg céd Neg=attacks
‘In truth, the Lady of the Wilderness does no slaughter, if someone else does
not attack.’ (SJ/JB)
RV 10.146.5ab
ná vedab. yó
who-NSg her-GSgcl udder-ASg Neg knows
stánān utá
thereto=PTCL her-GSgcl teats-APl as well
both-ISg=PTCL=him-DSgcl she yields milk
dā́tuṃ céd áśakad
to give céd he was able cow-ASg
‘Whoever knows not the udder of her, and likewise the teats of her, to him
she yields milk with both, if he has been able to give the cow.’ (Whitney) AVŚ
But, as with subordinating ca, we find unexpectedly “late” instances of céd as well.
a. [arthíno]
yánti céd árthaṃ
having-a-task-NPl proceed céd task-ASg
dadúṣo rātím
they will go to PTCL giver-GSg generosity-ASg
‘If those having a task proceed to their task, they will attain the generosity of
the giver.’
RV 8.79.5ab
14 Preliminaries to the investigation of clitic sequencing in Greek and Indo-Iranian
b. [abandhv éke
dádataḥ prayáchanto]
kinless-NPl some-NPl giving-NPl bestowing-NPl
dā́tuṃ céc chíkṣānt
to give céd they are able this-NSg heaven-NSg indeed
‘If some, without kin, giving, bestowing, are able to give, this is truly heaven.’
AVŚ 6.122.2cd
c. héḍaṃ
paśūnā́ṃ ny èti
wrath-ASg cattle-GPl PV comes
brāhmaṇébhyó ’dadad
Brahmans-DPl not-giving-NSg cow-ASg
[devā́nāṃ níhitaṃ
gods-GPl deposited-ASg portion-ASg
cén nipriyāyáte
mortal-NSg céd keeps (for himself)
‘The mortal not giving a cow to the Brahmans goes down to the wrath of the
cattle, if he keeps to himself the deposited portion of the gods.’
AVŚ 12.4.21
And note that we can use our analysis of these “late” instances to make our interpretations of certain Vedic passages more precise. Look at the AB passage (from the Śunaḥśepa
legend) in (19).
(19) r̥nam
debt-ASg him-LSgcl he pays
ca gachati
immortality-ASg andcl he goes to
putrasya jātasya
father-NSg son-GSg born-GSg
cej jīvato
he should see ced living-GSg face-ASg
‘A debt he payeth in him, and immortality he attaineth, that father who seeth the
face of a son born living.’ (Keith)
AB 7.13.4
Keith, whose translation I have provided, takes [putrasya jātasya… jīvato mukham] ‘(the)
face of a son born living’ as a (discontinuous) constituent, the direct object of the verb
paśyet. That is, his analysis (ignoring for a moment the ced, to which we will turn momentarily) is that the subject and predicate divide like this:
(20) [pitā] [putrasya jātasya paśyet jīvato mukham]
There are two possibilities for where pitā could be under Keith’s interpretation: it could
have been fronted into the Topic position, or, of course, it could be in some position
lower than C (in Focus, or in IP, e.g.). If it were below C, the output of the syntax (now
with ced reintroduced) would have been as below, with the “prosodic flip” indicated:
Mark Hale
[CP ced [pitā ced putrasya jātasya paśyet jīvataḥ mukham]]
If pitā were in Topic, we would have instead expected:
(22) [Topic pitā] [CP ced [putrasya ced jātasya paśyet jīvataḥ mukham]]
Neither of these is the sentence in the text. It is clear what the structure must be if the
placement of ced is to fit with all the other evidence for the use of this particle in early
(23) [Topic pitā putrasya jātasya] [CP ced [paśyet ced jīvataḥ mukham]]
‘when the father of a (just) born soni sees the face of (himi ) living’
7 hí, gár, and clitic sequencing
In my dissertation (Hale 1987), I dealt fairly extensively with the data from Vedic hí
‘because, since’. I noted that while the vast majority of instances of hí are in “second
position” (appropriately defined), there were a number of counterexamples. Note that
hí occupies, at the end of the syntactic computation, the very same position (“C”) as
subordinating ca and céd.
I won’t bother citing second position instances of hí – as I said, the vast majority of
the approx. 630 attestations of the particle are in that slot, properly defined. Some of
the not terribly numerous exceptions are given in (24) below. Several interesting issues
arise, so I cite a healthy number of the exceptions.
a. urukramásya
hí bándhur itthā́
wide-striding-GSg this-NSg hí bond-NSg thus
‘for exactly that is the bond to the wide-striding one’ (SJ/JB)
RV 1.154.5c
b. asmā́ñ ca tā́ṃś
ca prá hí néṣi vásya
us-APl andcl them-APl andcl PV hí lead better-ASg PostP
‘lead both us and them forth to a better state.’ (SJ/JB)
RV 2.1.16c
c. tribhíḥ pavítrair
dhí arkáṃ
three-IPl purifiers-IPl he purified hí chant-ASg
‘Since he [=Agni?] purified the chant with three purifying filters,’ (SJ/JB)
RV 3.26.8a
d. ákṣetravit
hí áprāṭ
not-knowing-the-field-NSg knowing-the-field-ASg hí asked
‘Because the one not knowing the field asked the field-knower,’ (SJ/JB)
RV 10.32.7a
14 Preliminaries to the investigation of clitic sequencing in Greek and Indo-Iranian
In all of the “exceptions” I will cite here, we can analyze the data just as we did in the
case of ca and céd: the first constituent of the clause is in the Topic position above the
CP, hí is in C itself, and is not “properly hosted” by a tonic element within its domain
on its left, and thus undergoes inversion. Thus we have [Top urukramásya] in (24a), [Top
asmā́ñ ca tā́ṃś ca] in (24b), [Top tribhíḥ pavítrair] in (24c), and [Top ákṣetravit] in (24d).
In my dissertation, I rather unwisely said, regarding examples such as these, that the
poets were able to treat the caesura as equivalent to a clause-boundary, and thus place hí
in second position after the caesura, rather than after the actual start of the clause. This
is a not a particularly good idea, giving the meter far too much power to determine the
data – certainly far more than I would be willing to concede at this stage of my research
on the matter.
We can give a much more sensible assessment of this data if we instead note that the
boundary between the element in Topic and the start of the CP-domain is marked by an
intonational reset (or pause), and that the natural place to align this pause within the
rhythmic structure of the verse line is at the caesura. In all of the examples above, the
Topic ends at the caesura of a trimeter line (this will be true of the examples I cite below
as well).
As usual, there are many other interesting things going on with these examples as
well. For example, in support of the topicalization analysis, we see in an example such as
(24a) a discontinuity (urukramásya…bándhur). We need to account for this discontinuity,
and movement is the way to do it in our model – topicalization provides the relevant
explanation for that movement. We will see additional examples of this type below.
Finally, and returning to the matter of clitic sequencing, we may be able to learn something important about how exactly the “prosodic flip” works to trigger specific orderings
from examples such as those in (25).
a. índro
ánu hí tvā
Indra-NSg knowing-NSg PV hí you-ASgcl kept an eye on
this-ISg=I-NSg Agni-VSg instructed-NSg have come hither
‘Because the knowing Indra has kept you in his sights, instructed by him
have I come here, o Agni.’ (SJ/JB)
RV 5.2.8cd
b. sadyó jajñānó
ví hīm
at once being-born-NSg PV hí=themcl kindled-NSg he observed
‘for immediately upon being born, he, kindled, observed them’
RV 10.45.5c
Recall that pronominal clitics occupy the lowest position in the C-domain (or the highest
in IP), so one possible structure for what the syntax would have sent to the prosody for
(25a) would be:
[Top índro vidvā́m̐] [CP hí [tvā [ánu cacákṣa]]]
Mark Hale
In this structure, neither hí nor tvā can be properly hosted on their left, with the expected
“prosodic inversion” being thus triggered:
(27) [Top índro vidvā́m̐] [CP hí [tvā [ánu hí tvā cacákṣa]]]
However, as we all also know, there are many “exceptions” to the syntactic “weak
pronoun fronting” that seems to be responsible for making pronominal clitics targets
for Wackernagel’s Law-type effects in archaic IE languages. If the tvā of (25a) were to
represent one of these exceptions, and thus be unfronted, the most likely input structure
for the prosody would have been:
[Top índro vidvā́m̐] [CP hí [ánu tvā cacákṣa]]
Which would have been operated on by the prosody so as to create:
(29) [Top índro vidvā́m̐] [CP hí [ánu hí tvā cacákṣa]]
These two possible analyses have quite different implications for how the system I
have assumed gives rise to clitic sequencing. In the analysis in (26) we would be looking
at the effects of iterative prosodic inversion events, and the examples would reveal an (as
far as I can see somewhat unexpectedly) “outside-in” processing (hí flips in first, then
Under the analysis in (28), we are looking at the relationship between the resolution
of the hosting needs of an unmoved tvā relative to the ability of an inverting hí to “slip
in” between tvā and tvā’s potential (and ultimate) host ánu. The details of the processes
involved under the latter set of assumptions are too complex for me to deal with in this
context, but there is evidence that that approach does represent the correct analysis of
examples such as (25ab).
Recall our earlier discussion of the Attic Greek mén gár clitic sequence. gár is of
course essentially isofunctional with Rigvedic hí. In addition, Thomson’s (1939) wellknown paper on the “postponement of interrogatives” in Attic drama supports the idea
that one could still front into a high Top position in this language. This has a specific
entailment, since the WH-elements Thomson talks about are in CP, and the topics he
deals with are higher, and since gár is in C, there should be Attic drama cases exactly
like the “postponed” subordinating ca, céd, and hí examples we walked through earlier.
And there are.7
a. pròs
mḕ psaúsēi
in light of these-APl Neg should touch any-NSgcl Greek-GPl
Note that in (30c) the articular infinitive construction has a proclitic article, and the first prosodic word
after which gár ‘flips’ is thus tō̃ i=plouteĩn.
14 Preliminaries to the investigation of clitic sequencing in Greek and Indo-Iranian
[spʰagē̃i paréksō gàr dérēn
knife-DSg I submit to gár neck-ASg bravely
‘In light of these things, no one of the Greeks need touch me, because I will
bravely submit my neck to the knife.’
Eur. IphA 1559–1560
b. hōn
oúte métron
oút᾿ aritʰmós estí moi·
which-GPl PTCL Neg measure-NSg Neg number is me-DSgcl
gàr eis hámillan
trouble-DSg trouble-NSg gár to competition-ASg comes
‘Of which (woes) there is neither measure nor number for me, because woe
comes into competition with woe.’
Eur. Tro 620–621
c. kaì nḕ dí᾽
eí tí
ésti lampròn
and by Zeus-ASg if something-NSg PTCL is splendid-NSg and
ḕ kʰaríen
antʰrṓpoisi, dià
or elegant-NSg men-DPl through you-ASg it comes about
gár estʰ᾽ hupḗkoa.
everything-NPl Art-DSg=being rich-INF gár is subservient-NPl
‘and, by Zeus, if something is splendid and beautiful, or elegant for men, it
comes about through you (=Wealth), because everything is subservient to
being rich.’
Ar. Plutus 144–146
In cases involving both mén, which marks focus,8 and gár, meaning ‘because’, the interpretation of scope within the clauses indicates that we are dealing with a structure
such as ‘because (gár) one the one hand (mén) … on the other hand (dé) …’ When there is
nothing for the gár to lean leftwards on, we get the surface order mén gár. This indicates
sequential “prosodic inversion” of the form:
(31) [CP gár [mén [X´ mén gár Ý Ź]]]
But this is an “inside-out” (mén first, then gár) resolution of the hosting needs of these
elements. If the Vedic mechanisms are the same – and all indications are that they are
– then this is clear evidence against the analysis in (26), favoring the (28) analysis. The
implications of this prosodic “tucking in” have not been explored in any significant detail.
8 Conclusions
If we tie the domain of a clitic like hí or gár to its semantic scope – which we can easily
do by positioning it via the syntax, whose job, after all, is to create precisely these kinds
The particle mén normally has a contrasting element, marked by the particle dé. I translate the contrastive
relationship between these two elements as ‘on the one hand X, on the other hand Y’ below.
Mark Hale
of scope relations – we need not worry about finding the structure for it to be in “second
position” in. If the clitic cannot be hosted on its left in situ, that structure will be the
one which provides the nearest prosodic host to the right of the syntactic position of the
clitic, regardless of what entity that is.
As with other syntactic entities that take arguments, we sometimes do have to specify
the nature of those arguments. But it doesn’t follow from that that we need to do it
stipulatively – it isn’t chance that there is no word-level hí or gár ‘because’ clitic. A word
doesn’t express the kind of things ‘because’ needs to take as an argument to generate a
coherent semantics.9 But that same word would work fine as an argument of ‘and’.
Given a sufficiently rich understanding of the semantics of a particular enclitic or postpositive, we should be able to deduce the nature of the kinds of syntactic entities it can
take as an argument. No stipulation should be needed. Of course we are far from having
this kind of understanding of the meaning of many Vedic and Attic Greek enclitics.
To the extent we can determine with some degree of confidence the syntactic position
of the enclitic elements we are interested in, we are in an excellent position to examine
their surface position (which may be the same as their syntactic position, but may be
perturbed by the “prosodic inversion” process) with a view to determining the detailed
mechanics of the interactions involved when multiple 2P elements are present in a string.
The more explicit a conception we have of the relevant algorithms, the easier this task
will be. One of the strengths, in my view, of the model assumed here is that its parts are
all clearly enough defined that it should be easy to discover those instances, if any, in
which stipulation may be, unfortunately, required.
By contrast, approaches which leave vague the processes that give rise to clitic sequencing are revealing in that shortcoming their general inadequacy. Getting prosodic
positioning to interact in the required way with, on the one hand, syntactic positioning
(which all grammatical theories require) and, on the other, with semantic interpretation
(ditto), is a very non-trivial problem: it goes to the core architecture of the grammar.
Working out the details of one’s assumptions in this domain cannot be left as an exercise to future work – one needs to be formulate a clear notion about such things going
in. When there are multiple clitics we see overtly the failure of inexplicit models (such
as Hale 1987, and a lot of subsequent work), but those same problems are present, if
obscured, in the case of simple clitics as well.
Yes, I know about the prepositional ‘because’ phenomenon. If you think about what such strings mean,
and assume that their meaning is representationally present (but not all pronounced), as in
Q: Who slew Vrtra?
A: Indra.
in which ‘Indra’ means ‘Indra slew Vrtra’ (because it can be a lie, and only propositions can be false, not
nouns), then you’ll see why I don’t think this is a problem. Anyway, there’s no evidence that the speakers
of Rigvedic Sanskrit could say: ‘Indra slew Vrtra. Because, the waters.’
14 Preliminaries to the investigation of clitic sequencing in Greek and Indo-Iranian
Anderson, Stephen R. 2005. Aspects of the theory of clitics. Oxford: Oxford University
Bartholomae, Christian. 1886. Arische Forschungen, Bd. II. Halle.
Goldstein, David M. 2016. Classical Greek syntax: Wackernagel’s Law in Herodotus. Leiden:
Hale, Mark. 1987. Studies in the comparative syntax of the oldest Indo-Iranian languages.
Harvard University PhD thesis.
Hale, Mark. 1996. Deriving Wackernagel’s Law: Prosodic and syntactic factors determining clitic placement, in the language of the Rigveda. In Aaron L. Halpern & Arnold M.
Zwicky (eds.), Approaching second: Second position clitics and related phenomena, 165–
197. Stanford: CSLI.
Halpern, Aaron. 1995. On the placement and morphology of clitics. Stanford: Center for
the Study of Language (CSLI).
Hettrich, Heinrich. 1988. Untersuchungen zur Hypotaxe im Vedischen. Berlin: de Gruyter.
Hock, Hans Henrich. 1996. Who’s on first? Toward a prosodic account of P2 clitics. In
Aaron L. Halpern & Arnold M. Zwicky (eds.), Approaching second: Second position clitics and related phenomena, 199–270. Stanford: CSLI.
Jamison, Stephanie & Joel Brereton (eds.). 2014. The Rigveda. Oxford: Oxford University
Keydana, Götz. 2011. Wackernagel in the language of the Rigveda: a reassessment. Historische Sprachforschung 124. 80–107.
Lowe, John J. 2014. Transitive nominals in Old Avestan. Journal of the American Oriental
Society 134(4). 553–577.
Thomson, George. 1939. The postponement of interrogatives in Attic drama. CQ 33(3/4).
Wackernagel, Jacob. 1892. Über ein Gesetz der indogermanischen Wortstellung. Indogermanische Forschungen 1. 333–436.
Wackernagel, Jacob. 1920. Vorlesungen über syntax: mit besonderer berücksichtigung von
griechisch, lateinisch und deutsch. Birkhäuser.
Chapter 15
Nominal verbs and transitive nouns:
Vindicating lexicalism
Paul Kiparsky
Stanford University
Event nominalizations and agent nominalizations provide evidence that all affixation is morphological, and that phrasal categories are projected from words in the syntax. Departing
from both transformational and earlier lexicalist approaches to nominalizations, I first argue on the basis of English and Finnish evidence that gerunds are not DPs built on heads
that embed an extended verbal projection (Baker & Vinokurova 2009; Kornfilt & Whitman
2011), but IPs that need Case. They are categorially verbal at all levels of the syntax, including having structural subjects rather than possessor specifiers. Their nominal behavior is
entirely due to the unvalued Case feature borne by their Infl head, which they share with all
participial verb forms. I then argue that agent nominalizations are categorially nominal at
all levels of the syntax, and that the verb-like case assignment of transitive agent nominalizations is due to the verbal Aspect feature borne by their nominalizing head. Vedic Sanskrit,
Northern Paiute, and Sakha evidence is shown to favor this analysis over B&V’s analysis
of intransitive agent nominalizations as nominal equivalents of Voice heads and transitive
agent nominalizations as Aspect heads. The two “mixed” categories – gerunds and transitive
nominalizations – thus prove to be formally duals: respectively verbs with Case and nouns
with Aspect.
1 Nominalizations
The earliest generative work derived all nominalizations syntactically (Chomsky 1955;
Lees 1960). Chomsky (1970) then argued that only -ing gerunds are derived syntactically,
while all other types of event nominals, such as refutation, acceptance, refusal, are derived morphologically in the lexicon from bases that are unspecified between nouns and
verbs. The suffix -ing was shown to serve both as the gerund formative and as one of
the formatives that derive lexical event nominals. Chomsky’s main argument was based
on the fact that gerund phrases have the structure of verb phrases whereas other event
Paul Kiparsky. 2017. Nominal verbs and transitive nouns: Vindicating lexicalism. In
Claire Bowern, Laurence Horn & Raffaella Zanuttini (eds.), On looking into words (and
beyond), 311–346. Berlin: Language Science Press. DOI:10.5281/zenodo.495452
Paul Kiparsky
nominals have the structure of noun phrases.1 The differences are completely systematic.
Unlike derived event nominals, gerunds are modifiable by adverbs, assign structural case
to their complements (Table 1a,b), disallow articles and other determiners (Table 1c) and
plurals (Table 1d), allow aspect (Table 1e) and negation (Table 1f), and they have a grammatical subject which is assigned a Th-role as in finite clauses (Table 1g), and which may
be an expletive (Table 1h). In (Table 1g), reading refers to an event and her to its agent,
the reader. In the derived nominal, her could also be a sponsor, organizer, or some other
participant of a reading event not necessarily identical with the reader (Kratzer 1996;
2004), and reading could also mean ‘manner of reading’, ‘interpretation’. Without an
of complement, derived nominals are also interpretable in a passive sense: in Mary’s
confirmation, Mary could be the confirmer or the confirmee.
Table 1: Gerunds vs. Nominals.
(-ingV )
*her quick signing the
her immediately reciting it
Determiners * the/a/this/each performing it
(-ingN , -ion, -al, -ance…)
her quick signing of the
* her immediately recital of it
the/a/this/each performance
of it
her readings of it
* her readings it /
* her reading its
by her having sung it
* by her having sung of it
by her not approving it
* by her not approval of it
we remembered her reading it we remembered her reading
of it
it(s) seeming to me that I
*it(s) appearance to me that I
The lexicalist line of analysis continues to be developed in different ways (Malouf
2000; Blevins 2003; Kim 2016). But many recent treatments have reverted to a uniformly
syntactic derivation of nominalizations, in which nominalizing heads project a nominal
structure and have a verbal complement whose type determines the nominalization’s
properties. The differences between the two types in Table 1 is captured by introducing
Chomsky also contrasted the uniformity, regularity and full productivity of gerunds with the morphological and semantic diversity, idiosyncrasies, and limited productivity of derived event nominals. As Anderson
2016 notes, these points played a subsidiary role in Chomsky’s argument. Indeed, they are not compelling
criteria by themselves, for there is no shortage of productivity and regularity in the lexicon, and syntax
has its share of idiosyncrasy.
2 I can’t help but feel a little despondant due to it seeming to me that the TIE/fo and the T-70 make the original
TIE and T-65 somewhat redundant (Internet), evidence that “explains away” its seeming to me that p is the
case (James Pryor, The Skeptic and the Dogmatist, Noûs 34: 534, 2000). The variation between Poss-ing and
Acc-ing gerunds seen here is briefly addressed in 2.3 below.
15 Nominal verbs and transitive nouns: Vindicating lexicalism
them at different levels in the functional structure. The gerund -ingV is structurally high,
and derived nominals in -ingN , -ion, -al, -ance are structurally low.
Kornfilt & Whitman (2011) dub this the Functional Nominalization Thesis (FNT),
and propose a typology of four levels of nominalization, CP, TP, vP and VP. In this
typology, English gerunds are TP nominalizations, while derived nominals are VP nominalizations.3 This paper vindicates a uniform treatment of nominalizations in a different
way: all true nominalizations are derived lexically; gerunds are not nominalizations at
all – they are neither DPs nor NPs but IPs that need Case.
As my point of departure I take Baker & Vinokurova’s (2009) theory, which extends
the FNT from event nominalizations to agent nominalizations. For gerunds, B&V posit
the structure (1a), based on the version of the DP analysis originated (along with the DP
itself) by Abney 1987. The DP’s complement here is an NP headed by the gerund nominalizer -ingV , below which the structure is entirely verbal: an AspP which hosts aspectual
material and certain adverbs, and which has a vP complement whose v (v=Voice) head
assigns structural case and introduces an external agent argument. This external agent
argument shows up as a genitive in D head. B&V do not say exactly how it gets there
in (1a); perhaps it is base-generated in D and bears a control relation to the PRO in the
Spec-vP position where the agent role is assigned.
For derived nominals B&V propose the structure (1b), where the head (-ingN , -ion, etc.)
takes a bare VP complement. Because it has no Asp or v projection, it contains neither
adverbs, agents, nor structural case.4
The structures in (1) take care of the contrasting properties Table 1b, e, f, and g, but
leave the remaining four properties (Table 1a, c, d, and h) to fend for themselves. On the
one hand, the DP in (1a) provides too little structure: expletive it-subjects are believed to
occupy Spec-IP or Spec-TP, but (1)a provides no Spec-IP or Spec-TP for them.5 On the
other hand, the DP, needed in the analysis as a site for the gerund’s subject, generates
unwanted structure. Since DPs can have plural heads, adjective modifiers, determiners,
and quantifiers, the DP analysis wrongly predicts that Table 1a, c, and d should be grammatical. To maintain it one must somehow prevent functional projections like AP, QP,
and NumP from appearing in DPs that have NP complements that have AspP complements, while allowing them in other kinds of DPs, and one must prevent the head of a
A syntactic derivation of gerunds from TP/IP is also developed by Pires 2006. The aspectual content of the
gerund is treated in Pustejovsky 1995; Alexiadou 2001, and Alexiadou, Soare & Iordăchioaia 2010. Alexiadou
and her co-workers conclude that gerunds are imperfective Aspect heads that dominate VoiceP and vP,
while nominalizers are n heads that also dominate VoiceP and vP, but under NumberP and ClassifierP,
housing adjective modifiers, determiners, and plural.
4 All analyses have to contend with the fact that certain adverbs can occur as postmodifiers with derived
nominals, and even with some underived ones (Payne, Pullum & Huddleston 2010); they cite examples such
as the opinion generally of the doctors, a timber shortage nationally, the people locally, and the intervention
again of Moscow. We shall see similar Finnish data in §3 below.
5 Expletive there, which likewise appears in gerunds, may sit in a lower subject position, since it is sensitive
to the argument structure of the predicate – the absence of Cause according to Deal 2009, who puts it
in the specifier of v. Like expletive it, there does not appear in derived nominals (*there’s appearance to
be a problem). On Deal’s analysis, the distribution of expletive there is consistent with my IP analysis of
gerunds, but adds no further support to it.
Paul Kiparsky
DP whose complement is an NP whose complement is an AspP from being an article or
a demonstrative pronoun.6
a. his reading the book (high)
b. the reading of the book (low)
of the book
the book
Contrary to what the FNT seems to promise, then, the morphosyntactic properties of a
nominalization cannot be fixed just by locating its nominal head in a universal hierarchy
of verbal functional categories, or even in a language-specific one. In that approach to
mixed categories, it seems that the functional content that a given nominalizing head
may combine with must be specified on an item-specific basis. But not just any arbitrary
mixed category is possible. Consider the awesome unused power unleashed by the FNT.
If functional N heads can convert AspPs into NPs in the syntax, as in (1a), why aren’t
there such things as Q heads with vP complements (*[some [he read it]vP ]QP ), let alone
multiple verbalizing and nominalizing syntactic heads interspersed to generate phrases
in which layers of verbal and nominal structure alternate in various combinations?
Some of the overgeneration could be curbed by by eliminating the DP layer, or by eliminating the NP layer
and having D select for AspP directly. But these projections cannot be struck from (1) because their heads
are essential to the analysis. The D head serves as the site of the structural subject, and the N head houses
the nominalizer -ing. Neither of these elements can be accommodated in the Asp head, for that is required
for the aspectual auxiliary have.
15 Nominal verbs and transitive nouns: Vindicating lexicalism
The empirical problem of overgeneration is a direct result of the theoretical approach
behind the FNT-style analysis. The derivation of gerunds in (1a) involves syntactic affixation of -ing to the phrasal projection AspP.7 A lexicalist perspective rules out affixation
to phrases. It dictates an entirely different kind of derivation, in which the gerund suffix
-ing is added to verbs in the morphology to build words (e.g. reading), which are then
inserted in terminal nodes in the syntax. On this view, a gerund phrase is the syntactic
projection of a gerund, not of a determiner as in (1a). On that basis we can build a simple
and restrictive theory of nominalizations that explains all the data in Table 1.
The key idea is that gerunds are participles, and that participial suffixes, -ing included,
are Infl heads that differ from finite and infinitive Infl heads in that they bear a Case
feature. The extended projection of a gerund is then an IP with a Case feature, which
needs to be checked (or, from a non-lexicalist perpective, valued) in the syntax. The Case
feature restricts participial phrases to two syntactic functions: arguments – gerunds – in
positions where their value for Case can be checked by a predicate, and participial modifiers in positions where their value for Case can be checked by head-modifier agreement.
Lexicalism excludes not only FNT-style analyses of gerunds, but every kind of syntactic affixation to phrasal categories. This means that no syntactic process can have the
effect of changing the category of a word. That holds for all types of nominalization:
event nominalizations, result nominalizations, and agent nominalizations. All “mixed
categories” must then arise from morphological specifications of lexical heads, rather
than from syntactic embedding as in (1). In §3 I support this more general prediction by
showing that transitive agent nouns do not have an embedded vP projection and that
their verbal properties come from a Tense/Aspect feature on the agent suffix.
I assume that a phrasal constituent is a projection of its head, which inherits its category (Noun, Verb, etc.), its inflectional features (such as Aspect and Case), and its thematic roles (Agent, Patient, Instrument, Event, etc.).8 Mixed categories are verbs, nouns,
and adjectives that have an extra phi-feature. Their extended projections behave like
extended projections of ordinary verbs, nouns, and adjectives, modulo the properties enforced by that feature content. The language-specific syntax of gerunds is determined by
their Case feature. A gerund that can bear any Case projects a phrase with the distribution of a DP. A gerund that has a partially specified Case feature projects a phrase that is
restricted to positions compatible with the specified values of the feature. For example,
Finnish gerunds are restricted to internal argument positions (section 2.2). Similarly, the
verbal properties of transitive agent nouns are due to a Tense/Aspect feature assigned
to these nouns by the agent affix that forms them. This feature may likewise be lexically
unvalued and specified by additional aspectual morphology (as in Northern Paiute), or
inherently specified on the agent noun affix itself (as in Sanskrit and Sakha), see 3.3. Since
the mixed categories under lexicalist assumptions are projected from a single head, we
correctly predict the absence of mixed categories in which verbal and nominal structure
A similar earlier proposal is Yoon (1996).
E.g. ⟦-er⟧ = λPλxλe[P(e) ∧ Agent(e,x)] (the set of human individuals that are the Agent of some event),
⟦-ee⟧ = λPλxλe[P(e) ∧ human(x) ∧ Undergoer(e,x)] (the set of human individuals that are the Undergoer
of some event).
Paul Kiparsky
is alternately layered in weird combinations, of vPs that function as DPs, and of the other
abovementioned monstrosities.
A theoretical gain is that we need not divide nominalizations into a syntactic type and
a lexical type, as in standard lexicalist analyses. Once gerunds are recognized as IPs, we
can maintain that all nominalizations are derived morphologically in the lexicon. This
can be done either in a realizational morphology of the type pioneered by Anderson
(1992), or in a morpheme-based one such as the minimalist morphology of Wunderlich
(1996a). It remains to be seen whether the analysis can be recast in a DM-friendly syntaxbased format. What is clear is that it does not follow from any theory that countenances
structures like (1). To that extent at least, its empirical success constitutes new empirical
support for lexicalism.
I begin in §2 with “high” event nominalizations. I show that the lexicalist approach
correctly predicts the syntax of English and Finnish gerund phrases, including aspects
that go unexplained in FNT analyses, and that it curbs the typology in a good way. In
§3 I apply the same idea to agent nominals, and support the resulting analyses with data
from Vedic Sanskrit and Finnish that is new to the theoretical literature.
2 Gerunds
2.1 English gerunds
Gerunds and participles are formally identical in English (Pullum 1991; Yoon 1996; Huddleston & Pullum 2002; Blevins 2003). For example, they are the only verb forms that
overtly distinguish perfect aspect but not progressive or past tense. Given the modest
morphology of English this identity might be dismissed as an accident, but the testimony
of richly inflected languages, such as Finnish (2b), Classical Greek (2c), Sanskrit (2d), and
Latin (2e) leaves no doubt that participles are systematically used in two functions: adjectivally as modifiers and nominally as arguments.
a. English -ing participle
i. Modifier: I saw Bill reading the book. (⇒ I saw Bill.)
ii. Argument: I hated Bill’s reading the book. (⇏ I hated Bill.)
b. Finnish participles
i. Modifier:
hunaja-a syö-vä-n
remember-pst-1sg honey-prtc eat-ptc-gen bear-acc
‘I remembered the/a bear (that was) eating honey.’
ii. Argument:
karhu-n syö-vä-n
remember-prtc-1sg bear-gen eat-ptc-gen honey-prtc
‘I remembered that the/a bear ate honey.’
15 Nominal verbs and transitive nouns: Vindicating lexicalism
c. Classical Greek participles
i. Modifier:
Phílippo-n apéktein-a
the-acc act-unjustly-prtc-acc Philip-acc kill-aor-1sg
‘I killed the unjustly acting Philip.’
ii. Argument:
Phílippo-n eksḗlenk-s-a
act-unjustly-prtc-acc Philip-acc prove-aor-1sg
‘I proved that Philip acted unjustly.’
d. Sanskrit participles
i. Modifier:
rājān-am ā-ga-ta-ṃ
king-acc to-go-prtc-acc hear-pres-1sg
‘I hear the king (who has) arrived.’
ii. Argument:
rājān-am ā-ga-ta-ṃ
king-acc to-go-prtc-acc hear-pres-1sg
‘I hear that the king has arrived.’
e. Latin participles
i. Modifier:
ad Antiochu-m confug-i-t
Hannibal.nom defeat-prtc.masc-nom to Antiochus-acc flee-perf-3sg
‘Defeated, Hannibal took refuge with Antiochus.’
ii. Argument:
Hannibal.nom defeat-prtc.masc-nom fear.abl
‘Hannibal’s being defeated freed the Romans from fear.’
Traditional grammars of these languages treat participles as verb forms which are inflected for Case, for good reasons. Participles distinguish the verbal categories of voice
and tense/aspect, and they are formed off the same tense/aspect stems as the finite verbs.
They supply the periphrastic forms that complete gaps in inflectional paradigms. They
assign the same cases to their objects as the corresponding finite verbs and infinitives
do. They are modified by adverbs, not by adjectives. They select for the same prefixes as
the corresponding finite verbs and infinitives, with the same (often idiosyncratic) meanings. Those languages that disallow noun+verb compounds (such as classical Greek and
Sanskrit) also disallow noun+participle compounds. As I show below, participles have
structural subjects.
Paul Kiparsky
So there must be some property that distinguishes participles from finite verbs and
infinitives, and which supports the double function of participles as nominal arguments
and adjectival modifiers. The obvious candidate is Case. Suppose then that participial
formatives are Infl heads that need Case. On lexicalist assumptions, they are affixed in
the morphology to a verb to make a participle, which is then inflected for case if the
language has case morphology, and enters the syntax with a specified Case feature that
– like any Case feature – must be checked in the syntax. In a language that lacks case
morphology, such as English, the participle remains unvalued for Case, and projects an
IP with a Case feature that must be valued in the syntax. Both “checking” and “valuing”
can be formalized as identical operations of feature unification, or as optimal matching
in OT Correspondence Theory.
As an illustration consider first the derivation of gerunds in English. Prescinding from
vP, AspP, VoiceP, and other possible functional projections, their syntactic structure is
as in (3).
The man’s
Infl[uCase] combines with V in the same way as Tensed Infl does. How this happens depends on the model of grammar. If we assume both lexicalist syntax and lexicalist morphology, the case-needing Infl -ing is suffixed to V in the morphology to form a participle,
and the participle then projects a case-needing IP in the syntax, where the Case feature is
valued. In argumental participles (gerunds), it is valued by the governing Case-assigner,
and in participial modifiers it is valued by agreement with the nominal they modify.
If we assume minimalist syntax, we can comply with lexical morphology by using
spanning (Svenonius 2016), which allows the lexically generated participle to be inserted
under the two corresponding syntactic terminal nodes. In DM, -ing would be a syntactic
terminal that is postsyntactically Lowered onto V. Thus, the idea that gerunds are caseneeding IPs can probably be implemented in any grammatical architecture. However,
in non-lexicalist frameworks this analysis is merely motivated on empirical grounds.
In lexicalist frameworks that prohibit affixation to phrases it is required on principled
grounds as well.
In languages where participles are morphologically inflected for Case, such as (2b-e),
participles project an IP that bears a specified Case value that must be checked in the
15 Nominal verbs and transitive nouns: Vindicating lexicalism
syntax. The lexicalist approach now makes an interesting prediction: in such languages,
gerunds may be morphologically restricted to particular Case features, which restrict
their syntactic distribution to contexts compatible with those features. This prediction
is confirmed in Finnish. Finnish gerunds bear an oblique Case – glossed as Genitive in
(2b.ii) – which confines them to internal argument positions (section 2.2).9
Returning to English, the analysis of participial clauses as IPs that have the single
special property of needing structural Case explains at a stroke which clausal and nominal properties they have and which ones they lack. To start with the latter, it explains
why gerund phrases, unlike DPs, have no articles, quantifiers, or numerals, why they
cannot be modified by adjectives and relative clauses, why their head cannot be genitive
or plural.
* The/a/each compiling the corpus took over a year.
* Both/every compiling corporas took over a year.
* His two compilings corpora each took over a year.
* His careful compiling the corpus was a turning point.
* His editing texts that is funded will take a year.
* His compiling corpora’s results were dramatic.
The missing categories are just the ones that would originate in a DP.
As for the nominal properties that gerunds do have, they are accurately covered by the
generalization that gerunds appear in Case positions. They function as subjects, objects,
and predicates, as objects of prepositions (6e,f,g), and as objects of a small set of transitive
adjectives (6f,g), all diagnosed as Case positions by the fact that full-fledged DPs occur
in them.
a. [ Bill’s leaving Paris ] was unexpected.
b. I regret [ Bill’s leaving Paris ].
c. The problem is [ Bill’s leaving Paris ].
There is no agreement relation between the genitive subject and the gerund in (4).
Paul Kiparsky
Because of Bill’s leaving Paris we’ll be hiring new personnel.
We are worried about Bill’s leaving Paris.
This event is worth my visiting Paris.
It’s no good my playing this sort of game.10
This does not mean that all transitive verbs take gerund complements. Particular verbs
can select for whether they take gerunds, that-clauses, or infinitive complements, just
as they can select for whether they take DPs:
a. * I said Bill’s leaving Paris.
b. I said it/something/several things.
What the analysis correctly predicts is that gerunds, unlike that-clauses and infinitives,
appear only in Case positions:
b. i.
c. i.
* I hope Bill’s leaving Paris.
* I hope it.
I hope that Bill is leaving Paris.
I hope to leave Paris.
* It is rumored Bill’s leaving Paris.
* The proposal is rumored.
It is rumored that Bill is leaving Paris.
It is rumored to be happening.
* It seemed / was expected Bill’s leaving Paris.
* It seemed / was expected this event.
It seemed / was expected that Bill would leave Paris.
It seemed / was expected to happen.
A further consequence is that the subjects of gerunds are IP specifiers. If overt, they
are Genitive or Accusative,11 just as the subject of a finite IP is Nominative, and the
overt subject of an infinitive requires a Case-assigning for. Crucially, they are true structural subjects analogous to subjects of finite clauses, not necessarily “agents” as in B&V’s
Table 1a, nor “possessors” with their varied functions as in derived nominals. This prediction is confirmed by three generalizations. Unlike genitive specifiers of nouns (including
derived nominals), but like structural subjects of finite clauses, the specifiers of gerunds
can be expletives:
It(s) seeming to you that you dreamt is not evidence of it(s) being the case
that you dreamt.
b. * It(s) appearance to you that you dreamt is not evidence of it(s) truth that you
Cf. It’s no good this sort of game. (Dickens, Our Mutual Friend)
On Acc+ing gerunds see the brief and inconclusive remarks in §2.3 below.
15 Nominal verbs and transitive nouns: Vindicating lexicalism
Like structural subjects of finite clauses, they are subject to control (Huddleston & Pullum 2002: 1190):
a. Mary remembered locking the door. [the rememberer is the locker]
b. Mary remembered the/a locking of the door. [the rememberer might not be
the locker]
like structural subjects of finite clauses, and unlike genitive agents of nominals, they
cannot be paraphrased with of or by:
the Persians’ quick run = the quick run of/by the Persians
the Persians’ quick running = the quick running of/by the Persians
the Persians’ quickly running ≠ *quickly running of/by the Persians
the Persians’ quickly attacking the Greeks ≠ *quickly attacking the Greeks
of/by the Persians
e. the Persians quickly attacked the Greeks ≠ *of/by the Persians quickly
attacked the Greeks
To summarize: the analysis of gerunds as IPs with Case explains the cross-linguistically common convergence of nominal and adjectival functions in a single morphological
class of verbal forms. By not positing any DP or NP structure over the IP it avoids the
overgeneration problem that FNT-type analyses face. It excludes the possibility of multiple alternating verbalizing and nominalizing syntactic heads to which the FNT opens
the theoretical door, and gets rid of the constraint that heads of DPs whose complements
are NPs whose complements are verbal projections may not be articles or demonstrative
pronouns. It provides the basis for a uniform structure for all DPs, and for a uniform lexical derivation of all nominalizations. It correctly predicts that gerunds and participles
have subjects – specifiers of Infl that are structural counterparts to the subjects of finite
clauses. What is important is that the analysis is not motivated merely by these empirical arguments; it is a consequence of lexicalism, and, if correct, supports the lexicalist
organization of grammar.
The question arises whether there might be a CP layer above the IP, headed by a
null complementizer. This additional structure is not justifiable for English, because the
distribution of gerunds differs from that of any type of CP. First, gerunds need Case,
whereas CPs do not (Vergnaud 1977). Secondly, gerunds are permitted in clause-medial
position, while that-clauses and other CP clauses must extrapose.12
That gerund phrases are full IPs with a structural subject, that they bear Case, and
that, unlike derived nominals, they have no DP or NP projection, and in particular no
possessor-type Specifier, makes many additional predictions that are testable in morphologically richer languages. They turns out to be abundantly supported, as demonstrated
for Finnish in the next section.
However, the case marking of participles in inflected languages could be considered as a kind of complementizer, as conjectured for the inherent case affix -n on Finnish gerunds in §2.2.
Paul Kiparsky
2.2 Finnish gerunds are IPs
Finnish participial propositional complement clauses are the closest functional counterparts of English gerunds, and I will call them gerunds here. They are not DPs with
possessors but IPs with true structural subjects. Their Case is inherently marked by the
oblique suffix -n, arguably functioning as a complementizer, which restricts them to internal argument positions. This illustrates how the typology of gerunds emerges from
variation in what cases they can bear.13
Unlike English gerunds, Finnish gerunds are never external arguments. Thus they can
be objects of transitive verbs such as say, think, want, prove, remember and hear, and
subjects of presentational intransitives like appear and become evident, but they cannot
be subjects of such predicates as be obvious, prove, and mean.14
become-clear-pst.3sg Matti-gen shoot-perf.prtc-gen bear-accgen
‘It became clear that Matti had shot the/a bear.’
b. * Mati-n
Matti-gen shoot-perf.prtc-gen bear-accgen anger-pst.3sg
‘That Matti had shot the/a bear angered Liisa.’
This distribution suggests that the ending -n that participles bear in their gerundial function, glossed “gen” in (12), marks an object Case that is compatible with internal arguments but not with external arguments. Historically, it is probably the old dative ending,
which has fallen together phonologically with the genitive, but persists as a morphosyntactically distinct type of genitive which (unlike the structural genitive) cannot function
as a subject (Kiparsky In press).
As shown in (13), Finnish gerunds behave more like bare finite CP clauses with että(that-) than like DPs, whether nominal DPs (13c) or pronoun-headed finite clauses with
se että- (it that-) (13d).
a. Huomas-i-n / ymmärrä-n
/ luule-n / otaksu-n tilante-en
notice-1sg / understand-1sg / think-1sg / assume-1sg situation-gen
be-prtc-gen difficult-gen
‘I noticed / understand / think / assume that the situation is difficult.’ [lit. ‘the
situation’s being difficult.’]
The data and analysis of Finnish gerunds presented in this section is condensed from my treatment of
Finnish nonfinite complementation in Kiparsky (In press), to which I refer the reader for the details.
In the glosses, accgen and accnom both refer to morphosyntactic Accusative structural case. The subscripts
show three different morphological case realizations of this morphosyntactic Case. They will become important shortly, but for now the reader may ignore them.
15 Nominal verbs and transitive nouns: Vindicating lexicalism
b. Se-n huomat-tiin
/ ymmärre-tään / luul-laan / otaksu-taan
It-gen notice-pass-pst / understand-pass / think-pass / assume-pass
be-prtc-gen difficult-gen
‘It is noticed / understood / thought / assumed to be difficult.’
c. Huomas-i-n / ymmärrä-n
/ *luule-n / *otaksu-n häne-t
notice-pst-1sg / understand-1sg / think-1sg / assume-1sg him-accacc /
that-accgen thing-accgen
‘I noticed / understand / think / assume him / this point (fact).’
d. Huomas-i-n / ymmärrä-n
/ *luule-n / *otaksu-n se-n,
notice-pst-1sg / understand-1sg / think-1sg / assume-1sg it-accgen that
on hankala.
situation.nom is difficult
‘I noticed / understand / think / assume that the situation is difficult.’
The distinction between verbs that allow DP objects (huomat- ‘notice’ and ymmärtä- in
(13)) and verbs that do not allow DP objects (luule- ‘think’ and otaksu- ‘assume’ in (13))
is correlated with factivity, but the correlation is not exact and my argument does not
depend on it.
Since gerunds are not DPs but case-marked IPs, their genitive subjects behave like
structural subjects and not like genitive specifiers of DPs. This is shown by five arguments.
The first argument that the genitive specifier of gerunds is a grammatical subject is
that it gets assigned exactly the same Th-roles as the subjects of the corresponding finite
clause, not the diverse range of interpretations that “possessors” of derived nominals
receive (see above under Table 1). So Matin in (12a) picks out the agent of the shooting
event, whereas the specifier Matin of the derived nominal (14) could be, among other
things, the organizer or theme of the rescue.
remember.pres-1sg Matti-gen rescue-nom-acc
‘I remember Matti’s rescue.’
The second argument comes from extraction. The subjects of gerunds can be extracted
as readily as objects:
a. Kene-n väit-i-t
ampu-nee-n hän-tä?
who-gen claim-pst-2sg shoot-pfp-gen he-part
‘who did you claim shot at him?’
b. Ke-tä
häne-n ampu-nee-n?
who-part claim-pst-2sg he-gen shoot-pfp-gen
‘who did you claim he shot at?’
Paul Kiparsky
But possessors cannot be extracted (16a), and neither can genitive specifiers of tenseless nonfinite complements such as the third infinitive (16b) and the second infinitive
(16c) (the Left Branch Condition, Ross 1967: 127).
a. * Kene-ni väit-i-t
t i karhu-n / että
who-gen claim-pst-2sg shoot-perf.prtc-gen bear-gen / that
t i karhu?
shoot-pass.pst bear.accnom
‘Whose bear did you claim (that) was shot?’
b. * Kene-ni väit-i-t
ei ampu-ma-n
karhu-n paina-nee-n
who-gen claim-pst-2sg shoot-3inf-gen bear-gen weigh-perf.prtc-gen
500 kilo-a?
500 kg-part
‘The bear shot by whom did you claim weighed 500 kg?’
c. * Kene-ni itk-i-t
ei ampu-e-ssa
who-gen claim–pst-2sg shoot-2inf-iness bear-accgen
‘Who did you weep while he shot the/a bear?’
A third diagnostic which shows that gerunds have structural subjects and not possessors is that they do not undergo possessor agreement. Nouns and infinitives agree with
their genitive specifiers, as exemplified for nouns in (17a), for the second infinitive in
(17b), and for the third infinitive in (17c).
a. (Minu-ni ) karhu-nii
(My-gen) bear.nom-1sg weigh-pst-3sg 500.acc kg-part
‘My bear weighed 500 kilograms.’
b. Matti
(minu-ni ) ampu-e-ssa-nii karhu-n
Matti.nom weep-pst.3sg (my-gen) shoot-2inf-1sg bear-accgen
‘Matti wept as I shot the/a bear.’
c. (minu-ni ) ampu-ma-nii karhu
(my-gen) shoot-3inf-1sg bear.nom
‘the/a bear I shot.’
But gerunds do not possessor-agree with their subjects, as we can see in (18a,b) and (with
a raised subject) in (18c).
a. Matti
minu-ni ampu-nee-n
Matti.nom know-pst.3sg me-gen shoot-prf.prt-gen
(*ampu-nee-nii )
(shoot-prf.prt(-gen)-1sg) bear-accgen
‘Matti knows that I’ve shot the/a bear.’
15 Nominal verbs and transitive nouns: Vindicating lexicalism
b. Selvis-i
häne-ni ampu-nee-n
become clear-pst-3sg he-gen (shoot-prf.prt(-gen))
(*ampu-nee-nsai )
(shoot-perf.prtc(-gen)-3p) bear-accgen
‘it became clear that he had shot the/a bear.’
c. Näytä-ti ampu-nee-n
(*ampu-nee-sii )
seem-2sg shoot-prf.prt-gen (shoot-perf.prtc-(gen)-2sg) bear-accgen
‘you seem to have shot the/a bear.’
Of course the subjects of gerunds cannot subject-predicate agree with the gerunds like
nominative subjects of finite clauses agree with the finite verb, for genitive subjects never
subject-predicate agree in Finnish.
The fourth argument that gerunds have structural subjects comes from the distribution
of accusative Case morphology. Descriptively, Finnish morphosyntactic Case is realized
as morphological case as follows.15
a. The subject of a participial clause is always Genitive.
b. The object of a participial clause can be morphosyntactic Accusative or
Partitive. Partitive is assigned to objects under the same conditions as in
finite clauses:
i. Objects under overt or implicit negation are Partitive.
ii. Objects of certain predicates (such as love and touch) are Partitive.16
iii. Otherwise objects are Accusative.
Morphosyntactic Partitive is always realized as morphological partitive. And now comes
the essential and trickiest part. Morphosyntactic Accusative is realized by three morphological cases:
a. as morphological accusative on personal pronouns,
b. otherwise as morphological genitive if the object is plural, or if the clause has
a subject with structural case (this last condition is called Jahnsson’s Rule),
c. otherwise as morphological nominative.
Clause types that lack subjects with structural case for purposes of Jahnsson’s Rule include imperatives, bare infinitives (“to see Naples and to die”), passives (which in Finnish
do not involve “promotion” of the object), and clauses with “quirky case” subjects.
Since the argument to be presented below uses Jahnsson’s Rule as a diagnostic for the
presence or absence of a structural subject, I will gloss the examples in such a way that
the reader can see whether Jahnsson’s Rule has taken effect in them. This means glossing
not only morphosyntactic Accusative Case, but whether morphosyntactic Accusative
For details see Kiparsky 2001; a sophisticated OT treatment of the variation is developed by Anttila & Kim
16 The class of partitive-assigning predicates is often called “telic” (e.g. Kratzer 2002). This is not quite correct;
for an attempt at a more accurate formulation see Kiparsky 2005a.
Paul Kiparsky
Case is realized as morphological accusative case or nominative case. So I will mark
morphosyntactic Case by the main gloss and morphological case with a subscript on it.
For example, in (21) both objects bear morphosyntactic Accusative Case, realized in (21a)
as morphological genitive case and in (21b) as morphological nominative case.
a. Matti
Matti.nom shoot-pst(3sg) bear-accgen
‘Matti shot the/a bear.’
b. ammu
shoot-imper bear-accnom
‘shoot the bear!’
Through the rest of the text in this section I use capitalization to distinguish morphosyntactic Case (such as Accusative) from morphological case (nominative, accusative, etc.).
At last we are ready for the argument. Nonfinite complement clauses are translucent
to the triggering of Accusative and Partitive Case and to the realization of Accusative
case as genitive or nominative, in the sense that (19) and (20) can be conditioned either
within the gerund clause or in the larger domain of the higher clause with its gerund
complement. So in (22a) the object of the lower clause, which contains no negation,
can have either Accusative Case (realized as morphological genitive case by (20a)), or
Partitive Case from the negated main clause by (19b.ii). In (22b) the morphosyntactic
Accusative Case on the object of the gerund is realized either as morphological genitive
case because the main clause has a subject, or as morphological nominative case, because the participle, being passive, is subjectless (Jahnsson’s Rule, (20b)).17 In (22c) the
morphosyntactic Accusative Case on the object can only be realized as morphological
nominative case because both the matrix verb and the participle are subjectless.
a. En
heidä-n ampu-nee-n
Not-1sg know-perf.ptc they-gen shoot-perf.prtc-gen /
/ karhu-a
shoot-prs.prtc-gen bear-accgen / bear-part
‘I didn’t know that they had shot / were (would be) shooting the/a bear.’
b. Ties-i-n
metsä-ssä ammu-tu-n
/ karhu
know-pst-1sg forest-illat shoot-pass.prtc-accgen bear-accgen / bear.accnom
‘I knew a bear to have been shot in the forest.’
c. Eilen
*karhu-n / karhu
Yesterday turn.out-pst.3sg shoot-pass.prtc-gen bear-accgen / bear.accnom
‘It turned out yesterday that a bear was shot.’
The variation between case governed locally within the subordinate clause and in the larger domain that includes the main clause is sensitive to as yet poorly understood semantic, stylistic and discourse factors. The
distribution of the Partitive in particular is affected by factivity and the scope of negation (“Lausetyyppien
käsittely transformaatioteoriassa [The treatment of sentence Types in Transformational grammar]” 1970:
31, Hakulinen & Karlsson 1979: 365). For example, in (22a), the Partitive registers surprise or skepticism,
and in (22b) the Accusative (realized as nominative) is likely to be interpreted factively.
15 Nominal verbs and transitive nouns: Vindicating lexicalism
The crucial case is (23), where the morphosyntactic Accusative Case of the object may be
realized as morphological genitive case. Since the matrix verb is subjectless, the object’s
realization as morphological genitive case must be licensed by the subject of the gerund,
Matin. Therefore the subject has structural Case.
(23) Ilmen-i
/ karhu.
turn.out-pst.3sg Matti-gen shoot-perf.prtc-gen bear-accgen / bear.accnom
‘It turned out that Matti shot the/a bear.’
This completes the fourth argument that the genitive subject of gerunds is a structural
In contrast, the fact that “quirky” genitive subjects induce the nominative form of the
object tells us, by Jahnsson’s Rule, that they are non-structural:
a. Hänen pitää
osta-a auto.
He-gen must-3sg buy-1inf car.accnom
‘He has to buy the/a car.’
b. Hänen on
helppo nosta-a tämä
He-gen be-3sg easy lift-1inf this.accnom sack.accnom
‘It is easy for him to lift this sack.’
This is as expected, since they are not assigned structurally but idiosyncratically by particular predicates.
A fifth argument that gerunds have structural subjects is that they can have a generic
null subject proar b . In Finnish proar b can be a subject (Hakulinen & Karttunen 1973)
but it cannot be a possessor: contrast (25a) and (25b). So, the fact that gerunds can
have a generic proar b subject, as seen in (25c), is another datum in support of the claim
that gerunds have structural subjects and not possessors. Moreover, gerunds can be
subjectless under the same conditions as subjects of finite clauses. For example, gerunds
can have the impersonal passive form, see (25d).
Siellä ∅ voi
there pro can-3sg dance-1inf
‘One can dance there.’
b. * On
mukava katsel-la
∅ valokuv-i-a.
be-3sg nice pro photo-pl.part
* ‘It’s nice to to look at one’s photos.’ (OK without ∅: ‘It’s nice to look at
c. Siellä väite-t-ään
∅ voi-va-n
there claim-pst.pass pro can-pres.prtc-gen dance-1inf
‘It is claimed that one can dance there.’
d. Siellä väite-tt-iin
there claim-pst.pass can-pass-pres.prtc-gen dance-1inf
‘It was claimed that there is dancing there.’
Paul Kiparsky
I conclude that Finnish gerunds are IPs like English gerunds, albeit with a different
syntactic distribution due to their oblique Case specification.
2.3 Desultory remarks on Acc-ing
The English “Acc-ing” construction differs in many ways from the “Poss-ing” gerund
considered here so far. I have no serious analysis of it to offer. Its behavior resembles AccInf (“ECM”) constructions in some ways. First, unlike gerunds with genitive subjects, it
is degraded by intervening adverbs, extraposition, and fronting, under roughly the same
conditions as nominal objects (“Quantification, Events, and Gerunds”; Pires 2006):
a. We anticipated (*?eagerly) him leaving Paris.
b. (We anticipated his resignation, but) *?him/his leaving Paris we did not
This is the same pattern as:
a. We believe (*?strongly) him to have told the truth.
b. (We believed him to have been mistreated, but) *?him to be telling the truth
we did not believe.
Acc-Inf gerunds allow extraction, like Acc-Inf complements and unlike Poss-ing gerunds:
a. Which city do you remember him/*his describing? (Portner 1995: 637, citing
L. Horn)
b. Who do you resent Bill/*Bill’s hitting? (Williams 1975: 263)
c. Who/*whose do you resent hitting Bill? (cf. *Who do you resent (it) that hit
d. Who do you believe to be telling the truth?
e. What do you believe him to be saying?
Another frequently noted difference between the constructions is that the genitive subject of gerunds is preferentially human, and cannot be expletive there at all, whereas the
accusative is unrestricted in this respect, again like Acc-Inf subjects.
a. There (*there’s) being no objection, the proposal is approved.
b. ? I imagined the water’s being 30 feet deep.
Accusative subjects of gerunds do not seem to be getting their case from the main verb,
since they can appear in gerunds that function as subjects. Possibly the accusative case
assigner is a null preposition or complementizer, an analog of the overt for of for-to
15 Nominal verbs and transitive nouns: Vindicating lexicalism
3 Agent nominals, transitive and intransitive
Like the FNT, my alternative theory of nominalizations is in principle applicable to every
type of nominalization, including agent nominalizations and result nominalizations. The
mixed category that most gravely challenges analyses of agent nominals is transitive
agent nouns, which function as nominals except for assigning structural case to their
objects and allowing some adverbial modifiers. I will make a case that, just as gerunds are
categorially verbal at all levels of the syntax and their noun-like behavior is entirely due
to a nominal Case feature borne by their Infl head, such transitive agent nominalizations
are categorially nominal at all levels of the syntax and their verb-like behavior is entirely
due to a verbal feature borne by their nominalizing head, namely Aspect. Gerunds and
transitive nominalizations thus prove to be duals in a sense – respectively verbs with
Case and nouns with Aspect.
I show that this idea predicts the distinction between transitive and intransitive agent
nouns, whereas the functional properties of nominalizations neither correlate with each
other as the FNT predicts, nor match the height of their nominalizing heads in syntax
or word structure. In Vedic Sanskrit (sections 3.1-3.3) and in Finnish (2.2), high agent
nominalizations do not assign structural case if they lack Tense/Aspect features, and even
low agent nominalizations do assign structural case if they have Tense/Aspect features.
3.1 Agent nominals and subject nominals
In their illuminating study based on the FNT approach, Baker & Vinokurova (2009) propose an analysis and typology of agent nominalizations similar to the one I have called
into question for event nominalizations. They begin by noting an asymmetry between
agent and event nominals. “High” event nominals like (1a) have no agent noun counterpart such as (30).
B&V claim that this is a systematic gap, and propose to explain it on the basis of two
key assumptions. First, agentive nominalizing morphology is added by a nominal head
immediately above VP.18 Secondly, in some languages, such as English, structural case
is assigned to objects by an active Voice/v head, whereas in other languages, structural
case is assigned configurationally (dependent case).19 Together, these assumptions rule
out transitive agent-denoting nominalizations, such as (30) *the reader the book. Instead,
they require the structure (31). Here the agent nominalizer pre-empts the case-assigning
active Voice morpheme in v that assigns structural case to objects in English, but (by
hypothesis) has no case-assigning force itself.
It is fair to ask why it is added there and not in a higher position. B&V hint that this is “a position apparently
forced on it by the natural (iconic) semantic composition of the clause” (Baker & Vinokurova 2009: 521),
but this remains to be justified.
B&V equate Voice with v, following Kratzer (2004), but contra Alexiadou (2001); Alexiadou, Soare & Iordăchioaia (2010); Harley (2012), among others.
Paul Kiparsky
(30) * the reader of the book
the book
of the book
15 Nominal verbs and transitive nouns: Vindicating lexicalism
The analysis further predicts that, since voice markers cannot attach to unaccusatives,20
such agent nouns cannot attach to unaccusative verbs.
B&V then draw a distinction between agentive and non-agentive “agent” nominalizers
– let’s call the latter subject nominalizers. Subject nominalizers do assign structural
case, and can be attached to unaccusative verbs. B&V (p. 547) analyze them as “nominal
equivalents of an aspect head”, in the sense in which agentive nominalizers like -er are
nominal equivalents of a voice head. Their example is Gikũyũ -i, another example is
Northern Paiute -dɨ (Toosarvandani 2014). B&V propose the structure (32):
(32) Subject nominalizers (high)
Ni (/Asp)
acc case
As an immediate challenge to the FNT in the domain of agent nominalizations, B&V
note that otherwise low agent nominalizations unexpectedly assign structural case in
some languages. For B&V, these languages must be special in that they assign structural
case by a dependent case mechanism, whereas languages in which low agent nouns have
oblique complements assign structural case by little v. The need to maintain two entirely
distinct mechanisms of structural case assignment on the basis of evidence that cannot
loom large in the learner’s experience would be another disappointing consequence of
the FNT.21 We’ll also see that B&V’s analysis of agent nominals imposes a functional
In fact an incorrect premise: unaccusative verbs passivize in numerous languages, including Finnish and
Sanskrit (Kiparsky 2013).
21 Levin, Preminger & Omer (2015) propose that all structural case can be assigned by dependent case, provided that the algorithm is parametrized in certain ways. However, they do not touch on the case variation
in objects of agent nominalizations, and the parametrization of structural case assignment that they propose does not account for it, as far as I can tell.
Paul Kiparsky
overload on little v that makes the FNT’s various criteria for syntactic height mutually
In summary, B&V’s proposal generates the typology of agent nouns shown in Table 2.
In the remainder of this section I show that the predicted correlations do not hold for
agent nouns of Vedic Sanskrit and Finnish, and propose a much simpler alternative that
does justice to the data.
Table 2: Typology of agent nouns predicted by Baker & Vinokurova (2009).
Agent Nominalizers (low, v)
always agentive
no unaccusatives
structural case only if dependent case
no adverbs
no Aspect
no Voice
Subject Nominalizers (high, Asp)
non-agentives OK
unaccusatives OK
structural case
adverbs OK
compatible with Aspect
compatible with Voice
3.2 Vedic agent nouns
Vedic and Pāṇinian Sanskrit has a large number of agent noun suffixes, which fall into
two clearly demarcated types. A minimal pair that highlights the contrast are the two
agent noun types in accented -tár-N and preaccenting ′-tar-V .22 Agent nouns in accented
-tár-N have genitive objects and get only adjective modifiers, never adverbs, e.g. (33a).
Agent nouns in preaccenting ′-tar-V (boldfaced in (33)) regularly assign structural case
to their objects and, can get certain aspectual adverb modifiers, such as punaḥ ‘again’ in
a. tvā́-ṃ hí satyá-m … vid-má dātā́r-am iṣ-ā́m
you-acc prt true-acc … know-1pl giver-acc good.thing-pl.gen
‘we know you as the true giver of good things.’
(RV. 8.46.2)
b. íṣ-kar-tā víhruta-m púnaḥ
fixer-nom wrong-acc again
‘the maker right again (of) what has gone wrong.’
(RV 8.1.12)
Both suffixes are true nominalizers: they form nouns, not verbs. They have a complete
nominal case and number inflection paradigm, take denominal derivational suffixes, such
Their Indo-European provenance is guaranteed by Greek and Avestan cognates (Lowe 2014). The following
exposition of their contrasting semantics, morphophology, and syntax draws on the generalizations and
evidence in Kiparsky 2016, to which the reader is referred for details.
23 Other agent nouns with verbal properties are attested in early Vedic include -i-V RV 9.61.20 jághnir vr̥trám
‘killer of V̥ tra’, -(i)ṣṇu-V RV 1.63.3 dhr̥ṣṇúr etā́n ‘bold against them’, -u-V AV 12.1.48 nidhanám titikṣuḥ
‘enduring poverty’, -∅-V RV 1.1.4 yáṃ yajñám … paribhū́r ási ‘the sacrifice that you embrace’.
15 Nominal verbs and transitive nouns: Vindicating lexicalism
as derived feminines, and can be compounded.24 They allow adjectival modification (in
addition to adverbial modification, in the case of ′-tar-V ). These nominal properties are
unsurprising for the noun-like -tár-N formations; that they hold also for the more verblike ′-tar-V is documented in (34).
a. āśúṃ
quick-acc win-er-acc
‘the quick (Acc.) winner (Acc.)’
b. táṣṭeva
pṣṭyāmayī ́
iva pr̥ṣṭya-āmay-ín
carve-er.nom like back-ache-ed.nom
‘like a notalgic (Nom.) carpenter (Nom.)’
(RV 8.99.7)
(RV 1.105.18)
Semantically both -tár-N and ′-tar-V are agent nominalizers, not subject nominalizers: they are never added to non-agentive verbs or unaccusatives of any kind, and the
meaning of the nominalization is canonically agentive.25 So by these criteria both nominalizations are “low” in the sense of B&V, not Gikũyũ-type “high” nominalizations.
The agent nominalizers ′-tar-V and -tár-N form a privative semantic opposition, missed
in the modern philological literature but correctly delineated already by Pāṇini, whose
description turns out to tally perfectly with the Vedic data. The unmarked member of
the opposition is -tár-N , which simply denotes agency (like English -er). The marked
member ′-tar-V has two additional meaning components:
a. ′-tar-V denotes agency in ongoing time.
b. ′-tar-V denotes habitual, professional, or expert agency.
The criteria of the FNT make contradictory predictions. Since both nominalizations
are agentive, both should be structurally low “little v” heads. On the other hand, ′-tar-V
nominalizations, which have the verbal properties of assigning structural case and allowing adverbial modification, should be structurally high, while -tár-N nominalizations,
which have strictly nominal properties, should be structurally low. Neither of these is
the case. In fact, as far as the case and adverb properties are concerned, the structure is
just the opposite of what is predicted: verbal ′-tar-V is low and nominal -tár-N is high.
This is shown by four arguments (details in Kiparsky 2016).
The first argument that verbal ′-tar-V is low and nominal -tár-N is high is their morphological position in the word. ′-tar-V always follows the bare verbal root directly, without
E.g. kṣirá-hotar- ‘milk-offerer’ (ŚBr.), and neṣṭā-potā́rau ‘leader and purifier’ (TS.), co-compounds (Kiparsky
2010b) denoting pairs of priests.
25 Thus, the following roots do not take either -tár-N and ′-tar-V or any other agent suffixes for that matter:
as ‘be’, ā́s ‘sit’, śī ‘lie’, sru ‘flow’, plu ‘float’, tras ‘tremble’, vyath ‘sway’, bhraṃś ‘fall’, svap ‘sleep’, kṣudh ‘be
hungry’, tr̥ṣ ‘be thirsty’, svid ‘sweat’, r̥dh ‘flourish’, ru(d)h ‘grow’, pyā ‘swell’, riṣ ‘sustain damage’, mr̥ ‘die’,
śam ‘become calm’, mad ‘get drunk’, mud ‘rejoice’, hr̥ṣ ‘get excited’, dhr̥ṣ ‘dare’, bhī ‘fear’, hīḍ ‘be angry’,
krudh ‘become angry’, gr̥dh ‘be greedy’, ruc ‘shine’, śubh ‘shine, be beautiful’, bhā ‘shine’ bhās ‘gleam’, dyut
‘to strike’ (of lightning), pat ‘to fall’.
Paul Kiparsky
any other intervening suffix; it cannot be added to compound or prefixed bases. -tár-N ,
on the contrary, can be separated from the root by verb-to-verb suffixes commonly analyzed as Voice/v heads (causative, denominative, intensive, desiderative). It is affixed to
the whole verb base, including the extended root plus any prefix that combines with it:
a. RV cod-ay-i-tr-ī-́ ‘impeller (fem.)’ (caus. cod-áy-a-ti ‘impels’)
b. TS pra-dāp-ay-i-tár- ‘bestower’ (caus. prá-dāp-ay-a-ti ‘bestows’),
c. ni-dhā-tár- ‘one who sets down’ (ní-dadhā-ti ‘sets down’)
The morphological data point to the respective constituent structures in (37):
a. Low: [ Prefix [ Root ′-tar-V ] ]
b. High: [ [ Prefix [ Root (Caus)… ] ] -tár-N ]
The second argument that verbal ′-tar-V is low and nominal -tár-N is high comes from
word accentuation. The morphological conditioning of accent placement provides a convenient probe into the constituent structure of words. In Vedic and Pāṇinian Sanskrit,
the accentuation of words is computed cyclically from the accentual properties of the
morphemes from which they are composed. Morphemes may be accented or unaccented,
and at the word level, all accents but the first in a word are erased (Kiparsky 2010a). Both
of our agent suffixes (like the majority of derivational suffixes) belong to the accentually
dominant type: they erase the accent off the bases to which they are added. The crucial
fact for present purposes is that dominant affixes exercise this erasing effect exactly on
the stems to which they are added, no more and no less. Thanks to this property we can
use accentuation to diagnose constituent structure in morphologically complex words.
The empirical generalization is that prefixes always prevail over low (bare-root) suffixes, including ′-tar-V , whereas high suffixes always prevail over prefixes, dictating the
place of the word accent. The reason is that prefixes are added after the low suffix ′-tar-V :
(38) Prefixation to nouns with the the low suffix ′-tar-V :
add dominant preaccenting ′-tar-V
prá-[bhár-tar] add accented prefix
prábhartarerase all accents but the first
On the other hand, -tár-N is accentually dominant, causing all accents on its base to be
deleted, and attracting accent to itself. This shows that it is added to the entire stem
including the prefix, causing the resulting word to be accented on the suffix.
(39) Suffixation of high -tár-N to prefixed verbs:
add accented prefix
[apa-bhar]-tárN add dominant accented -tár-N
apabhartárThe third argument that verbal ′-tar-V is low and nominal -tár-N is high comes from
tmesis, the splitting of prefixes from stems. Prefixes can be separated from verbs and
from nominals formed with low suffixes like verbal ′-tar-V .
15 Nominal verbs and transitive nouns: Vindicating lexicalism
a. sáttā ní yónā (= nísattā yónā) ‘a sitter down in the womb’ (RV 9.86.6)
b. úpa sū́re ná dhā́tā (= sū́re nópadhātā) ‘like the Placer of the Sun’ (RV 9.97.38).
Prefixes are never separated from nominals formed with high suffixes such as nominal
-tár-N .
The explanation comes from the same constituent structure that accounts for the accentual difference: low suffixes such as the agent suffix ′-tar-V are added directly to the
root to form a noun, which can then be composed with a prefix (see (41a)), while high
suffixes such as the agent suffix -tár-N are added to the entire verb, which may already
bear a prefix and/or another suffix (41b,c).
It will be seen the prefix is an immediate constituent of the word in (41a), but not in
(41b) or in (41c). The natural generalization is that a prefix can only be split if it is an
immediate constituent of the word.
The fourth argument that verbal ′-tar-V is low and nominal -tár-N is high comes from
selectional properties of prefixes. Prefixes that only combine with verb roots require
high -tár-N , because the right-branching constituent structure (41a) would require them
to combine with nouns.26 Conversely, prefixes and other elements that cannot be combined with roots, only with nouns, require the right-branching constituent structure
(41a), which is available either with -tár-N or with ′-tar-V .27
Many examples are given in Kiparsky 2016. One will have to suffice here. The interjection hiṁ ‘the sound
hmm’ cannot be compounded with nouns. It can only combine with the root kṛ ‘do’, ‘make’. The agent
noun from hiṁ-kṛ- ‘to make the sound hmm’ must therefore have the high suffix -tár-N , viz. hiṁkartár-.
27 Again we must make do with a couple of examples. There is no compound verb such as *para-apara-i- ‘to
go far and near’ from which párāpara-etar- ‘one who goes far and near’ might be derived. In fact párāpara‘far and near’ is never compounded with verbs. Instead, the agent noun is a nominal compound formed
from para-apara- ‘far and near’ plus e-tár- ‘goer’ (← i-tár-N ). Another illustration of this generalization
is that the negation a- combines only with nouns. From hótar ‘priest’ (← hu-′tar-V ) we get á-hotar- ‘a
Paul Kiparsky
The above arguments establish the morphological constituency displayed in (37) and
(41). But Distributed Morphology is a resourceful theory that makes available various
movement operations that cause mismatches between morphology and syntax. So could
the morphologically low nominalizing morphemes be spelled out high where B&V predict they should be, and then undergo Lowering to their actual position? And conversely,
could the high nominalizing morphemes be spelled out low as predicted, and then undergo Raising to their actual position? The answer is negative on both counts.
The way morphologically low suffixes such as the agent suffix ′-tar-V could be syntactically high for purposes of the FNT is by DM’s Lowering operation, which applies
before Vocabulary insertion to adjoin a head to the head of its complement (Embick &
Noyer 2001):
(42) X0 lowers to Y0 :
[XP X0 …[YP …Y0 …]] → [XP …[YP …[Y0 Y0 +X0 ]…]]
In English, Lowering of T is assumed to adjoin T to v (Embick & Marantz 2008). But if
′-tar-V is generated in the N head of (32) and is lowered from there into the v, it would, on
B&V’s assumptions, have the properties of a subject nominalizer, forming nouns from
unaccusative and non-agentive verbs. But it does not have any properties of a subject
nominalizer – it is a true agent nominalizer, as we showed above.
The other thing required to maintain the FNT is to generate -tár-N low in v (as in (31))
and then raise it to its actual high position. This is not going to work in the DM model
either, for morphological head-raising can only adjoin a head to the next head above it
(Harizanov & Gribanova 2016), and in this case it would have to skip intervening heads,
including the v head that may be occupied by causative and other V→V suffixes, which
must not raise. Moreover, in more recent DM (Embick 2010), phonology is cyclically
interleaved with morphology, and this would cause problems with the abovementioned
accent erasure and tmesis phenomena.
It should also be noted that ′-tar-V is overwhelmingly preferred when its special meaning and morphological restrictions allow its use, and -tár-N is used elsewhere. Moreover,
other agent suffixes supersede each of them with particular roots and/or in particula
special meanings. Since competition in DM obtains only between morphemes that have
the same meaning and are realized in the same slot (such as English plural -s and -en),
all these competing suffixes would have to be generated in the same syntactic position.
The conclusion from the Vedic data is that the nominalizations’ verbal vs. nominal
properties do not correlate with structural height of their heads in the word or in the
syntax. In fact, the majority of morphologically high nominalizers in Vedic have nominal
properties, and the majority of morphologically low nominalizers have verbal properties
– the opposite of what the FNT predicts.
3.3 Aspect in Vedic agent nouns
A preliminary survey of nominalizations suggests that the Aspect feature of a nominalizer is the best predictor of verbal properties. Consider the following alternative to the
15 Nominal verbs and transitive nouns: Vindicating lexicalism
The Aspect hypothesis
Nominalizations assign structural case if and only if they have Aspect, either as
an inherent feature of the nominalization, or in virtue of combining with overt
Aspect morphology.
By aspect features I mean outer aspect features such as imperfective and habitual, not
inner aspect (Aktionsart), such as telicity. Sanskrit agent nouns in ′-tar-V are inherently
present/imperfective and habitual. Those in -tár-N , English agent nouns made with -er,
and Finnish agent nouns made with -ja have no inherent aspect: a driver can be a habitual
or professional driver, or just someone who happens to be at the wheel on a particular
The Aspect hypothesis is not implausible a priori because Aspect features are crosslinguistically known to affect case assignment – think of split ergativity based on imperfect vs. perfect aspect, and accusative vs. partitive objects in Finnish depending on gradability (Kiparsky 2005a). It looks promising for Vedic Sanskrit in particular because nominalizing endings with verbal properties, such as ′-tar-V , are added to the bare root, in the
same morphological position as the Aorist and Perfect Tense/Aspect suffixes. It is also
consistent with the fact that Northern Paiute deverbal nominalizations, which assign
structural accusative case, can have overt aspectual morphology below them (Toosarvandani 2014: 793, fn. 6).
The Aspect hypothesis is compatible with a lexicalist treatment of morphology. A
Distributed Morphology analysis of Vedic agent nouns is problematic because of the
conflicting criteria for structural height. In addition, they show a type of competition
between morphemes that DM rejects. The semantically nondescript -tár-N , structurally
low by B&V’s criteria but high in word structure, is the default (elsewhere) case. The semantically restricted ′-tar-V suffix, structurally high by B&V’s syntactic criteria but low
in the word morphology, is strongly preferred whenever it is applicable, namely to denote habitual agency in ongoing time with morphologically simple verbs. Elsewhere the
default is -tár-N , structurally low by B&V’s syntactic criteria but high in the word morphology – for past or future agency, or occasional agency, or when the verb is morphologically complex (causative, intensive, desiderative, denominative, or prefixed. Suppose
then that a structurally low agent is added in a syntactic derivation. The derivation must
crash if and only if a structurally high agent can be successfully added in a competing
derivation. But DM does not allow rules that spell out syntactico/semantic features in
different positions to compete with each other. Moreover, if we assume bottom-up morphological spellout of the syntax (by cycles or phases), the syntactically low agent would
have to “know” about the upstairs high agent in order to be blocked by it. On the other
hand, in a morphological theory of word-formation, morphologically low items naturally
block morphologically high items. Besides, blocking of affixes with general meanings by
affixes with special meanings regardless of the locus of affixation is straighforward in
lexicalist approaches such as those of Wunderlich (1996b; 2001) and Kiparsky (2005b).
Gerunds are arguably inherently imperfective (Alexiadou 2001; Alexiadou, Soare & Iordăchioaia 2010).
Paul Kiparsky
3.4 The Finnish subject nominalizer -ja
B&V’s formulation of the FNT entails that agent nominalizations don’t assign case and
subject nominalizations do (§2). We have seen that Vedic falsifies the first of these claims.
Finnish (among many other languages) falsifies the second. The fully productive Finnish
suffix -ja is not an agent nominalizer, but a subject nominalizer, which is to say a high
nominalizer in B&V’s typology. It can go on non-agentive/unaccusative verbs, and freely
attaches to causatives, often assumed to be under v, as well as denominatives, reflexives,
inchoatives, and inner aspect morphemes such as frequentatives and semelfactives, thus
testing positively for high position by several diagnostics.
a. kuolija ‘one who dies’, ‘dier’, eläjä ‘one who lives’, ‘liver’, toipuja ‘one who
gets well’, ‘convalescent’, olija ‘one who is’, osaaja ‘one who is able to’,
syntyjä ‘one who is born’, hikoilija ‘one who sweats’, putoaja ‘one who falls’,
turpoaja ‘one who swells’, pelkääjä ‘one who fears’, luulija ‘one who
supposes’, tuntija ‘one who knows, expert’, muistaja ‘one who remembers’,
jääjä ‘one who remains’, palelija ‘one who feels cold’, tarvitsija ‘one who
needs’, hukkuja ‘one who drowns’
b. i. Frequentative -ele-: kys-eli-jä ‘inquirer’, from kys-ele- ‘to make inquiries’
(cf. kysy-jä ‘asker’, from kysy- ‘ask’). Similarly ryypiskelijä ‘tippler’,
lähentelijä ‘harasser’, myyskentelijä ‘peddler’, rehentelijä ‘bragger’,
riitelijä ‘quarreler’.
ii. Causative: laula-tta-ja ‘one who makes sing’, from laula-tta- ‘to make
sing’, cf. lauja-ja ‘singer’, from laula- ‘to sing’.
iii. Inchoative + causative: selv-en-tä-jä ‘clarifier’, ← selv-en-tä- ‘make clear,
clarify’ (← selv-en- ‘become clear’ ← selvä ‘clear’).
iv. Causative + frequentative: sopi-ja ‘agreer’, sovi-tta-ja ‘fitter, arranger’,
sovi-tt-el-ija ‘reconciler, negotiator’
v. Reflexive: puolusta-ja ‘defender’ puolusta-utu-ja ‘(self-)defender’
vi. Denominative: testamentt-aa-ja ‘bequeather’ (← testamentt-at- ‘to
bequeath’ ← testamentti ‘testament’)
However, nominals in the suffix -ja do not assign structural case, whether they have any
of these suffixes below them or not. Their object complement (unlike that of passive
verbs) can only receive genitive case.
a. palkinto-j-en (*palkinno-t) saa-ja
prize-pl-gen (prize-pl.acc) get-er.(nom)
‘the/a winner of the prizes’
b. minu-n (*minu-t) käv-el-ytt-eli-jä-ni
me-gen (*me-acc) walk-freq-caus-freq-er.(nom)-1sg.poss
‘(the) one who frequently takes me around for walks’
Nominals in -ja do take oblique nominal modifiers, as do all nominalizations, including
action nominalizations, and even to some extent ordinary basic nouns.
15 Nominal verbs and transitive nouns: Vindicating lexicalism
a. Saksa-sta
voitta-ja-na palaa-ja
/ pal-uu
Germany-elat win-er-essive return-er.(nom) / return-ing.(nom)
‘(the) one who returns / a return from Germany as a winner’
b. hallitse-va-ssa asema-ssa
/ ol-o
govern-in-iness position-iness be-er.(nom) / be-ing.(nom)
‘(the) one who is in a governing position’
c. palatsi Cannesi-ssa
palace Cannes-iness
‘the/a palace in Cannes’
They generally do not take adverbs, except for certain perfectivizing adverbs (47d,e):
a. * nopea-sti juoksi-ja
quick-ly run-er.(nom)
‘(the) one who runs/ran/will run quickly’
b. * kilpailu-n
taas voitta-ja
competition-gen again win-er.(nom)
‘(the) one who wins/won/will win the/a competition again’
c. * aina matkusta-ja
always travel-er.(nom)
‘(the) one who always travels’
d. kilpailu-sta
pois jää-jä
competition-elat away remain-er.(nom)
‘one who does not join the competition’, ‘eliminee’
e. viime-ksi tuli-ja
lat-transl come-er.(nom)
‘the last to arrive’
In (48) (an example adapted from the internet) the adverb jälleen ‘again’ appears with an
agent noun.
Cannesi-ssa jälleen palkinno-tta jää-jä
Cannes-iness again prize-abess remain-er.(nom)
‘one who ended up prizeless again in Cannes’ (lit. ‘a remainer prizeless again…’)
Possibly it modifies not the nominalization but the abessive modifier palkinno-tta ‘prizeless’.
Since Finnish -ja must be high in order to get a non-agentive interpretation and to
scope over every kind of verb-to-verb suffix, it should assign structural case, which it
doesn’t. So it does not fit into B&V’s syntactic typology, and constitutes a problem for
the FNT generalization. In this case, morphological raising or lowering, even if they
were available and motivated, would not help to resolve the contradiction.
Paul Kiparsky
The lexicalist alternative, however, holds up. Like English -er, nouns in -ja are morphologically incompatible with overt Aspect or Voice morphology, and they refer indifferently to prospective, present, or past events, hence have no inherent Aspect features.
maksaja ‘payer’ (one who has paid, is paying, or will pay), similarly ostaja ‘buyer’,
vuokraaja ‘renter’, maahanmuuttaja ‘immigrant’, lähtijä ‘goer’, siittäjä ‘inseminator’
Since -ja has no Aspect features and no verbal properties, the fact that it doesn’t assign
structural case is consistent with the Aspect hypothesis but inconsistent with the FNT.
I conclude that Finnish -ja supports the lexicalist analysis of nominalizations.
3.5 The Sakha agent nominalizer -AAccY
Baker & Vinokurova (2009: 536) note that the correlation predicted by the FNT breaks
down for Sakha agent nominalizations, which have structural accusative objects but
otherwise conform to the low type, in that they have no Aspect morphology or adverbs.
Their solution is that Sakha accusative case is not assigned by Voice/v but by a different
mechanism, dependent case assigment.
(50) Dependent case assignment (Marantz 1991; Baker 2015)
a. If there are two distinct NPs in the same spellout domain such that NP1
c-commands NP2 , then value the case feature of NP2 as accusative unless NP1
has already been marked for case.
b. If there are two distinct NPs in the same spellout domain such that NP1
c-commands NP2 , then value the case feature of NP1 as ergative unless NP2
has already been marked for case.
Their main argument that Sakha accusative is dependent case is that objects of passives
receive accusative case. This argument depends on the fragile assumption that the implicit agent of Sakha passives is a syntactically visible but phonologically null NP, which
receives nominative case and serves as the NP1 that triggers the assignment of accusative
case to the object of the passive by (50a).
I am skeptical of this solution for both theoretical and empirical reasons. It would
be strange for UG to offer two entirely different methods of structural case assignment,
since their empirical differences are rather obscure, and offer learners of most languages
little core data to choose between them. Secondly, the analysis of impersonal passives as
having an invisible nominative agent subject is excluded on general grounds by any kind
of demotion analysis of passive, including the typologically grounded theory proposed
in Kiparsky 2013. Finnish provides empirical evidence against the idea that objects of
passives receive dependent structural case because of a syntactically visible but phonologically null nominative implicit agent. The object of passives in Finnish is assigned
structural case as in Sakha, but the case cannot possibly be assigned by the dependent
case algorithm (50), for the implicit agent of passives in Finnish is invisible to case assignment, as clearly demonstrated by Jahnsson’s Rule (20b), see e.g. (22c,d).
15 Nominal verbs and transitive nouns: Vindicating lexicalism
Our approach predicts the transitivity of Sakha agent nouns in -aaccy out of the box.
The reason is that they have an aspect feature. Agent nouns in Sakha -aaccy denote
specifically habitual or generic agents. B&V (p. 531) illustrate this generalization with
the following examples:29
ynaq-y ölör-ööccü
cow-acc kill-ag.nom
‘a killer of cows, a butcher’
b. * Misha-ny ölör-ööccü
Misha-acc kill-ag.nom
‘the killer of Misha’
The habitual aspect feature of the Sakha agent nouns licenses its structural case assignment just as the habitual aspect feature of agent nouns formed by the Sanskrit agent
suffix ′-tar-V does, as opposed to aspectually void agent nouns in -tár-N , Finnish -ja, and
English -er. This accounts fully for the case data without resorting to the unsupported
syntactic height distinctions demanded by the FNT.
Summing up our conclusions about agent nominalizations so far: the syntactic FNT
is falsified by Vedic agent nominalizations in one direction, and by Finnish subject nominalizations in the other, and it requires an otherwise unsupported parametric choice
between two heterogeneous structural case assignment algorithms. The analysis reveals
that little v can’t do all of the following things: (1) introduce Agents, (2) host voice heads
or agent nominalizer heads, (3) host causative V affixes, (4) host aspectual material, and
(5) assign structural case. In agent nominals it is not possible to place nominalizing heads
above or below little v in a consistent way that satisfies all of (1)-(5). (1) and (5) cannot be
reconciled with an agent nominalizer that assigns structural case such as Sanskrit ′-tar-V ,
or with a subject nominalizer that does not assign structural case such as Finnish -ja. The
Sakha nominalizer -AAccY can dominate causatives (high) but not aspectual adverbs –
a conflict between (3) and (4) – and introduces agents (low) but assigns structural case
(high) – a conflict between (2) and (5).
These difficulties fall away if we assume that that agent nominals are nouns, and that
nouns assign structural case if and only if they have Aspect features.
4 Conclusion
The Functional Nominalization Thesis claims that so-called “mixed categories” arise when
a nominal head is affixed to an extended verbal projection that is its syntactic complement. My findings instead support a lexicalist approach, in which mixed categories are
projections of a nominal or verbal heads with an extra phi-feature. Their extended projections behave like normal extended projections modulo the properties enforced by that
This component is foregrounded in the related habitual participle function of the same suffix: e.g.
salaj-aaccy means both ‘manager’ (agent noun) and ‘habitually managing’ (participle), see Vinokurova
(2005: 123).
Paul Kiparsky
In §2 I argued that gerund phrases are not DPs/NPs with AspP complements. They
are not even nominalizations. They are participial phrases – IPs with a Case feature
that is checked or valued in an argument position. In all other respects their syntax is
entirely clausal: they lack DP material such as articles, demonstratives, quantifiers, and
adjectives, and they are formally built like IPs, complete with structural subjects. The
lexicalist analysis explains these properties.
In §3 I argued that agent nominalizations that assign structural case to their objects are
not nouns with vP complements (or with any other phrasal complements), but deverbal
nouns derived by agent suffixes that have an Aspect feature. The Aspect feature makes
the nouns transitive, and modifiable by aspectual adverbs. Otherwise their syntax is
entirely nominal. The merit of this analysis is that it tightly correlates the transitivity of
agent nouns with their aspectual meaning. Also, by relieving the burden on little v, it
eliminates the mismatches between word structure and syntax that we found in Vedic
and Finnish agent nouns under the FNT analysis.
The lexicalist approach retains the key idea of the FNT without the typologically unwarranted overgeneration caused by allowing syntactic affixation. It preserves a uniform
mechanism of structural case assigment, a unified analysis of true nominalizations, and
the insights that originally led Chomsky to lexicalism.
I am grateful to Vera Gribanova and to an anonymous reviewer for their very helpful
comments and queries.
Abney, Steven. 1987. The English noun phrase in its sentential aspect. Massachusetts Institute of Technology PhD dissertation.
Alexiadou, Artemis. 2001. Functional structure in nominals: nominalization, and ergativity.
Amsterdam: John Benjamins.
Alexiadou, Artemis, Elena Soare & Gianina Iordăchioaia. 2010. Number/aspect interactions in the syntax of nominalizations: a distributed approach. JL 46. 537–574.
Anderson, Stephen R. 1992. A-morphous morphology. Cambridge: Cambridge University
Anderson, Stephen R. 2016. The role of morphology in transformational grammar. In
Andrew Hippisley & Gregory T. Stump (eds.), The Cambridge handbook of morphology,
587–608. Cambridge: Cambridge University Press.
Anttila, Arto & Jong-Bok Kim. 2016. Locality and variation in Finnish structural case.
Natural Language and Linguistic Theory. DOI:doi:10.1007/s11049-016-9352-x
Baker, Mark & Nadya Vinokurova. 2009. On agent nominalizations and why they are
not like event nominalizations. Language 85. 517–556.
15 Nominal verbs and transitive nouns: Vindicating lexicalism
Baker, Mark C. 2015. Case: Its principles and parameters. Cambridge: Cambridge University Press.
Blevins, James P. 2003. Remarks on gerunds. In Peter Sells & C. Orhan Orgun (eds.),
English morphology and the web of grammar. Stanford: CSLI Publications.
Chomsky, Noam. 1955. The logical structure of linguistic theory. ms. published 1985.
Chicago: University of Chicago Press.
Chomsky, Noam. 1970. Remarks on nominalization. In R. Jacobs & P. Rosenbaum (eds.),
Readings in English transformational grammar, 184–221. Waltham, MA: Ginn.
Deal, Amy Rose. 2009. The origin and content of expletives: evidence from “selection”.
Syntax 12. 285–323.
Embick, David. 2010. Localism versus globalism in morphology and phonology. Cambridge:
MIT Press.
Embick, David & Alec Marantz. 2008. Architecture and blocking. Linguistic Inquiry 39(1).
Embick, David & Rolf Noyer. 2001. Movement operations after syntax. Linguistic Inquiry
32(4). 555–595.
Hakulinen, Auli & Fred Karlsson. 1979. Nykysuomen lauseoppia. Vol. 350. Suomalaisen
kirjallisuuden seura.
Hakulinen, Auli & Lauri Karttunen. 1973. Missing persons: on generic sentences in
Finnish. In Cls9: papers from the ninth regional meeting, 157–71. Chicago Linguistic
Society, Chicago.
Harizanov, Boris & Vera Gribanova. 2016. Whither head movement? Workshop on the
Status of Head Movement in Linguistic Theory. Stanford.
Harley, Heidi. 2012. External arguments and the mirror principle: on the distinctness of
voice and v. Lingua 125. 34–57.
Huddleston, Rodney & Geoffrey Pullum. 2002. The cambridge grammar of the English
language. Cambridge: Cambridge University Press.
Kim, Jong-Bok. 2016. The syntactic structures of Korean: a construction grammar perspective. Cambridge: Cambridge University Press.
Kiparsky, Paul. 2001. Structural case in Finnish. Lingua 111. 315–376.
Kiparsky, Paul. 2005a. Absolutely a matter of degree: The Semantics of Structural Case in
Kiparsky, Paul. 2005b. Blocking and periphrasis in inflectional paradigms. In Geert Booij
& Jaap van Marle (eds.), Yearbook of morphology 2004, 113–135. Dordrecht: Springer.
Kiparsky, Paul. 2010a. Compositional vs. paradigmatic approaches to accent and ablaut.
In Stephanie W. Jamison, H. Craig Melchert & Brent Vine (eds.), Proceedings of the 21st
annual UCLA Indo-European conference. Bremen: Hempen.
Kiparsky, Paul. 2010b. Dvandvas, blocking, and the associative: The bumpy ride from
phrase to word. Language 86. 302–331.
Kiparsky, Paul. 2013. Towards a null theory of the passive. Lingua 125. 7–33.
Kiparsky, Paul. 2016. The agent suffixes as a window into Vedic grammar. In Joshua
Katz Dieter Gunkel Brent Vine & Michael Weiss (edd (eds.), Sahasram ati srajas: Indo-Iranian and Indo-European studies in honor of Stephanie W. Jamison.
Paul Kiparsky kiparsky/Papers/ Ann Arbor: Beech
Stave Press.
Kiparsky, Paul. In press. Paradigm effects and opacity. Stanford: CSLI.
Kornfilt, Jaklin & John Whitman. 2011. Afterword: nominalizations in syntactic theory.
Lingua 121. 1297–1313.
Kratzer, Angelika. 1996. Severing the external argument from its verb. In J. Rooryck &
L. Zaring (eds.), Phrase structure and the lexicon. Dordrecht: Kluwer Academic Press.
Kratzer, Angelika. 2002. The event argument and the semantics of verbs. Manuscript, University of Massachusetts, Amherst.
Kratzer, Angelika. 2004. Telicity and the meaning of objective case. In Jacqueline
Lecarme & Jacqueline Guéron (eds.), The syntax of time. Cambridge, MA: MIT Press.
Lausetyyppien käsittely transformaatioteoriassa [The treatment of sentence Types in
Transformational grammar]. 1970. Ajatus (Journal of the Finnish philosophical society]
XXXII. 150–159.
Lees, Robert B. 1960. The grammar of English nominalizations. International Journal of
American Linguistics 26(3).
Levin, Theodore, Preminger & Omer. 2015. Case in Sakha: are two modalities really necessary? Natural Language and Linguistic Theory 33. 231–250.
Lowe, John J. 2014. Transitive nominals in Old Avestan. Journal of the American Oriental
Society 134(4). 553–577.
Malouf, Robert. 2000. Mixed categories in the hierarchical lexicon. Stanford: CSLI Publications.
Marantz, Alec. 1991. Case and licensing. In German Westphal, Benjamin Ao & Hee-Rahk
Chae (eds.), Proceedings of the 8th Eastern States Conference on Linguistics (ESCOL 8),
234–253. Ithaca: CLC Publications.
Payne, John, Geoffrey K. Pullum & Rodney Huddleston. 2010. The distribution and category status of adjectives and adverbs. Word Structure. 31–81.
Pires, Acrisio. 2006. The derivation of clausal gerunds. Syntax 10. 165–203.
Portner, Paul. Quantification, events, and gerunds. In Emmon Bach, Eloise Jelinek, Angelika Kratzer & Barbara Partee (eds.), Quantification in natural language, 619–659.
Dordrecht & Boston: Kluwer Academic Publishers.
Pullum, Geoffrey. 1991. English nominal gerund phrases as noun phrases with verbphrase heads. Linguistics 29. 763–799.
Pustejovsky, James. 1995. The generative lexicon. Cambridge, MA: MIT Press.
Ross, John. 1967. Constraints on variables in syntax. Massachusetts Institute of Technology Ph.D. dissertation.
Svenonius, Peter. 2016. Spans and words. In Daniel Siddiqi & Heidi Harley (eds.), Morphological metatheory. Amsterdam & Philadelphia: John Benjamins.
Toosarvandani, Maziar. 2014. Two types of deverbal nominalization in Northern Paiute.
Language 90. 786–833.
Vergnaud, Jean-Roger. 1977. Open letter to chomsky and lasnik. LingBuzz 461. http://ling.
15 Nominal verbs and transitive nouns: Vindicating lexicalism
Vinokurova, Nadezhda. 2005. Lexical categories and argument structure. Utrecht: LOT
Williams, Edwin S. 1975. Small clauses in English. In John Kimball (ed.), Syntax and semantics, vol. 4, 249–273. New York: Academic Press.
Wunderlich, Dieter. 1996a. A minimalist model of inflectional morphology. In Chris
Wilder, Hans-Martin Gärtner & Manfred Bierwisch (eds.), The role of economy principles in linguistic theory, 267–298. Berlin: Akademie Verlag.
Wunderlich, Dieter. 1996b. Minimalist morphology: the role of paradigms. Yearbook of
Morphology. 93–114.
Wunderlich, Dieter. 2001. How gaps and substitutions can become optimal: the pronominal affix paradigms of Yimas. Transactions of the Philological Society 99. 315–366.
Yoon, James Hye Suk. 1996. Nominal gerund phrases in English as phrasal zero derivations. Linguistics 34. 329–356.
Chapter 16
On mechanisms by which languages
become [nominative-]accusative
Ashwini Deo
The Ohio State University
New Indo-Aryan languages are characterized by accusative (DOM) objects in ergative, perfective clauses. This paper traces the emergence of this ergative—accusative marking pattern
with the goal of determining whether it is to be considered part of a single “de-ergativization”
trajectory, in which languages gradually lose aspects of their ergative orientation in analogy
to the non-ergative portion of the grammar. Data from Middle Indo-Aryan suggests that accusative marked objects — a deviation from the classic ergatively-oriented sub-system — cannot be analyzed in terms of the analogical extension of any existing nominative-accusative
model or as a reduction of markedness. In contrast, the empirical facts of Indo-Aryan diachrony align better with the possibility that such deviations have to do with independent
changes in the broader argument realization options for the language. This is consistent with
Anderson’s (1977; 2004) claim that a significant part of the explanation for ergativity-related
patterns lies in patterns of diachronic change rather than abstract structural considerations
of Universal Grammar.
1 Introduction
The term ergative is used to refer to a grammatical relation marking pattern in which
the object of a transitive verb patterns with the single argument of an intransitive verb
(surfacing with absolutive case), while the transitive subject patterns distinctly (surfacing with ergative case) (Dixon 1979; 1994; Comrie 1978; Plank 1979). It has sometimes
been claimed that there is a clear asymmetry between the pervasiveness of ergative–
absolutive vs. nominative–accusative marking systems across sub-domains of grammars
in languages.
No ergative language is fully consistent in carrying through the ergative principle
throughout its entire morphology, syntax, and lexicon: all languages that exhibit
ergative patterning in their commonest case-marking system also exhibit some
accusative pattern somewhere in the rest of their grammar. (Moravcsik 1978, p.237)
Ashwini Deo. 2017. On mechanisms by which languages become [nominative]accusative. In Claire Bowern, Laurence Horn & Raffaella Zanuttini (eds.),
On looking into words (and beyond), 347–369. Berlin: Language Science Press.
Ashwini Deo
A possible way of interpreting this stated generalization is to take it to refer to the presence of accusative case-marking in ergative languages – that is, that every language with
an ergative-nominative case marking or agreement pattern also exhibits a nominative–
accusative pattern in some subsystem of the grammar. However, this interpretation is
clearly not borne out since several languages exist that have ergative case but lack accusative case marking altogether.1 Coon & Preminger (to appear) interpret the above
claim to mean that even in languages which show a high number of ergative characteristics, there can generally be found some portion of the grammar in which the ergative
pattern is lost, and transitive and intransitive subjects are treated alike. In this case, the
term “ergative pattern” seems to refer, not to surface morphological properties, but more
broadly to syntactic properties like control and binding with respect to which the highest
arguments of a clause may pattern alike. Split-ergativity is a term reserved specifically
for morphological marking patterns and refers to the systematized occurrence of a mixed
indexing system, which is ergatively organized in well-defined syntactic-semantic configurations with nominative–accusative marking elsewhere in the language. The question of how such systems arise in natural languages and change (or persist) through time,
as well as the possible diachronic reasons for the parameters on which the split is based,
can only be answered by an investigation of split-ergative languages for which we have
some clear diachronic record available.
Anderson (1977, and later in 2004)has suggested that to the extent we have such information, changes involving ergative orientation seem to be “consequences of relatively
superficial phenomena.” According to him, ergative patterning is not a deep syntactic
property of linguistic systems but rather an emergent effect arising from several distinct
trajectories in the morphological systems of languages. In effect, there is no principle
that determines an “ergative” or “accusative” pattern; rather languages may innovate or
lose specific cases such as ergative or accusative, with such patterns arising more as emergent effects of the change and not as abstractly determined invariant objects. This paper
examines one such emergent effect in trajectories associated with systems containing
ergative case – the emergence of overt accusative (object) marking in ergative clauses.
New data from Late Middle Indo-Aryan (MIA) and Early New Indo-Aryan (NIA) suggests
that transitions resulting in deviations from the classic ergatively-oriented sub-system
in a split ergative language cannot be analyzed uniformly in terms of the analogical extension of any existing nominative-accusative model or as a reduction of markedness.
In contrast, the empirical facts of Indo-Aryan diachrony align better with the possibility that such deviations have to do with independent changes in the broader argument
realization options for the language. This is consistent with Anderson’s claim that a
significant part of the explanation for ergativity-related patterns lies in patterns of diachronic change rather than abstract structural considerations of Universal Grammar
(contra Delancey 1981; Dixon 1994; Tsunoda 1981).
An anonymous reviewer points to languages like Chukchi, Tabassaran, Chamalal, Tzutujil, Central Yupik
Eskimo, and Burushaski that lack an accusative case, and therefore lack nominative-accusative “patterning”
in terms of case marking.
16 On mechanisms by which languages become [nominative-]accusative
2 Morphosyntactic changes in Middle Indo-Aryan
2.1 The emergence of ergativity
One well-discussed source for ergative marking in natural languages is a passive clausal
structure that gets reanalyzed as active. Oblique marking on the optionally surfacing
agent is reanalyzed as ergative case while the unmarked subject of the passive clause
surfaces as absolutive object, identical to the subjects of intransitive clauses. Indo-Aryan
languages bear the most concrete diachronic record for such a passive–to-ergative shift
scenario. In the history of these languages, a passive construction with resultative semantics was reanalyzed as an active, ergative clause with perfective aspectual reference at
least by the time of Epic Sanskrit (Old Indo-Aryan (OIA)) and Early MIA (Andersen 1986;
Peterson 1998; Condoravdi & Deo 2014 a.o.).2 In the oldest Vedic texts, the -ta-affixed
form of the verb serves to describe a result-state brought about by a preceding event
when it is used predicatively in an adjectival passive construction. The -ta forms (boldfaced) in (1a) agree with the nominative patient while the agent remains unexpressed. In
(1b), the agents and instruments are overtly expressed in the instrumental case.
a. stīr-ṇáṃ
át-tave eat-inf
‘The Barhis has been strewn for thee, O Indra; the Soma has been pressed
(into an extract). The barley grains have been prepared for thy two
bay-horses to eat.’ (Ṛgveda 3.35.7)
b. nṛ-bhir
‘It (the Soma) has been washed by men, pressed with the help of stones,
strained with wool-filters.’ (Ṛgveda 8.2.2)
As shown in (2), the -ta form agrees with the sole (nominative) argument of intransitive verbs. This results in a difference in the marking of the subject arguments of transitive and intransitive verbs. In (1) the verb does not agree with the instrumental agentive
arguments. In (2), in contrast, the verb śri-taḥ has a nominative subject soma and agrees
with it in number and gender.
The Indo-Aryan branch of Indo-European inherits the deverbal result stative form with the affix -ta (allomorph -na) (reconstructed for Indo-European as *-to/-no). -ta, attested at all stages of OIA and MIA,
attaches directly to the root, and the resulting stem is adjectival, inflecting for number and gender like any
other adjectival forms.
Ashwini Deo
adhi śri-taḥ on
‘Soma rests (is supported) in the heaven.’ (Ṛgveda 10.85.1)
This resultative -ta construction (sometimes in periphrasis with tense auxiliaries) is
the source of the ergative pattern observed in the perfective aspect in the later languages.
In later stages of OIA, the construction was extended to marking the perfect aspect and
it exhibited existential as well as universal perfect readings (Condoravdi & Deo 2014).
By the time of Epic Sanskrit (late stage of OIA), the -ta construction became a frequently
used device for marking past perfective reference. The agent argument in these cases
is most frequently overt and marked with instrumental case. Past eventive reference is
indicated by the presence of past referring frame adverbials like purā ‘formerly’ and tadā
‘then’. Perfective clauses containing intransitive verbs occur with nominative subjects
(3c). All the examples below are from the Mahābhārata, one of two epics that constitute
the record for this stage of the language.
a. purā
ca eva dṛṣ-ṭaṃ
formerly and ptcl everything
‘Lord, formerly, in the age of the Deva (Gods), I saw everything.’
(Mahābhārata 3.92.6a; Deo 2012)
b. hṛ-tā
tadā t-ena then
tu na tark-itaḥ ptcl neg
‘Then he stole that cow, but did not consider the fall (consequences).’
(Mahābhārata 1.93.27e; Deo 2012)
c. jaratkāruḥ
Jaratkā accompanied
‘Jaratkāru went to heaven accompanied by his ancestors.’ (Mahābhārata
The main change between Epic Sanskrit (OIA) and the later MIA stage of the language
concerns the erosion and simplification of the rich tense-aspect system (Pischel 1900;
Bloch 1965). Inflectional past referring forms such as the aorist, the inflectional perfect,
and the imperfect disappeared from the language, leaving the -ta construction as the
only past referring device.3 This loss of the inflectional system has often been cited as a
reason for the increase in the frequency and scope of the participial construction, which
Traditional grammarians do provide instances of the inflectional perfect and the aorist during this period,
but they only occur as isolated, unanalyzed forms for a few verbs like āha-‘say-aor’ and akāshi -‘do-aor’.
16 On mechanisms by which languages become [nominative-]accusative
in turn led to the unmarking of the stative nature of the construction. The change to
an ergative alignment was certainly complete at the Mid to Late MIA stage (Hock 1986;
Bubenik 1998). The examples below from an archaic MIA Mahāraṣṭrī text Vasudevahiṃḍī
(ca. 500 AD) shows this ergative alignment. The verb agrees with the nominative subject
in (4a). In (4b) the verb agrees with the nominative marked object while the agentive
argument (‘that running one’) appears in the instrumental.
a. pat-to
ya seṇiyo
ta-m and Seṇ
‘And King Seṇiya reached that place.’ (Vasudevahiṃḍī KH. 17.1)
b. t-eṇa
palāyamāṇ-eṇa purāṇakuv-o
taṇadabbhaparichinn-o diṭ-ṭho
‘That running one noticed an old well covered with grass.’ (Vasudevahiṃḍī
KH. 8.6)
Indo-Aryan diachrony after the MIA stage has often been characterized as involving a
progressive loss of ergative alignment and gradual drift towards a nominative-accusative
marking in perfective clauses. There are three observed ways in which the descendent
systems deviate from the proto-ergative system of MIA: (a) Loss of ergative morphology
in pronominal and nominal paradigms4 ; (b) Subject agreement (replacing or in addition
to object agreement); (c) Accusative marking on a privileged class of objects, i.e. the
spread of differential object marking.
It is logical to think of the implementation of any of these changes independently or
together as the “de-ergativization” of an ergative system in analogy to the non-ergative
portion of the grammar. Indeed, the patterns seen in individual NIA languages, such as
suppression of overt ergative case (e.g. in Old Hindi and Marathi); nominative subjects
(e.g. in Bangla) and agreement with overt ergative subject (e.g. in Nepali) are all analogizable to existing marking patterns in the language such as unmarked subjects, nominative subjects, and subject agreement. However, the emergence of accusative marking on
objects of transitive, perfective clauses poses a puzzle for a straightforward analogical
In fact, data from some Early NIA languages, e.g. Hindi, reveals that the original instrumental marking
observed on transitive subjects for the MIA ergative system is entirely lost for all nominal and pronominal
expressions in some stages of Indo-Aryan. The ergative pattern of agreement is nevertheless retained. The
example in (i) is from the work of Kabir, a poet from the 15th century CE. There is no overt ergative marking
on the 3rd person subject but the agreement on the verb is with the feminine object argument (explicit or
unpronounced) chādar ‘sheet’.
(i) jo
which gods-men-sages.∅erg
‘Which sheet the Gods, men, and sages, all wore, (that sheet)…’
Ashwini Deo
extension narrative for de-ergativization. The puzzle arises from the evolution of case
marking in MIA, to which we now turn.
2.2 Syncretism in nominal case marking
A critical change between the OIA and MIA stages, particularly in the Late MIA period, is
the restructuring of the nominal case system. Notable here is the loss of morphological
contrast between nominative and accusative as well as between the genitive and the
dative cases. The syncretized set of case-endings for full nouns are given in Table 1.
Table 1: Case-endings for full nouns.
-u, a, aṃ
-eṃ, iṃ, he, hi
-hu, ahu, aho
-ho, aho, ha, su, ssu
-i, hi, hiṃ
-a, aĩ
-e(h)ĩ, ehi, ahĩ
-hũ, ahũ
-na, hã
Table 2 contains an example of inflected -a stems with the noun putta ‘son’.
Table 2: Inflected a-stems with putta ‘son’.
The pronominal system retains more contrasts and syncretism between the nominative and accusative is observed only in the plural sub-part of most pronominal paradigms.
Table 3 (culled from Clercq 2010) provides inflectional forms for some pronominal expressions to illustrate.
The loss of contrast between the nominative and accusative cases in most paradigms
in a relatively free-word order language leads to heavy reliance on semantic cues from
the linguistic material to determine grammatical relations. Consider the following examples from the Paumacariu, an 8th century text in verse, to illustrate the syncretic
nominative-accusative marking (glossed nom).5 In (5), a sequence of parallel clauses,
This is a Jaina rendition of the Epic Sanskrit text Rāmāyana. The edition used is the H.C. Bhayani edition published by the Bharatiya Vidya Bhavan between 1953 and 1960. The text is available in searchable
electronic format, input by Eva De Clercq at Ghent University. The reason for using a late MIA text is to
identify properties of the system that is as close to the grammars of the Early NIA system as possible.
16 On mechanisms by which languages become [nominative-]accusative
Table 3: Inflectional forms for pronominals.
1st pronoun
mahu, majjhu
amhẽ, amhaiṃ
amhẽ, amhaiṃ
amha, amhaha
2nd pronoun
paiṃ, taiṃ
tahu, tujjha
tumha, tumhaha
3rd pronoun
so, su; sā
taṃ; sā
taho, tahu; tāhe
te, tāu
te; tāu
tāhaṃ; tāhaṃ
whether the first-occurring nominative expression realizes the grammatical subject or
the grammatical object is determined by the meaning of the clause.6 In (6), the relative
pronoun, which refers to a human participant, disambiguates the grammatical structure.
(5) #kiṃ tamu
ṇa vālu ravi#
qes neg young
#kiṃ vālu davaggi ṇa ḍah-ai
qes young neg
#kiṃ kari
ṇa vālu hari#
qes neg young
#kiṃ vālu ṇa ḍaĩk-ai
qes young neg
‘Does the young (rising) sun not destroy darkness? Does the young fire (spark)
not burn down the forest? Does a young lion (cub) not shatter the elephant?
Does the young snake not bite?’ (Paumacariu
(6) jo
ghañ ṇisi-bhoyaṇu
ummah-ai ptcl
‘One who gives up eating in the evening (he) attains a spotless body and name.’
Accusative marking is clearly visible only on first and second person singular pronouns in imperfective clauses as shown in the examples in (7).
The #…# marks clause boundaries in the sequence.
Ashwini Deo
a. suggīu
‘Lord Suggiu remembers you.’ (Paumacariu
b. jai ṇa vihāṇa-e
if neg
‘If I do not capture you tomorrow…’ (Paumacariu
c. jo
maiṃ muevi aṇṇu besides
‘(The one) who adores another one besides me…’ (Paumacariu
Syncretism rooted in sound change is also observed between the nominative and instrumental forms (the case form that gets re-interpreted as ergative when appearing with
agentive arguments in perfective clauses) of the first and second person plural pronouns
as in Table 4.
Table 4: Nominative and instrumental pronominal forms.
teṃ, teṇẽ
Despite this syncretism, agreement is uniformly with the nominative argument – with
the nominative object in constructions based on the -ta form and with the nominative
subject elsewhere. The examples in (8) illustrate this pattern with the first and second
person plural pronouns amhẽ and tumhẽ. (8a) contains the syncretized pronoun amhẽ
which triggers agreement in the imperfective aspect while the same form fails to trigger
agreement in (8b). In (8c) the second person plural syncretic form used in an imperative
clause triggers agreement while it fails to trigger verb agreement in the perfective (8d).
a. amhẽ
‘We are going to our forest-exile.’ (Paumacariu
16 On mechanisms by which languages become [nominative-]accusative
b. ki-u
ko avarāh-o we.syncr what
‘What crime have we done?’ (Paumacariu
utthar-aho tumhẽ
c. jiha
in.which.way in.that.way
yourselves in the way that you can.’ (Paumacariu
d. tumhẽ
‘That, which you thought (would happen), happened.’ (Paumacariu
These patterns of syncretization within the nominal inflectional system of MIA are difficult to reconcile with a story in which there is a straightforward extension of an existing
alignment pattern in the language to a marked sub-system of the grammar. Although
there is a contrast between the nominative and accusative cases in MIA, it is exhibited
only in selected parts of the pronominal system (a subset of the singular pronouns) and
therefore seems to be rather weak evidence for extending the accusative marking pattern
to ergative clauses. A reviewer argues that the regular presence of such a case-marking
pattern in imperfective clauses, however limited in terms of its application, should not
be seen as “weak” evidence for a nominative accusative pattern. I concede that it is indeed theoretically possible that the pattern observed in a small subset of imperfective
non-ergative clauses gets extended to perfective, ergative clauses. However, neither existing grammars of MIA (Pischel 1900; Vale 1948; Clercq 2010) nor an examination of the
textual data indicate any presence of accusative marked object arguments in perfective
transitive clauses at this stage in the language. Even pronominal objects (9a)–(9b) and
human-denoting full noun phrase objects (9c)–(9d) of canonical transitive verbs, which
obligatorily appear with overt accusative marking in the NIA languages, are uniformly
marked nominative at this stage.7
a. hauṃ ṇikkāraṇe
rām-eṃ without.reason Rā
‘Rām drove me out (of Ayodhya) without any reason.’ (Paumacariu
b. cakkesar-eṇa
kema tuhũ
di-ṭṭhī how
‘How were you noticed by Cakkesara (Rāvaṇa)?’ (Paumacariu
c. viṇivār-iu
rāhav-eṇa rāvaṇ rā
‘Rāhava (Rāma) dissuaded Rāvaṇa’ (Paumacariu
Thus, there are no positive instances with pronominal forms maiṃ, taiṃ, taṃ etc. being used instead of
hauṃ, tuhuṃ, or so/su etc. in ergative clauses with pronominal objects at even the latest stages of Middle
Ashwini Deo
d. di-ṭṭhu
rāhavacand-eṃ jaṇaddaṇ rā
‘Rāhavacanda saw Jaṇaddaṇa.’ (Paumacariu
Moreover, no language of the later stage (Early NIA) has an ergative-accusative marking pattern which uses the pronominal forms of late MIA in ergative clauses that accusative marking on objects. While the issue needs to be more closely investigated, it
seems reasonable to look for an alternative source for accusative marking in ergative
clauses than the template offered by MIA.
3 Differential object marking: A New Indo-Aryan
The previous subsection established that accusative marking of the MIA variety is both
weakly present and shows no evidence of being extended to perfective ergative clauses
at later stages of Indo-Aryan. This leaves the possibility that the incidence of object
marking in ergative clauses – a pervasive phenomenon in the Modern NIA languages –
begins with the Differential Object Marking pattern – which is considered to be an NIA
innovation. Differential Object Marking (henceforth DOM) in Indo-Aryan languages is
sensitive to animacy and referentiality features of arguments. It is obligatory on 1st and
2nd pronominal objects, and on 3rd person animate-denoting pronominals. It is optional
with animate-denoting full NPs where the absence of object marking correlates with a
non-referential interpretation of the NP. In the Modern NIA languages, this semantically
driven pattern of object marking does not distinguish between ergative and non-ergative
clauses; i.e. the case marking on objects is entirely independent of any overt or covert
presence of case on the subject.
Logically, one can imagine two ways in which an ergativity-insensitive object marking pattern can emerge in a system. It could be that the DOM pattern first emerges in
Late MIA or Early NIA in non-ergative clauses. Such a pattern is then later extended
analogically to ergative clauses as part of the de-ergativization trajectory characterizing Indo-Aryan diachrony. The second possibility is for the DOM pattern to emerge
simultaneously in both ergative and non-ergative clauses and gradually extend to different classes of verbs. On this latter scenario, the presence of DOM in ergative clauses
is not part of the larger de-ergativization trajectory that characterizes NIA diachrony,
but rather attributable to independent developments that introduce overt marking on
direct objects into the case system.8 The empirical facts of Late MIA and Early NIA texts
support the second scenario. In what follows, I will suggest that the emergence of DOM
in both ergative and non-ergative clause types of MIA amounts to the extension of an
inherited OIA marking pattern observed with the class of so-called “double object” verbs.
The effects on agreement in languages which exhibit such a changed case-marking pattern may be different.
In Modern NIA we see both default agreement in ergative clauses when both subject and object are casemarked (e.g. in Hindi, Marathi) or continued object agreement despite overt accusative marking on the
object (e.g. in Gujarati, Marwari).
16 On mechanisms by which languages become [nominative-]accusative
3.1 Double object verbs in Old Indo-Aryan
A class of verbs in OIA exhibits a double object pattern in which the theme or goal and
another participant of the denoted event are marked in the accusative case. Semantically,
this is a diverse class and includes at least the subclasses in Table 5.
Table 5: Double object verbs in Old Indo-Aryan.
Verbs of speaking
brū ‘speak’, vac ‘say’, kath ‘tell’
Verbs of asking
pṛcch ‘ask’, yāc ‘request, solicit’, bhikṣ ‘beg’, prārth ‘plead’
Verbs of teaching
upa-diś ‘teach’, anu-śās” ‘teach’, ā-diś ‘direct’
Causatives of some
khād-aya ‘cause to eat’, pā-yaya ‘cause to drink’, darś-aya
‘cause to see’, śrāv-aya ‘cause to hear’
jī ‘win, duh ‘milk’, daṇḍ ‘punish’, nī ‘lead’
(10) contains examples from OIA (Epic Sanskrit) involving verbs of speaking in imperfective, non-ergative clauses. In (10a), the pronominal denoting the addressee if the verb
of speaking event tvāṃ is accusative as is the information communicated, nidarśanam
‘the teaching’. (10b), from a proximal location in the text, is similar.
a. atas tvā-ṃ
hence karṇ
punaḥ again
‘Hence, O Karṇa, I tell you this teaching (advice) again.’ (Mahābhārata
b. śalyo
punaḥ karṇ-aṃ
ś again karṇ
‘Śalya again spoke out his advice to Karṇa’ (Mahābhārata 8.28.1c)
An alternative realization for pronominal animate-denoting higher arguments of double object verbs is as dat/gen clitics.
a. hanta te
kathay-iṣy-āmi nām-āni
ptcl you.dat/ here
Ashwini Deo
‘Ah, I will tell you the names of the wise ones.’ (Mahābhārata 1.48.4a)
b. … īś-ate
te you.dat/
‘The Lord alone reigns [over time and death and this universe of mobile and
immobile objects], this truth I tell you.’ (Mahābhārata 5.66.13c)9
In ergative, perfective clauses, this higher argument may surface variably: either as
the nominative subject of the passivized verb form (examples in (12)) or as a dat/gen
marked clitic pronoun (examples in (13)).10
a. uk-to
‘I was spoken to by the beasts at night.’ (Mahābhārata 3.244.11a)
b. ta-yā…
‘He was made to hear (these) words by her.’ (Mahābhārata 2.2.6a)
c. sa
mayā varadaḥ
‘He, the boon-granting one, was solicited by me for (fulfilling my) virtuous
desire.’ (Mahābhārata 1.78.3c)
a. sāṃkhyadarśan-am
etāvad uk-taṃ
sāṃkhyadarś so far you.dat/
‘Thus far, the Sāṃkhyadarśana was spoken to you, O best of kings.’
(Mahābhārata 12.295.1a)
b. tad etat
mayā vo
thus, you.dat/
‘Thus, I have told you all this, O great sages.’ (Mahābhārata 1.20.12a)
The previous line of verse completes the translation: kālasya ca hi mṛtyoś ca jaṇgamasthāvarasya ca (Mahābhārata 5.66.13a)
In (12a), the passivized subject is covert and the nominative case marking of the pro-dropped subject is
inferred from the agreement on the auxiliary verb asmi.
16 On mechanisms by which languages become [nominative-]accusative
c. upadiṣ-ṭo
hi me
yogo ptcl I.dat/
‘The method of penetrating into this (military) array has been taught to me
by my father.’ (Mahābhārata 7.34.19a)
d. brahmacary-am idaṃ bhadr-e
mama this
ca ādiṣ-ṭaṃ Dharmarā and
‘Good lady, this twelve-year celibacy has been commanded of me by
Dharmarāja.’ (Mahābhārata 1.206.21a-c)
The argument realization pattern illustrated in (11) and (13), where the higher argument of a double object verb surfaces with dative or genitive marking in both ergative
and non-ergative clauses, is fairly robust in OIA. The alterations to the nominal case system in MIA described in Section 2.2, have no effect on this pattern, since the syncretized
dat/gen remains available for overt marking throughout the period. Crucially, given
the organization of the MIA case system, this dative/genitive marking is the only reliably present overt marking on non-subject arguments in both ergative and non-ergative
clauses at this later stage. Based on the data from MIA, it seems most reasonable to conjecture that this template triggers the reanalysis of dat/gen as accusative marking on a
subset of direct objects.
3.2 Double object verbs in Middle Indo-Aryan
In (14) are given examples of the OIA double object verbs in their MIA incarnations.
Notice that themes surface with the syncretized nominative–accusative case (glossed
nom) while the non-theme higher argument (the addressee of the speech verb in (14a)–
(14b) and the causee in (14c)) appear with the syncretized dat/gen marking.11
a. sabbhāv-eṃ
ema rāma-dat/
‘He said this to Rāma with goodwill.’ (Paumacariu
b. mārui
Mā valadeva-dat/
‘Māruti told the news to Valadeva.’ (Paumacariu
c. ta-ho
ajju jamattaṇu
he-dat/ now
‘Now, I will show him the prowess of Yama (the god of death).’ (Paumacariu
(14a) and (14c) have subject pro-drop.
Ashwini Deo
A look at perfective, ergative clauses in MIA containing double object verbs reveals
overt dat/gen marking on the non-theme argument and unmarked themes. (15a) contains the causative of a perception verb, while (15b)–(15c) contain verbs of speaking. Just
like OIA, there is no difference between ergative and non-ergative clauses vis-à-vis the
realization of non-subject arguments.
a. paḍ-e
daris-āv-iya Sita-dat/
‘(He) showed the image of Sita on a screen (painting) to Bhāmaṇḍala.’
b. kah-iu
parama-jiṇind-eṃ I.dat/
‘The great Jinendra told (this) to me.’ (Paumacariu
c. ta-ho
maiṃ parama-bheu
‘I have told you this great secret.’ (Paumacariu
In addition to the non-theme arguments of double object verbs, the syncretized dat/
gen marking also appears on possessor and goal arguments of standard ditransitives
(examples in (16)) and on themes of verbs that describe a reciprocal experience (examples
in (17)).
a. kikkindha-ho
‘She garlanded Kikkindha (lit. put a garland on)’ (Paumacariu
b. paripes-iu
aṇaraṇṇa-ho chief-dat/ Anaraṇya-dat/
Ayodhyā-dat/ king-dat/
‘(He) sent a letter to Anaraṇya, the king of Ayodhya’ (Paumacariu
c. aṅgutthala
ṇav-evi samapp-iu
tāvahñ mahu bow-ger then I.dat/
‘(After) I handed her the finger ring, having bowed to her, (she) gave me this
precious gem.’ (Paumacariu
16 On mechanisms by which languages become [nominative-]accusative
d. diṇṇa
maiṃ dasaraha-taṇay-aho dasaraha-son-dat/
‘I have given my daughter to the son of Dasaraha (Daśaratha).’ (Paumacariu
a. salil-u
jiha milai ocean-dat/ as
‘Just as the water meets the ocean’ (Paumacariu
b. tāvehñ gayaṇa-ho oar-evi
then, sky-abl descend-ger Añjanā-dat/ Vasantamāla.nom
‘Then, having descended from the sky, Vasantamālā met Añjanā.’
Critically, the syncretized dat/gen marking is the only reliable signal of non-subject
arguments in MIA and it appears without discernible difference in distribution in both
ergative and non-ergative clauses. It does not however appear, for the most part, on
theme/patient arguments of canonical transitive or ditransitive verbs – animate or otherwise. (18a)–(18b) are examples of ergative clauses with animate-denoting subjects while
(18c) contains a non-ergative clause.
a. hā vahue-vahue mañ
alas bride.voc
‘Alas, O bride, I drove you out without testing you in any way.’ (Paumacariu
b. ṇi-u
tettahe sapparivāru there
jettahe where
‘(She) took the lord of the three worlds there where Purandara was with his
family.’ (Paumacariu
c. muṇivara
‘The king will drive out the sages.’ (Paumacariu
3.3 The emergence of DOM
The key suggestion I make here is that the Indo-Aryan differential object marking pattern emerging between late MIA and Early NIA amounts to the generalizing reanalysis of
syncretic dat/gen marking on non-subject non-theme arguments as accusative marking
Ashwini Deo
on (a privileged class of) objects. The data that provide evidence to enable such a reanalysis are clauses containing double object and other ditransitive verbs which either have
implicit (non-overt) theme arguments or where the arguments (in the case of verbs of
speech) are propositional. Such clauses are not very frequent but they do occur quite
reliably in MIA. Examples of non-ergative clauses are given in (19) and ergative clauses
are in (20).
a. akkh-ai
samīraṇa-putt-aho Samīraṇaputta-dat/
‘Sita told Samīraṇa-putta (this).’ (Paumacariu
b. kahai
that.dat/ Lavaṇa-dat/ capable-dat/
‘The great sage said to that Lavaṇa, who was capable in battle (thus).’
a. aṭṭhāvaya-giri-kampāvaṇ-aho
‘The messenger told (this) to Ravana, who was capable of causing the eight
territories (aṣṭapada) to tremble.’ (Paumacariu
b. to pamiṇipura-paramesar-aho daris-āv-iya
then pamiṇipura-lord-dat/
‘They showed (the boys) to the lord of Pamiṇipura (Padminipura), the king
Vijayaparvata.’ (Paumacariu
c. añjaṇ-ahe
jāya dih-i
Añjanā-dat/ birth
‘They handed him (the baby Hanumān) to Añjanā on the day of his birth.’
In clauses such as those in (19) and (20), the only overt non-subject argument carries
dat/gen marking. Moreover, this pattern of marking does not differentiate between
whether the subject carries ergative marking or is unmarked (nominative).
Consider a learner that must arrive upon the case inventory of a language based on
the observable input. The MIA system provides reliably present morphological evidence
for nominative, ergative, and dative/genitive case but no reliable evidence for accusative
case. It also provides robust data in which the only non-subject argument overtly expressed in a clause carries case marking (the syncretic dat/gen marking). It is possible
that the learner takes this subset of data as evidence for extending the dat/gen marking,
16 On mechanisms by which languages become [nominative-]accusative
reserved for non-theme arguments, to theme and patient arguments as well. The Differential Object Marking pattern evidenced in Early NIA emerges because the analogical
extension of the overt dat/gen marking is constrained by the semantic properties associated with the original class of arguments marked by it – animacy and referentiality. If
this hypothesis is correct, then we expect that there may be early data supporting this
extension of dat/gen case marking to direct objects – in effect, the reanalysis of dative
marking as accusative case, restricted to arguments meeting the criteria of high animacy
and referentiality.
In the previous subsection, it was claimed that as far as the MIA stage is concerned,
direct arguments of canonical transitive verbs do not, for the most part, surface with
dat/gen marking (examples in (18)). The caveat was provided precisely because the MIA
stage itself seems to exhibit some data which is possibly analyzable as emergent DOM.
The tentativeness with which this claim can be made emerges from three uncertainties
about the data: (a) Although the lexical verbs appearing with the dat/gen marked objects arguably have an argument structure corresponding to transitive verbs and their
translational equivalents in English are realized as canonical transitives, given the semantics of these verbs, it seems possible that they pattern either with ditransitives or
with “reciprocal” verbs” or with intransitives having accusative goal arguments in Sanskrit. Thus, it is necessary to investigate more closely whether these cases are early
DOM-instances or whether they should be reclassified as exhibiting previously occurring patterns (b) The object-marking pattern is very infrequent outside of the class of
double-object verbs, other ditransitives, and “reciprocal verbs”. (c) There is absolutely
no example of perfective clauses with ergative subjects in which the object appears with
dat/gen marking.
It is possible therefore that the human-denoting dat/gen marked NPs in the data
below are not the theme/patient arguments in a standard transitive template as they
appear to be; they may be better analyzed as recipient or goal arguments. I will leave
the adjudication of this issue for further research. But regardless of their status, they
provide further surface evidence to the language acquirer for an object marking case
“accusative” in the language.
In (21) and (22), we see that the human-denoting non-subject arguments of the transitive verbs khama ‘forgive’, pekkha ‘look at’, garaha ‘denounce, curse’, abhiṭṭa ‘attack’,
ḍhukka ‘approach’ and bhiḍ ‘battle’ appear with dat/gen marking. The examples in (21)
contain non-perfective clauses ((21b) is an imperative) while those in (22) illustrate the
argument realization pattern in perfective clauses.
a. ekkavāra ma-hu
one.time I-dat/ forgive-imp.2.g
‘O warrior (Lakshmana), please forgive me one time’ (Paumacariu
b. sundari
jujjh-ant-aho fight-part.dat/
‘O beautiful one, look at the battle.’ (Paumacariu
Ashwini Deo
c. ema jāma garah-anti
thus when jiṇinda-dat/
tāma dharaṇind-aho then dharaṇinda-dat/
‘When they were denouncing Jiṇinda thus, the seat of Dharaṇinda started to
shake.’ (Paumacariu
d. ham
dūsaṇ-aho Dūsaṇa-dat/
‘I will attack Dūsaṇa’ (Paumacariu
a. dhā-iu
abbhi-ṭṭu aṅ lakṣmaṇa-dat/
lavaṇ rāma-dat/
‘Aṅkuṣa ran to Lakshmaṇa (while) Lavaṇa attacked Rāma’ (Paumacariu
b. kattha
vi bhaḍ-aho
sivaṅgaṇa ptcl warrior-dat/
‘At some places (on the battlefield), she-jackals approached the (dead)
warriors.’ (Paumacariu
c. indai
haṇuvant-aho haṇuvant-dat/
‘Indai (Indrajit) battled with Haṇuvanta in the battlefield.’ (Paumacariu
3.4 The DOM pattern in Early New Indo-Aryan
Turning to the Early New Indo-Aryan stage (illustrated here with Old Marathi), we see
a clearly established animacy- and referentiality-sensitive DOM pattern in both ergative
and non-ergative clauses from the earliest period.12 The syncretic dat/gen marking of
MIA appears as a generalized oblique case and it is augmented with innovated postpositions that correspond to accusative and dative case markers. This trajectory, in which
the MIA case-system reduces to a nominative/oblique contrast and new postpositions are
innovated to convey the semantic and structural information associated with the older
cases, is shared across Indo-Aryan languages (Masica 1991; Bubenik 1996; 1998 a.o).
This period is represented here by two texts – Līḷācharitra (ca. 1286 CE, prose) and the Dnyāneśvarī (ca.
1287 CE, verse).
16 On mechanisms by which languages become [nominative-]accusative
Direct objects in Old Marathi surface with an innovated postpositional accusative
clitic, -teṃ, attached to the oblique stem (the reflex of the MIA dat/gen marker). The examples selected for presentation here contain transitive verbs whose animate-denoting
theme arguments in both ergative and non-ergative clauses appear with overt accusative
marking in (23)–(25).
a. āmhīṃ tuma=teṃ ne-unuṃ
‘We will take you (to Varanasi).’ (Līḷācaritra 1.25)
b. aiseṃ mhaṇ-auni yā=teṃ
śrīkarī-ṃ dhar-ūni āpuleyā
thus speak-ger this.obl=acc hold-ger
‘Having spoken thus, taking him by the hand, she took him to her house.’
(Līḷācaritra 1.34)
a. mhaṇoni prakāśā=ce=ni=hi
na dekh-atī
therefore light.obl=of=by=ptcl neg
‘Therefore, even by the strength of light, they do not see me.’ (Dnyāneśvarī
b. tehīṃ
‘He saw this one from the upper story (of the house).’ (Līḷācaritra 1.6)
a. āṇi te
āma=teṃ dhari-tī
And we-acc
‘And they (honorific) would catch us.’ (Līḷācaritra 1.18)
b. ekī-ṃ
ākāś-īṃ sūryā=teṃ dhar-ileṃ sun.obl=acc
‘Someone (might) catch the sun in the sky.’ (Dnyāneśvarī 10.0.37)
The examples in (26) contain the same non-animate denoting but referential argument
jaga ‘world’ that also receives accusative marking in both imperfective and perfective,
ergative clauses ((26a) and (26b) respectively).
a. maga āpu-leṃ
āṇī jagā=teṃ
then and world-acc
‘Then they proclaim their own deeds and denounce the world.’ (Dnyāneśvarī
Ashwini Deo
b. prabaṃdhavyāj-eṃ jagā=teṃ rakṣ-ileṃ
jāṇa world-acc
‘Know that (the Guru) has saved the world through this literary work.’
(Dnyāneśvarī 18.78.1765)
It is necessary to take a much closer look at the pattern of DOM seen in Old Marathi
languages and compare it on a verb-by-verb and argument-type by argument-type basis
with the MIA pattern. It is only such an investigation that can accurately establish the
nuanced differences between the impoverished accusative marking of Late MIA and the
innovated accusative marking of Old Marathi. Noteworthy is the fact that no reflexes
of the MIA accusative marking survive in the pronominal system of Old Marathi; only
traces of the syncretized gen/dat marking remain.
4 Conclusion
At first glance, the presence of accusative marking (DOM) in NIA ergative clauses could
be considered to be a case in which an existing template from the imperfective domain
is extended by analogy to the perfective ergative domain. However, a closer study of
the case-marking patterns of Late MIA reveals that there is no evidence for any direct
extension of the MIA accusative marking to ergative clauses. It is more likely the case
that the DOM pattern emerges in NIA languages as a reanalysis of the MIA dat/gen
marking that appears systematically on a specific subset of non-subject arguments into a
marker of accusative case. This reanalyzed accusative case is attested in both ergative and
non-ergative clauses in the earliest texts of Old Marathi, supporting the hypothesis that
accusative marking in ergative clauses is not part of any “de-ergativization” trajectory
in the history of Indo-Aryan but rather an emergent effect of across-the-board changes
in argument realization options for the languages.
Glosses are as follows. “-” stands for a morpheme boundary, “=” for a clitic boundary.
(Old Indo-Aryan Present)
impfct Old Indo-Aryan Imperfect
16 On mechanisms by which languages become [nominative-]accusative
discourse particle
verb particle
syncretic (nom/inst)
Support from the National Science Foundation (NSF BCS-1255547/BCS-1660959.) is gratefully acknowledged.
Andersen, Paul Kent. 1986. Die ta-Partizipialkonstruktion bei Aśoka: Passiv oder ergativ?
Zeitschrift für vergleichende Sprachforschung 99. 75–95.
Anderson, Stephen R. 1977. On mechanisms by which languages become ergative. In
Charles Li (ed.), Mechanisms of language change, 317–363. Austin: University of Texas
Anderson, Stephen R. 2004. Morphological universals and diachrony. In Geert Booij &
Jaap van Marle (eds.), Yearbook of morphology 2004, 1–17. Dordrecht: Springer.
Bloch, Jules. 1965. Indo-Aryan from the Vedas to Modern Times. Translated by Alfred Master. Paris: Adrien-Maisonneuve.
Bubenik, Vit. 1996. The structure and development of Middle Indo-Aryan dialects. Delhi:
Motilal Banarsidass.
Bubenik, Vit. 1998. Historical Syntax of Late Middle Indo Aryan (Apabhramsa). Amsterdam: John Bejamins Publishing Co.
Clercq, Eva de. 2010. The Apabhraṃśa of Svayambhūdeva’s Paumacariu. Delhi: Motilal
Comrie, Bernard. 1978. Ergativity. In Winfred P. Lehmann (ed.), Syntactic typology: Studies in the phenomenology of language, 329–394. Austin: University of Texas Press.
Condoravdi, Cleo & Ashwini Deo. 2014. Aspect shifts in Indo-Aryan and trajectories
of semantic change. In Chiara Gianollo, Agnes Jäger & Doris Penka (eds.), Language
change at the syntax-semantics interface, 261–292. Berlin: Mouton de Gruyter.
Coon, Jessica & Omer Preminger. To appear. Split ergativity is not about ergativity. In
Jessica Coon, Diane Massam & Lisa Travis (eds.), Oxford Handbook of Ergativity. Oxford: Oxford University Press.
Delancey, Scott. 1981. An interpretation of split ergativity and related patterns. Language
57(3). 626–657.
Deo, Ashwini. 2012. The imperfective–perfective contrast in Middle Indic. Journal of
South Asian Linguistics 5. 3–33.
Dixon, Robert M. W. 1979. Ergativity. Language 55. 59–138.
Ashwini Deo
Dixon, Robert M. W. 1994. Ergativity. Cambridge: Cambridge University Press.
Hock, Hans Henrich. 1986. Principles of historical linguistics. Berlin: Mouton de Gruyter.
Masica, Colin. 1991. The Indo-Aryan languages. Cambridge: Cambridge University Press.
Moravcsik, Edith. 1978. On the distribution of ergative and accusative patterns. Lingua
45. 233–279.
Peterson, John M. 1998. Grammatical relations in Pāli and the emergence of ergativity in
Indo-Aryan. München: Lincom Europa.
Pischel, Richard. 1900. Grammatik der Prākrit-Sprachen. Translated from German by Subhadra Jha, 1981. Delhi, India: Motilal Banarsidass.
Plank, Frans (ed.). 1979. Ergativity: Towards a theory of grammatical relations. New York,
NY: Academic Press.
Tsunoda, Takasu. 1981. Split case-marking patterns in verb types and tense/aspect/mood.
Linguistics 21. 385–396.
Vale, Ramchandra N. 1948. Verbal Composition in Indo Aryan. Poona, India: Deccan College Postgraduate & Research Institute.
Chapter 17
OF as a phrasal affix in the English
determiner system
Randall Hendrick
University of North Carolina at Chapel Hill
This paper examines the distribution of of in the complex determiner system of English
nominal expressions. The hypothesis is advanced that of in degree phrase inversion constructions (expressions like more influential of a book) is a phrasal affix in the sense of
Anderson’s classic work asserting the autonomy of morphological theory. The phrasal affix
analysis can be distinguished from syntactic accounts that treat of as a syntactic head of
phrase. It is argued that the phrasal affix account can capture of’s sensitivity to morphological sub-category features and to second position more naturally than the syntactic head of
phrase analysis. In contrast to the syntactic head analysis, the phrasal affix analysis also
has the advantage of explaining why of fails to exhibit typical syntactic properties such as
selection, dislocation, conjunction, and adjunction of focus particles. The analysis extends
neatly to the use of of in complex determiners involving fractions.
1 Introduction
Anderson (1992) has argued that the classical distinction between words and affixes can
be profitably extended to the phrasal domain, a line of argumentation that is extended in
Anderson (2005). On this view, some properties of constructions might be better understood if analyzed as involving affixes attached to phrases, parallel to the attachment of
affixes to words. Anderson identifies the possessive ’s as the prototype of such a phrasal
affix, but (special) clitics more generally are identified as phrasal affixes in this sense.
In the same spirit of enlarging the explanatory work of morphological theory into the
domain of phrasal syntax, Anderson (2005) argues carefully that some verb second phenomena should be given a morphological explanation parallel to second position clitics.
Both claims have proven provocative because they seem to compete with other rather
popular analyses that make use of syntactic movement.1 In this paper I will argue that of
The true difference between these types of analyses may be less than meets the eye if one assumes a
syntactic engine like that advocated in Chomsky (1995). There is a real substantive difference in the embrace
of optimality theoretic explanations in Anderson (2005), but I will ignore that issue here in the belief that
Randall Hendrick. 2017. OF as a phrasal affix in the English determiner system. In
Claire Bowern, Laurence Horn & Raffaella Zanuttini (eds.), On looking into words (and
beyond), 369–384. Berlin: Language Science Press. DOI:10.5281/zenodo.495454
Randall Hendrick
as it occurs in the pre-determiner system of English provides compelling corroboration
for the attempt to extend the range of morphological analysis into at least some areas of
phrasal syntax because alternative syntactic explanations for these instances of of are
relatively weak.
The center of our attention is on the formative of as it occurs in sentences like (1)(2). These sentences appear to have some string before the indefinite D a(n) as part of
the bracketed nominal constituent. They have been the focus of attention in a number
of works, including Bolinger (1972), Bresnan (1973), Hendrick (1990), and Kennedy &
Merchant (2000), among others. Kim & Sells (2011) present a wide variety of naturally
occurring examples of the construction. A range of proposals have been offered to provide syntactic accounts of such sentences. Such sentences seem closely related to (4), in
which the adjective abuts a directly without the cushioning of of , and (3), in which the
adjectival constituent appears to the right of a rather than to its left. Kennedy & Merchant (2000) label (4) and (1) as inverted DegP (i.e., inverted degree phrases), suggesting
that they originate as (3) and subsequently move to a higher position.
It is difficult to find [ more influential of a book ]
[ How long of a vacation ] did she take?
It is difficult to find [ a more influential book ]
(4) It is difficult to find [ more influential a book ]
The degree phrases that are implicated in this inversion are semantically fairly heterogenous: a great many, like (1), appear to be evaluative in the sense of Keenan & Faltz
(1985), but others, like (2), are extensional and amenable to a model theoretic interpretation.
There is another family of constructions that exhibit of preceding indefinite [D a(n)].
These are the complex determiners that have figured prominently in the study of the
semantics of determiners and general quantifiers. Keenan & Stavi (1986) noted a range
of complex determiners as part of their attempt to establish the semantic conservativity
of a large set of determiners. Many of those determiners include of to the left of a, and
denote proportions or frequencies, as noted in Peters & Westerståhl (2006). The examples
in (5) list some of the complex determiners involving of .2
They sang for half of an hour.
She ate two thirds of an orange.
The tornado damaged one quarter of an entire neighborhood.
the existence of phrasal affixes is conceptually independent of any commitment to optimality theoretic
2 Peters & Westerståhl (2006) observe sentences like two of ten interviewed persons have not answered the
question. The analysis in the text does not generalize to such an example because it does not exhibit [D
a(n)]. Unlike the examples examined in the text, a plural NP is involved here as well. The of in this example
would have to be accounted for separately. The fact that such examples can occur with out before of in
contrast to (1) could be evidence supporting this inference.
17 OF as a phrasal affix in the English determiner system
My principal claim is that of is inserted in (6) as an affix to D̄, formally marking the
relation between the head D and its specifier when α is phonologically non-null.
This affixed of is evinced in all of (1), 2 and (5). As a marker of a formal relation,
it itself is not semantically evaluated, and in this sense, it is not meaningful. As an
adjunct, It is also syntactically inert for movement processes that target specifiers, heads
or complements.3
I have framed my claim in terms of the DP hypothesis in (6), where D is the head of
the entire nominal phrase. If one were instead to assume the NP hypothesis, where N is
the head of the nominal phrase, we would have structures like (7). In this case we would
say that of is adjoined as an affix of D̄, formally marking the relation between D̄ and the
phrase α to its left when α is non-null.
2 OF as a phrasal affix
There is a prima facie case to be made that of in examples like (1) and (5) are phrasal
affixes. The insertion of of is sensitive to morphological properties and shows second
position effects common to other putative examples of phrasal clitics that Anderson
2.1 OF as sensitive to morphological properties
The inversion of degree phrases in examples like (1) does not apply generally. It is unnatural if the degree phrases are an instance of unmodified synthetic comparatives, as
illustrated in (8).
See Kaufman (2010) for a broadly similar analysis of Tagalog and other Austronesian clitic systems. In
lexical theories of the sort that Anderson has generally favored, the syntactic inertness of the phrasal clitic
will follow from the post-syntactic affixation of the phrasal clitic.
Randall Hendrick
a. We are shopping for a better/prettier/finer vase.
b. * We are shopping for better/prettier/finer a vase.
c. * We are shopping for better/prettier/finer of a vase.
When synthetic comparatives are modified, inversion of degree phrases, as in (9), is
acceptable for some speakers, although others reject the inversion of degree phrases
even when modified.
a. * We have never seen a any better/prettier/finer vase.
b. We have never seen any better/prettier/finer a vase.
c. We have never seen any better/prettier/finer of a vase.
Speakers who reject all synthetic comparatives in (8)–(9) treat inversion of degree
phrases as sensitive to that morphological class. Speakers who accept (9) in contrast to
(7) appear to treat the inversion of degree phrases as sensitive to the prosodic contour
of the phrase. Degree phrases composed of a single prosodic word, as in (8), are excluded. When more prosodic material is added, the same synthetic comparatives appear
acceptable once again, as in (9).
The inversion is also sensitive to the morphological properties of the D to its right.
The D cannot be definite, as shown by (10b) and (11b). Of the indefinite D’s, it is only
able to employ a(n), and resists co-occurring with some or no.
a. How long a novel did she write?
b. * How long the novel did she write?
c. * How long some novels did she write?
d. * How long no novel did she write?
a. She wrote more fascinating a novel this time.
b. * She wrote more fascinating the novel this time.
c. * She wrote more fascinating some novels.
d. * She wrote more fascinating no novel.
If a(n) has the features [+INDEF, -PLURAL] while some is [+INDEF, +PLURAL] and
no is [+INDEF] but lacks a feature specification for plurality, we can stipulate for the
explicitness of description that degree inversion is restricted to D’s that carry the feature
specification +INDEF, -PLURAL.4
It is worth noting that there is another formative some that is stressed and is definite. In addition, the
table is offered only to make explicit the morphological use of indefinite in the text. It is not a theoretical
claim. If we were to decide that we needed to specify no as PLURAL we could appeal to a feature CARD for
cardinality and stipulate that degree phrase inversion could only occur with D’s specified as -CARD. By the
same token, we might have reason for adopting a morphological theory that avoided features altogether.
In any scenario we need to restrict Degree Inversion to the D a(n) and not other members of the D class,
much as /n/ is added to /a/ and not other indefinites.
17 OF as a phrasal affix in the English determiner system
Table 1: Co-occurrence of indefinite D’s with count nouns
Indefinite D
Singular N
a novel
*some novel
no novel
Plural N
*a novels
some novels
no novels
2.2 OF as sensitive to second position
The of described in the preceding subsection is not capable of bearing stress. Nor is it
able to appear initially in a nominal domain. When it surfaces, it always occupies second
position in the nominal expression. To see this point, consider sentences like (12) that
contain two pre-nominal adjectives. If long is part of a wh-phrase, it preposes (as in
(12b)), and if romantic is part of a degree phrase, it may prepose, as in (12c). However, if
both are before a, the result is the unacceptable (12d).
a. Jane Eyre is a long romantic novel.
b. How long of a romantic novel is Jane Eyre?
c. Jane Eyre is more romantic of a long novel (than Middlemarch).
d. * How long more romantic of a novel is Jane Eyre?
We begin to capture these facts if we assume that the preposed wh-phrase in (12b) and
degree phrase in (12c) target a (unique) specifier position for movement (i.e. specifier of
DP). We might accomplish this by way of (13a) that gives a(n) an EPP feature, parallel to
what is often assumed for T. We could then posit a spell out rule that prefixes of .
a. There is an optional feature on [D a] that requires a filled specifier.
b. This feature is optionally spelled out as the phrasal prefix of .
3 Could OF be a syntactic head of phrase?
Some researchers have suggested that the of that occurs in the Degree Inversion structures is a syntactic head of phrase.5 On this view we should assign how long of a vacation
the structure in (14), whereas the proposal I have suggested in (13) is (15), if we assign of
to the category F and adopt an item and arrangement approach to affixation in order to
facilitate the comparison of the two theoretical viewpoints.6
This structure is advocated in Kennedy & Merchant (2000) and is assumed in other work, for example,
Borroff (2006). The labels of constituents differ in other analyses, in particular Matushansky (2002) and
Kim & Sells (2011), but the constituency is broadly the same. The crucial point is to contrast analyses that
treat of as a head of phrase from the analysis defended in the text where of is a phrasal affix.
In a word-and-paradigm approach of might not be given a structural position at all but simply spelled out
at the left bracket D̄. This point of view is probably closer to Anderson’s. I believe that the structure in (15)
gives some insight into why of occurs between the D and its specifier, as I explain at the end of this paper.
Randall Hendrick
how long
how long
how long
how long
Some authors, following the lead of Bennis, Dikken & Carver (1998), have tried to treat
of as parallel to a copular construction in which predicates have been claimed to invert
in (16).
a. Jill was a much better friend.
b. A much better friend was Jill.
Den Dikken (2006) offers an extended analysis in this vein over a range of languages.
However, I find compelling the argumentation in Heycock & Kroch (1999) that rejects
the inversion analysis in (16) in favor of viewing these as identity statements. On this
view there is no basis for the parallel between (16) and (14), as Heycock and Kroch note.
17 OF as a phrasal affix in the English determiner system
Four reasons lead us to avoid a structure like (14).7 First, heads of phrases are able to
stand in a selection relation with other heads of phrases. Yet, there is no lexical head
that ever selects the FP of (14). This fact is accidental if of is a head of phrase, but it is a
necessary consequence if it is not actually a head, as in (15).8 Second, the of phrase in
(14) is immune to syntactic dislocation. Other of phrases are able to prepose or postpose,
as shown in (17) and (18). It is also possible to strand of in examples parallel to (19). In
this sense, they are syntactically active.
Three of these books have been reviewed.
Of these books three have been reviewed.
The announcement of the Nobel Prize in Chemistry was delayed.
The announcement was delayed of the Nobel Prize in Chemistry.
(19) What book did the newspaper publish a review of?
Yet (14) stands in contrast to examples like (17) in being unable to be preposed, postposed, or stranded. They appear to be syntactically inert.9
a. More extensive of an experiment was in the planning stages.
b. * Of an experiment more extensive was in the planning stages.
a. More extensive of an experiment was delayed
b. * More extensive was delayed of an experiment
*What (kind of) experiment did the journal report more extensive of?
However, if of is not a head of phrase with its own phrasal projection as in (15), the lack
of syntactic dislocation is expected.
A third challenge to the analysis in (14) involves conjunction. Locative prepositions
like on can be repeated members of a conjunction with only, presumably for pragmatic
reasons related to defeasing conversational implicatures that would otherwise obtain.
The same seems true of of that introduces the complement of derived nominals like
Some work has suggested that the fronted how long must originate as a predicate adjective. See, for example,
Wood & Vikner (2011) and Troseth (2009). This seems unlikely, at least for English. Expressions like a
beautiful dancer are ambiguous between a predicative reading in which the dancer is beautiful and a nonpredicative reading in which the dancer dances beautifully. The ambiguity is preserved in how beautiful (of)
a dancer. This fact is surprising if the preposed degree phrase could only arise from a predicate adjective.
These facts are general and can be reproduced with a skillful manager.
8 This challenge can be given a sharper formulation if we accept the stipulation that the complement of V is
reserved for selected arguments of V. This stipulation requires that V select of , if of is a head of phrase. To
the extent that every V that selects DP also selects this of we miss a regularity in the system of selection, a
cost that is not incurred if of is provided as a phrasal affix. One can imagine mechanisms that allow of , if it
were a head of phrase, to inherit argument properties of its complement DP, but such mechanisms would
obscure the generalization that of is syntactically inert generally.
9 A reviewer observes that this second challenge can be met by stipulating that only maximal phrasal projections are available for movement. While this defense of the head of phrase analysis will side step the
second challenge, it does not generalize to meet the other challenges. For this reason the phrasal affix
account seems to me to offer a more straight forward explanation, all other things being equal.
Randall Hendrick
You may place the notebook on, and only on, the desk.
The tornado caused the destruction of, and only of, one residence.
You may place the notebook only on the desk.
The tornado caused the destruction only of one residence.
In contrast, the of that appears with degree phrase inversion is unable to appear in similar conjunctions. If of is syntactically and semantically inert, as the phrasal affix analysis
contends, it could not conjoin and co-occur with only to defease any conversational implicatures.
a. You can’t find more valuable of a proposal.
b. * You can’t find more valuable only of a proposal.
c. * You can’t find more valuable of, and only of, a proposal.
d. How long of a vacation did they take?
e. * How long of, and only of, a vacation did they take?
A fourth reason to be skeptical of the analysis in (14) is that it is unable to generalize
to explain the appearance of of in fractions. Fractions have been treated as complex
determiners since the classic work of Keenan & Stavi (1986) on the conservativity of determiners. Fractions exhibit this semantic property, a fact that is easy to explain if they
are (complex) determiners, but that looks accidental if they are assimilated to constructions that take of to be a head of phrase, like the partitive constructions discussed in the
next section. The use of of in the fraction in (25) parallels its use in degree phrase inversion. In both cases of is adjacent to the indefinite D a(n), in both cases of is optional,
and in both cases of lacks the ability to be lexically selected, or syntactically preposed
or stranded as illustrated in (25). The phrasal affix analysis in (15) can provide a unified
analysis, on the assumption that fractions occur in specifier of DP.10
a. The recipe called for one half of a cup.
b. * Of a cup the recipe called for one half.
c. * A cup the recipe called for one half of.
d. * The recipe called for one half of, and only of, a cup.
4 Partitive OF is a syntactic head of phrase
There is a second construction involving of which should be distinguished from of as a
phrasal affix in Degree Inversion structures. This second type of construction is exemplified in (26) and has traditionally been labeled a partitive.
It is tempting to extend this analysis of fractions to proportions like six of ten adults don’t vote. Proportions
like this do not make use of the D a(n) and are plural rather than singular. Without a more detailed analysis
of the proper structural analysis of cardinals like ten, and an understanding of why proportions admit some
form of dislocation (e.g. (out) of (every) ten adults six don’t vote) the extension of the analysis will have to
remain incomplete.
17 OF as a phrasal affix in the English determiner system
a. She invited ten of the students to the party.
b. Jack offered to buy any three of her paintings.
c. Sandy knows the answer to the last (one) of those questions.
The of in these partitives is a syntactic head of phrase, appearing in structures like (27),
where of heads a PP. With the exception of half (of) the, partitive of appears obligatorily,
like other heads of phrases and in contrast to the pattern of of with Degree Phrase
Inversion. In addition, the of phrase can be topicalized as a syntactic constituent. In this
respect it differs from of in Degree Inversion constructions like (20).
a. Of the students, she invited ten to the party.
b. Of her new paintings, Jack offered to buy any three.
c. Of those questions, Sandy knows the answer to the last (one).
The of phrase can also be co-ordinated, a standard test for constituency. Examples are
somewhat limited, presumably because of semantic restrictions on partitives. However,
naturally occurring examples are not difficult to find.11
a. If you accept it to be a part of you and of your life, you gain control of the
b. I try to be very aware of all of the students and of their strengths and weaknesses.
In contrast, Degree Inversion structures do not show the same support for co-ordination.
a. * More expensive of a house or of a car would be hard to find.
b. * How good of a novel or of a play did she write?
It is also possible to use expressions like only with partitive of , in contrast to the of in
Degree Inversion structures.
She invited ten of the linguistics students, and only of the linguistics students,
to the party.
b. * How good of a novel, and only of a novel, did she write?
(29a) occurs at as accessed
September 24, 2016. (29b) occurs at
htm and was accessed September 24, 2016.
Randall Hendrick
My claim that of in Degree Inversion structures is structurally distinct from of in
partitives can be better appreciated if it is contrasted with the classic analysis presented
in Bresnan (1973).12 Bresnan suggests that of should be inserted in examples like (33) by
an elegant rule like (32).13
∅ → of / Q D N
(33) She has enough of a problem as it is.
more of an egg
more of the egg
(32) is similar to the phrasal clitic analysis I suggested earlier in that of is inserted rather
than being treated as a syntactic head of phrase. It differs from the phrasal clitic analysis
in two ways: (i) it does not require Q to be complex, and (ii) it does not require D to be
The first of these differences is more apparent than real. I say this because Bresnan
argues that more originates as a syntactically complex constituent -er much, and that
this complex constituent is replaced in the course of the derivation with more. Bresnan
analyzes enough in a parallel complex structure, with the difference that enough necessarily co-occurs with a null specifier. If this complex structure is correct, there is a point
in the derivation where (34) is not a counter-example to our phrasal affix account of of ;
we only need to insure that the phrasal clitic of is inserted prior to -er much suppletion.
Consider now the second way in which rule (32) differs from the phrasal clitic analysis
that I suggested in §2. By inserting of before any D, the rule in (32) loses the important
empirical generalizations about the way in which of in (35) patterns like a syntactic
head of a phrase in terms of conjunction, the distribution of only, and the possibility
of movement. It does not code naturally the way in which the of in Degree Inversion
constructions patterns differently along these dimensions. In the absence of other evidence, considerations of simplicity would make (32) quite attractive, but in this case,
that very simplicity obscures important ways in which of patterns in distinct, rather
than haphazard, ways. Moreover these distinct patterns have a semantic reflex. Peters
& Westerståhl (2006) argue that semantic definiteness is a necessary characteristic of
partitives. From this perspective the distinction between (34) and (35) that the phrasal
affix analysis encodes is a natural one.
5 OF in complements of nominalization and syntactic
Case theory
Since Chomsky (1981) it has been popular to say that the presence of of after nominalizations like picture in (36) is required to provide the nominalization’s complement (Mary
Stockwell, Schachter & Partee (1973) also distinguished partitives with of from other constructions where
of surfaces. Selkirk (1977) distinguishes true partitives from pseudo-partitives like two pounds of turkey in
part by pointing to differences in the distribution of of in the two types of constructions.
Q in (32) is intended to denote quantified expressions like more and enough.
17 OF as a phrasal affix in the English determiner system
in this example) with Case. The details of the required derivation have rarely been made
precise.14 This imprecision is remedied in an interesting way in Kayne (2002). There,
Kayne gives a careful description of the insertion of of in (36) and links it to an explanation for why the of in derived nominals like (36) can show stranding effects like (37),
superficially in violation of subjacency requirements.
John was admiring a picture of Mary.
(37) Who was John admiring a picture of?
John was [V P admiring [DP a [N P picture of Mary ]]]]]