Technical Report An introduction to tag sequence Ted Briscoe

Technical Report An introduction to tag sequence Ted Briscoe
Technical Report
ISSN 1476-2986
Number 662
Computer Laboratory
An introduction to tag sequence
grammars and the RASP system parser
Ted Briscoe
March 2006
15 JJ Thomson Avenue
Cambridge CB3 0FD
United Kingdom
phone +44 1223 763500
c 2006 Ted Briscoe
Technical reports published by the University of Cambridge
Computer Laboratory are freely available via the Internet:
ISSN 1476-2986
An introduction to tag sequence grammars and the
RASP system parser
Ted Briscoe
Computer Laboratory
University of Cambridge
[email protected]
This report describes the tag sequence grammars released as part of the Robust
Accurate Statistical Parsing (RASP) system. It is intended to help users of RASP
understand the linguistic and engineering rationale behind the grammars and prepare them to customize the system for their application. It also contains a fairly
exhaustive list of references to extant work utilizing the RASP parser.
The first tag sequence grammar (TSG) was written in 1991 in an afternoon as part of an
experiment in grammar induction (Briscoe & Waegner, 1992). Since then, it has undergone many extensions and modifications. The basic idea behind the TSGs was that it is
hard to develop a wide-coverage and robust parsing system which relies on rich lexical
valency or subcategorization information, because it is hard to acquire such wide-coverage
lexicons. Accordingly in 1995 we used a TSG to parse text in order to try to automatically
acquire such lexical entries for the Alvey Natural Language Toolkit full grammar (Briscoe
and Carroll, 1997). In the following decade, especially since the first public release of
the robust accurate statistical parsing (RASP system in 2002 (Briscoe & Carroll, 2002)1 ,
the use of TSG parsed text has broadened to include experiments on topic classification
(Bennett & Carbonell, 2005; Wang, 2005), sentiment classification (Khan, 2006), question
answering (Leidner et al, 2003; Watson et al, 2003), text anonymization (Medlock, 2006),
summarization (de Souza, 2005), information extraction (Callmeier et al, 2004; Grover
et al, 2005; Weeds et al, 2004a; Vlachos et al, 2006), anaphora resolution (Preiss, 2003;
Preiss & Briscoe, 2003; Preiss et al, 2004), word sense disambiguation (Clark & Weir,
2002; McCarthy et al, 2004; Preiss, 2004; Pustejovsky et al, 2004), cue phrase detection
(Abdalla, 2005), identifying rhetorical relations (Sporleder and Lascarides, 2005), computing entailments (Andreevskaia et al, 2005), resolving metonymy (Nissim & Markert,
2003), extracting robust miminal recursion semantics (Copestake, 2003), detecting ellipsis
This paper, describing the RASP system and announcing its public release, has been cited 74 times
(Google Scholar 18/10/05).
(Nielsen, 2004), detecting telic roles (Yamada & Baldwin, 2004), detecting verb alternations (Tsang & Stevenson, 2004), detecting the countability of English nouns (Baldwin &
Bond, 2003), detecting compositionality in phrasal verbs (Bannard, 2005; McCarthy et al,
2003), crosslingual PP attachment (Agirre et al, 2004), Japanese-English machine translation (Tanaka & Baldwin, 2003), measuring lexical and phrasal distributional similarity
(Curran, 2003; Weeds et al, 2004b, 2005), measuring syntactic development in child language (Sagae et al, 2005; Buttery & Korhonen, 2005), as well as further work on valency
acquisition (Baldwin, 2005; Korhonen, 2002, Korhonen et al, 2003; Yallop et al, 2005).
In this report, I describe and motivate the analyses adopted in these grammars and
provide details of the various output representations available. The aim is to provide
enough information that this report can act as a self-contained introduction to the current grammar for users of the RASP system (Briscoe and Carroll, 2002)2 . For more
detailed information, it provides references for further reading. The Grammar Development Environment (GDE) provides facilities for viewing all the grammar constructs as
well as interactively parsing examples to view the derivations produced (see Carroll et al
(1991); $RASP/manuals/gde). I also discuss performance and customization issues for
new applications and datasets.
The Grammar Formalism
The grammar formalism is a subset of that used in the Alvey Natural Language Toolkit
(ANLT) formalism (Briscoe et al, 1987) and TSGs have been developed in the ANLT
Grammar Development Environment (Carroll et al, 1991).
The ANLT formalism defines a metagrammar, based on the notation of Generalized
Phrase Structure Grammar (GPSG, Gazdar et al, 1985) which compiles into an object
grammar which can be thought of as a syntactic variant of a Definite Clause Grammar,
that is, a phrase structure grammar in which categories consist of sets of attributes with
values which are unified by Prolog-style term unification. The object grammar is compiled from the metagrammar by ‘expanding out’ PS rules with additional features (see
Carroll et al, 1991 for details). In the TSGs, I make use of feature declarations, feature
defaults and propagation rules, category delcarations and phrase structure rules (i.e. no
use of GPSG Metarules or ID/LP rules). Furthermore, all features take a finite-set of
atomic, ground values, so the feature system is an abbreviatory device over the class of
context-free (CF) grammars (though we do not compile out to CFGs). Below, I introduce
and motivate the formalism informally – see Briscoe et al (1987) or Carroll et al (1991)
for more precise and formal definitions.
Feature Declarations
Feature declarations take the following form:
They define the set of allowable attributes together with their legal values. There are
41 such declarations in the current grammar and the attribute with the most values is
VSUBCAT with 31. The attribute names (and values) are for the most part based closely
on those used in GPSG and also the ANLT full grammar (Grover et al, 1993) so these
works can be consulted for more detailed descriptions of the use of and motivation for
these features.
Category Declarations
Category declarations take the following form:
CATEGORY CVw2 : [~INV, N -, V +, BAR 0] => {VSUBCAT, PRT}.
They define the allowable attributes of a category or the legal extension of a set of
attribute-value pairs. The antecedent can be a category pattern which specifies absence
of an attribute (such as ĨNV above) or a disjunction of values of an attribute. Category
declaration antecedent pattterns are matched with underspecified categories in PS rules
during construction of the object grammar. Together with feature declarations, category
declarations constitute the typing system which defines the possible forms of grammatical
categories in the grammar.
Phrase Structure Rules
Phrase structure (PS) rules consist of a rule name, rule and optional semantic specification:
PSRULE T/txt-sc1 : Tp --> V2[FIN +, CONJ @c, IMP @i] (Ep).
PSRULE S/pp-np_s : V2[WH @w, INV @i] --> P2[WH -,
PSUBCAT NP] ( +pco ) H2[FIN +, WH @w, INV @i]
: 3 : (ncmod _ 3 1).
The first rule above, called ‘T/txt-sc1’, is one of two start rules in the grammar whose
mother is the root category aliased ‘Tp’, and defined simply as (TOP +). It states that
a matrix clause can consist of a finite sentence, optionally followed by end-of-sentence
punctuation (Ep, e.g. ?). The CONJ and IMP attributes appear with variable values to
ensure that imperatives and sentences which begin with a conjunct can occur as matrix
clauses. In effect, they prevent the application of default feature value declarations in this
context (see below).
The second rule, ‘S/pp-np s’, states that a matrix clause can consist of a preposed prepositional phrase (PP) with noun phrase (NP) complement, optionally followed by a comma,
and then by a finite clause (the head daughter in this rule). The values of the (non-head)
features WH and INV are bound between mother and head daughter in the rule as their
propagation is not covered by more general feature propagation rules (see below). The
category of the head daughter is abbreviated H2 (which expands to H(ead) +, BAR 2)
to ensure that the HFC V propagation rule applies. The attribute H and its values are a
metagrammatical convenience and are removed from the object grammar after the metagrammar compilation process (Carroll et al, 1991). This rule also contains a semantic
specification which states that the third daughter (3) is the semantic value of the rule
and that the rule licenses inference of a grammatical relation of non-clausal modification
(ncmod) between the head daughter and the first daughter (see § 7 for more details).
Semantic specifications as well as PS rule patterns use the convention that the mother
category is 0 and the daughter categories can be referenced numerically from left to right.
There are 678 PS rules in the current grammar of which 157 are marked as rare or
peripheral as opposed to core rules of English grammar.
PS Rule Names
PS rule names in TSGs are intended to be mnemonic to aid grammar debugging. The
general convention is that they consist of an upper case mother category followed by a
slash and a sequence of lower case daughter categories delimited by underscores. I mostly
use more traditional labels for categories in rule names and X-bar ones as abbreviations
in the grammar proper. The daughter categories can also be given mnemomic ‘featural’
information separated by hyphens. So in ‘S/pp-np s’ above, the rule has an S mother
and PP and S daughters, while the PP is restricted to have a NP complement. The rule
names are not intended to be an output representation, although the system can output
trees with nodes decorated with rule names for debugging. Many changes to the grammar
will result in rule name additions or modifications, therefore applications based on this
output representation will tend to be unstable.
A final feature of rule names is that some end in ‘-r’ to designate their peripheral or rare
status. These rules are included to increase coverage of genre specific or otherwise marked
constructions. The ‘-r’ designation can be exploited when creating the probabilistic parser
from the grammar and training data.
Feature Propagation Rules
Feature propagation rules propagate designated attribute values between mother and
daughter categories in PS rules.
PROPRULE AGR1 : V2 --> [N +, V -], [H +, BAR 1]. PLU(1) = PLU(2).
PROPRULE WH1 : N1 --> [H +], U. WH(0) = WH(1).
The first rule above states that in any PS rule with V2 (S) mother, with a nominal first
daughter and with a V1 (VP) head daughter, the attribute PLU must agree in value.
The effect of this rule is to bind variable values on PLU between the two daughters or to
propagate a specific value if one is specified in the PS rule. The second rule propagates
the value of the attribute WH between head daughters and N1 mothers in all N1 rules
with a head daughter. PS rule patterns in propagation rules are matched with PS rules
during object grammar compilation.
Feature Default Rules
Default rules assign default values to attributes in PS rules also in terms of a pattern
which matches a class of such rules.
DEFRULE CONJ1 : [] --> [N, V, ~CONJ], U. CONJ(1) = -.
DEFRULE AUX : V1 --> [H0], U. AUX(1) = -.
For instance the first rule above states that any daughter specified for the attributes N
and V but unspecified for CONJ in any rule should be given a default value of ‘–’. It is to
block application of this default rule that PS rule ‘T/txt-sc1’ above specifies CONJ with
a variable value. The second default rule states that in any VP (V1) rule with a head
daughter, that daughter is by default not an auxiliary verb. Default rules are applied after
feature propagation rules during object grammar compilation – see Carroll et al (1991)
for further details.
Features Sets and Aliases
Attributes can be organized into sets for the purposes of feature propagation in PS rules,
and attribute-value pairs can be abbreviated by defining aliases to create a mnemonic
CF ‘backbone’ in PS rules. For instance, the set declaration below, allows a succinct
statement of verbal head feature propagation underneath,
PROPRULE HFC_V : [V +, N -] --> [H +], U. F(0) = F(1), F in VHEAD.
enforcing binding of variables or identity of values on this set of attributes between mother
and head daughter in all PS rules with S or VP mothers. The alias declaration below
defines V2 (S) which can then be used as a mnemonic abbeviation in the rule underneath.
PSRULE S/np_vp : V2[WH @w, INV -] --> N2[WH @w] H1[FIN +].
Finally, TSGs contain WORD declarations which function as the lexicon by mapping
preterminal to terminal categories of the grammar. In TSGs these preterminals are the
PoS and punctuation tagset.
The PoS and Punctuation Tagset
TSGs exploit the comparative success of statistical part-of-speech (PoS) tagging (Church,
1988). The tagset utilized by the TSGs is based on the CLAWS tagset, as used in
the Susanne Corpus (Sampson, 1995). The full definition of the tagset is available in
these references, and Jurafsky and Martin (2000:Appendix C), the UCREL website at
the University of Lancaster, and the RASP distribution files ($RASP/manuals/).
The full CLAWS tagset contains over 170 PoS and punctuation tags. However, the current
grammar only utilizes 150 of these, of which around 50 are associated with an identical or
subsuming lexical category in the current grammar. Since the development of the Penn
Treebank (Marcus et al, 1993), it has been standard to use a tagset of around 50 labels
following Church (1988), who argued that this tagset size was optimal with respect to
local contextual resolution of ambiguity. In fact, there is no conclusive evidence that
labels from the Penn Treebank tagset can be assigned any more accurately than those
from CLAWS-style tagsets, and the CLAWS tagset contains useful additional information
which can be exploited by a TSG3 The tagger distributed with RASP can either be used
in forced choice mode to select for each word the tag with highest posterior probability
in the HMM or to return all tags with a posterior probabilility.
The tagger system contains a lexicon which has been developed to integrate with TSGs.
The original training material for the tagger (LOB, SEC, Susanne) contained inconsistencies which typically resulted in additional but infrequent associations of tags with
word-forms and omission of some additional tags which represent semi-productive lexical
alternations. For instance, many closed class items such as more or most were assigned a
wide variety of closed and sometimes open class tags (often as a consequence of confusion
or error on the part of the original annotator or variation of practice across corpora). A
systematic effort has been made to keep tag ambiguity in general, and for closed class
items in particular, to a minimum, and to relocate some of the (real) ambiguity in the
TSGs. Often low or even mid frequency open-class word forms occur in corpora only
in their most frequent realization (e.g. zip as a noun, but not verb) or morphologicallyrelated words are not present with a predictable tag (e.g. blurred as verb implies blur can
be a verb). The majority of these omissions have been semi-automatically corrected. The
tagger lexicon currently contains just over 50,000 word forms. The tagger utilizes a probabilistic unknown word model which has generally proved more accurate than attempting
to include rarer words in the lexicon itself.
The word declarations in the grammar associate tags with lexical categories. Most closedclass words (except prepositions) have a single usually fully-specified such category:
WORD AT : DT[PLU @x, POSS -, WH -].
; the, no
WORD DB : N0[NTYPE PART, PLU -, POSS -, WH -, CONJ -]. ; all, half
However, some may have several and may involve underspecification:
Elworthy (1993, 1994) demonstrates essentially identical tagging accuracies using CLAWS tags with
a 1st-order HMM tagger, and also argues that tagset size reduction can lead to worse performance since
it reduces the informativeness of the surrounding context.
WORD DD : DT[PLU @x, WH -, POSS -],
; any,enough,some,lot,rest
In the first case, as PLU is a binary-valued attribute, effectively 4 lexical categories are
assigned to DD. CSA is unique to as but underspecifies 12 lexical categories, as PSUBCAT
can take that number of values. The most ambiguous tags with respect to the current
grammar are those for main verbs as they are all underspecified for VSUBCAT, which
can take one of 31 values specifying allowable complementation via GPSG-style indexing
with 218 PS rules expanding VP (V1) in the current grammar.
The Text Grammar
In addition to a grammar of the syntactic relations which hold inside a sentence, a practical
system able to parse free text requires a text grammar which defines relations which
hold between elements of a text sentence, defined as whatever occurs between an initial
capitalized word and a full stop (i.e. the text unit returned by the RASP tokenizer for
tagging and parsing). For instance, in Kim fell – Sandy had pushed him – but luckily he
was unhurt, there are two sentences with syntactic sentences: Kim fell but luckily was
unhurt and Sandy had pushed him. These two sentences stand in a discourse relation of
elaboration which is signalled to some extent by the use of dashes.
The current grammar’s treatment of punctuation and text grammar draws heavily on
Nunberg’s (1990) approach, with some modifications to ensure computational tractability
and compatibility with the ANLT formalism. There are 53 PS rules with names beginning
T, Taph (text adjunct phrase), Tacl (text adjunct clause), etc., all occurring in the first
set of PS rules in the grammar file, which define how text units combine mediated by
punctuation. For instance, the first PS rule below states that a text sentence can consist
of a sentence followed by a non-clausal text adjunct with sentence-final punctuation.
PSRULE T/s_leta-cl : V2[FIN +, +ta] --> H2[-ta] Ta[+cl, +end]
: 1 : (ta colon 1 2).
PSRULE Taph/dash- : Ta[+da, +end, TXT PH] --> +pda Tph[-da, -do, -sc]
While the second rule states that a sentence-final text adjunct can consist of a dash
followed by a text phrase (in turn defined as any phrasal projection of the main grammar
not itself containing dashes, semicolons or full stops). These two rules are used to parse
a text sentence like Kim fell – Sandy had pushed him.
To a large extent the text grammar and main grammar are modular. However, there
are some places where commas are introduced directly by PS rules of the main grammar
in a construction-specific manner (see for example § 5.8 below). The text grammar is
motivated and an earlier version described in considerably more detail in Briscoe (1994),
while Briscoe and Carroll (1995) demonstrate that use of the text grammar with the
main grammar reduces average ambiguity and extends coverage of the overall system.
The treatment of quotation is dealt with in section 5.1.
The Main Grammar
The main grammar adopts the X-bar framework of Jackendoff as reinterpreted in GPSG.
Phrases and clauses are treated as projections of (lexical) head words and a two-level
(or bar) scheme is assumed so that the maximal projection of a lexical head has two
intervening constituent types (X0, X1, X2). The four major lexical categories (N, V, A,
P) all conform to this scheme (e.g. N0, N1, N2) and V1 (VP) is analyzed as the head of
V2 (S). Other minor categories are mostly treated as specifiers or left-attachments to X2.
Complements are right sisters of X0, and modifiers are right or left recursive sisters of
X1. A more thorough introduction to the X-bar schema used in the TSGs can be found
in Grover et al (1993). Note that the distinction between major and minor categories is
distinct from that between open and closed class lexical items, as prepositions are a closed
class but major lexical category.
The majority of PS rules in the current grammar respect this X-bar schema, using headedness to define properties of the mother via feature propagation and distinctions between
level and directionality of attachment to distinguish arguments from adjuncts and specifiers, as originally proposed by Jackendoff. However, there are marked rules in the current
grammar that are unheaded or otherwise override the schema to deal with peripheral or
marked constructions. For instance, the following PS rule allows a degree premodifier of
scalar adjectives, such as very or so to modify a progressive verb and form an AP, as in
a very damaging allegation.
PSRULE AP/dg_ing-r : A2[MOD -] --> A0[ATYPE DG] V0[VFORM ING, FIN -]
: 2 : (ncmod _ 2 1).
Many of these rules exist to compensate for omissions in the tagger lexicon and are therefore marked with ‘-r’ so that an alternative derivation treating damaging as an adjective
will be preferred where possible. However, in other cases such as coordination (see § 5.8)
the inherent inadequacies of the X-bar framework when combined with strict unification
of categories are circumvented by using unheaded rules.
Quotation is treated somewhat differently from other punctuation because it is inherently
not a (text) sentence-level phenomenon. The tagger normalizes different forms of quote
mark to a single tag. The parser preprocessor removes unbalanced quote marks usually
where the other mark crosses a sentence boundary as in “They’re here! Let’s surprise
them,” Kim said.. This is straightforward if there is a single quote mark in each sentence,
but more difficult if there are three (or more). Therefore, intelligent preprocessing of
the text at the document or paragraph level (before sentence boundary detection and
tokenization) may be worthwhile if extra-sentential quotation information is to be retained
and processed as accurately as possible.
Balanced quotes are handled by 21 PS rules at the end of the text grammar section of the
grammar file. Some of these rules simply allow (major) categories and their projections
to appear with quote marks around them and also allow nesting of quotation, as in: “Kim
‘Slasher’ Smith was arrested.”. Conventions of quote mark interchanging are tokenized
out by the tagger and not exploited by the parser, nor is the difference between begin/end
quote marks for the same reason. Most quoted constituents optionally allow commas or
end of sentence punctuation to occur inside the quote marks to handle quote transposition, as in: Is he a “reasonable man?”. Several of the rules combine quotation specific
constructions with a requirement for balanced quote marks, as in: “The situation,” Kim
opined, “is difficult.”. However in many cases these rules have non-quoted variants in the
main grammar to handle cases where the quotation is extra-sentential and the marks will
have been removed.
There are 89 PS rules expanding V2 (S) in various ways and named ‘S/...’, covering
(non-)finite clauses, small clauses, relative clauses, pre- and post-posing and modifying
constructions, subject-auxiliary inversion, vocatives, tag questions, clefts, imperatives,
locative inversion, sentential subjects, direct and indirect quotation, and so forth. Many
of these rules are preceded in the grammar file by short examples illustrating their intended
Preposing and postposing PS rules, ‘S/xp s’ or ‘S/s xp’ are separated by type of PP etc to
control ambiguity and some are marked ‘-r’ to avoid spurious ambiguity where there is an
alternative more likely derivation which cannot be distinguished (using only information
in tag sequences). Postposing rules require an obligatory comma, ‘S/s pco s’. The third
group of rules handles vocatives, reflexives, tag questions, and similar phenomena, once
again often requiring delimiting commas in order to avoid overlap with other derivations.
The fourth set of rules handle NP+VP combinations and agreement in finite and small
clauses followed by some rules which handle clauses with different types of clausal or
VP subjects, locative inversion and cleft constructions that cannot be handled by other
more general rules. There is one rule of subject-auxiliary inversion ‘S/sai’ for main verbs,
which treats the auxiliary as first daughter and (non-finite) clause as third daughter (with
optional negative operator between) and special case rules for be and ought.
There is one rule which introduces sentential complements with complementizers, one
for imperatives, and seven rules which deal with unbounded dependency questions (i.e.
non-subject wh-questions). The final set of rules deal with clausal coordination (see § 5.8
Verb and adjacent particle combinations are constructed with a ‘morphological’ rule which
combines a V0 and particle to form a V0 with PRT + (to avoid recursion). Otherwise,
verbal complement frames are defined by 218 ‘V1/v...’ rules which partition NP complements by noun type (NTYPE) defined in terms of CLAWS tagset nominal subtypes (see
tags NX(X) in the WORD declaration of the grammar file). Breaking down the rules
this way allows greater precision in the application of some rules, especially in contexts
like ditransitive ‘double NP’ constructions where it is often difficult to distinguish NP
boundaries (e.g. (elect) Tony Blair the prime minister vs. (give) him it). These rules also
treat heavy-shifted variants, particle movement, and diathesis alternations as separate
complementation patterns.
The second group of rules, ‘V1/vp xp’ or ‘V1/xp vp’ handles pre- and post-modification
of V1 and the third set coordination of VP.
The next set of VP rules deals with auxiliaries. There are 35 rules, V1/be...’ dealing with
copular and auxiliary be along with ‘VP-internal’ modification of predicative NP objects
of be. There are 11 rules, ‘V1/do..’ which handle auxiliary and main verb do, 22 for have,
‘V1/have...’, 2 for infinitive to, and 2 for modals. There are also a number of special rules
for non-constituent coordination with be (e.g. is a conservative and proud of it) and ‘-r’
rules allowing gaps and ellipsis with each type of auxiliary (e.g. what do you do e?). The
latter are needed as auxiliaries are effectively subcategorized in the grammar, unlike main
Noun Phrases
The first set of rules deals with numbers and some tokens containing numbers, such as
$10M, which receive a variety of PoS tags which are associated with the feature NTYPE
NUM in the grammar. Thse rules allow complex numbers to be built out of tokenized
components (e.g. 10 . 4) (though the RASP tokenizer doesn’t currently segment these),
ordinals, ranges, dates, times, and so forth. Some variants of such phrases involve combination with a NTYPE TEMP token as head (e.g. 4th July, 1989) or are coerced to this
type if recognizable as a date, and so forth. Numbers can also be coerced to pronouns
(e.g. the twenty who came).
The next set of rules license NP (N2) phrases without specifiers (determiners, etc) for
a subset of NTYPEs, such as plural common nouns, names, etc, and NPs with initial
determiners, partitives, and so forth (e.g. all/half (of ) the group). Some of the latter
rules are ‘-r’ because many partitives have very flexible syntactic realization and thus a
number of PoS tags, so these rules should only yield highly-ranked derivations if the tagger
is used in forced choice mode. There follow three rules handling possessives including one
‘-r’ rule which allows ellipsis (e.g. Kim went to the butcher’s).
Pronouns have separate tags,PPX(X), and are NTYPE PRO/REFL. The next set of 15
rules deal with post- or pre-modification of pronouns (e.g. (the) someone in the barn)
– these are mostly ‘-r’ to force a preference for higher attachment of the postmodifiers
where possible (e.g. I met someone in the barn). There follow 10 rules of more general NP
postmodification, many of these are also quite marked as most postmodification and/or
complementation is specified at the N1 level.
To cut down on ambiguity the rules dealing with nominal phrase combinations are broken
down by NTYPE, for example licensing combinations of names and NPs (e.g. girls’ band
Banamarama), comma-less appositives the singer James Browne etc. There are 18 such
often very specific rules and this is an area where further additions may be required to
deal with specific sublanguages.
The next group of rules are unheaded as they mostly deal with morphological conversion
(e.g. the pick up) or (elliptical) cases with adjectival heads (e.g. the poor in the ghetto).
However, some rules capture additional marked specifiers or premodifiers (e.g. about 10
Nominal phrases (N1) constructed from single nouns are subdivided by NTYPE to control
ambiguity. (Notably pronouns have no N1 phrasal level and therefore are not accessible
to the unmarked general rules of complementation and pre- or post-modification.) The
second set of rules deals with AP premodification, possessives without specifiers, and
unheaded cases that also need a N1 level analysis (e.g. the tedious pick up), these are
all ‘-r’ to avoid competition with the NP level analogues and to disprefer the additional
misanalyses that can result from these (more general) rules.
There are 6 rules which deal with N0+N1 combinations which can’t be subsumed by the
NP-level analogue rules discussed above. These are again subdivided by NTYPE to avoid
spurious ambiguities as much as possible. There follow 22 rules which deal with N0+N0
compounding, in the broadest sense, allowing many combinations of different NTYPES
but excluding many others. Once again these rules may need supplementing for new
The next set of 6 rules deal with noun complementation or alternatively postmodification where a left-recursive N1 level rule would overgenerate. These rules handle of PP
complements and sentential complements on various NTYPEs. The next 20 rules handle postmodification or occasional complementation of with verb phrases, and relative
clauses. Finally, there are 13 rules to handle nominal coordination at N2, N1 and N0
There are 31 PS rules dealing with prepositional complementation for various types of
preposition. Prepositions are distinguished in so far as the CLAWS tagset allows by the
value of PFORM. For example, with is IW and (PFORM with), as is CSA and (PFORM
as) etc. than and of are also (ADJ +) to prevent their attachment as adverbial modifiers.
The bulk of prepositions though (128 in the current tagger lexicon), are tagged as II in
the CLAWS scheme and are treated as (PFORM prep) in the grammar, so many PS rules
underspecify PFORM value. The 53 subordinating conjunctions tagged CS and the 17
temporal conjunctions tagged ICS, such as because, although, before etc, are treated as
prepositions with PFORM value ‘preps’.
Prepositional complements are distinguished by the attribute PSUBCAT which takes 12
values. However, many P1 rules are further broken down to distinguish, for example,
NTYPEs within (PSUBCAT NP) complements as reflected in the mnemonic rule names,
e.g. ‘P1/p np pl’, P1/p np-org’ etc. These provide increased precision in parsing, as
not all NP NTYPEs can occur as prepositional complements and they also allow induction of prepositional complementation from more reliable rules (e.g. ‘P1/p np-pro’
with him/her/it), and higher precision marked rules, such as ‘P1/p np-poss-ellip-r’ (at
the butcher’s). Some PSUBCAT values, such as ‘sing’ are often licensed as much by the
verb as the preposition (wonder/*wander about Kim leaving). In a deeper grammar, these
would be treated as part of the verbal complementation. TSGs factor the information
between VSUBCAT and PSUBCAT in order to reduce overall ambiguity and to allow
flexible combination of verbal and prepositional frames. However, the observed interactions of VSUBCAT and PSUBCAT, PFORM and specific verbs and prepositions can and
should be exploited in parse selection and when inducing valency lexicons.
There are separate rules for as and than due to the large range of elliptical complementation in comparative and equative constructions and some marked rules for moved
‘particles’, often tagged as prepositions, which are also used in preposition stranding constructions. These treat such prepositions / particles as intransitive – therefore recovery
of any complementation can only be done heuristically in the grammatical relations (GR)
output or postprocessing phase (see section 7).
PP (P2) rules allow for premodification (partly because ..., slightly after noon) and MOD
prevents these being treated as verbal arguments. Finally there are 7 rules of prepositional
The AP rules generalize over adverbs and adjectives except where ADV +/- is specified in
rules. The first 14 A1 rules handle premodification including some rare (idiomatic) cases
(very slow(ly), trouble free), premodified marked unheaded APs (completely destroyed
(city)), and numerical (comparative) premodifiers (one-third more difficult).
There are 15 rules of adjectival complementation mostly dealing with PPs distinguished
by PFORM including wh-PPs, but also clausal and predicative complements. These are
followed by 2 rules for PP modifiers which are (MOD +) to prevent overlap with argument
PPs. There are 7 rules for specific constructions involving adverbial premodification of
adverbs (completely accurately), postmodification by A0, etc. Finally there are 16 rules
handling A coordination including some for unlike categories in predicative complementation.
The attribute MOD is set to + when e.g. a N1 is postmodified and is used to constrain
order of attachment of pre- and post-modifiers and thus reduce spurious ambiguity. It is
also useful in some contexts (e.g. is this afternoon not available) to block attachment of
‘long’ postmodified NPs as interleaved adverbials, and so forth. The same feature is used
in PP, AP and VP phrases in a similar manner.
All the major categories have distinct coordination rules which follow an underlying
binary- and right-branching (X → X X) schema where the first conjunct is (CONJ −), the
second (CONJ +) and the mother unspecified for this attribute. Elsewhere in the grammar major categories default to (CONJ −). ‘XP/cj-beg xp’ rules introduce coordinate
markers at the beginning of coordinate constructions ((both (... and ...))). ‘XP/cj-end’
introduces conjunctions as sisters of (CONJ −) conjuncts (... (and (...))). There are 63
such rules in the current grammar as the rules must be specified for each major category
separately and because additional rules relaxing conjunct agreement are specified in a
category-specific way.
The main schema rules, ‘XP/xp xp-coord’ are not headed and feature propagation is
explicitly specified on a rule-by-rule basis to alleviate some of the problems of unificationbased approaches to this construction. In particular, for many categories more relaxed ‘-r’
versions of the rules are given that do not force the same degree of ‘agreement’ between
conjuncts (e.g. he and I (were), have you an alibi or did you murder Smith?, etc).
Unbounded Dependencies
All wh-NPs involving subject dependencies in relative clauses, wh-questions, and so forth
are treated as normal NPs. However, preposed non-subject wh-NPs, wh-PPs and some
(topicalized) normal NPs are handled by 5 PS rules which combine a preposed whconstituent with a clause (which may have undergone subject-auxiliary inversion). The
location of the gap in the clause is not represented as the grammar has no access to
(verbal) valency. However, the GR output does predict the role of the wh-constituent
heuristically and by underspecification – e.g. there is usually only one verb but we cannot
know whether it has one or more NP objects
The auxiliaries are treated somewhat differently as they are effectively subcategorized
in the tagset, so the grammar contains explicit gap rules ‘V1/be gap-r’, etc to allow for
missing preposed arguments. However, there is still no analogue to the GPSG / HPSG
gap features, which represent the dependency between preposed filler and gap.
Development Corpus
The development corpus in $RASP/prob/corpus/ indicates the current coverage of the
grammar and contains tagged files that can be parsed using the GDE fparse utility (*.tag)
or input to the RASP system (*.txt). The grs/ directory contains the correct GR output
for the correct derivations (*.trees1). Studying the files of derivations and GR output is
one good way to understand how the grammar works and what it is intended to produce.
Grammatical Relations
The current grammar outputs grammatical relations (GRs) based on a rule-to-rule encoding of GRs with PS rules. Historically, this developed from a procedural mapping
from TSG derivations to a GR scheme close to that recommended for parser evaluation
by EAGLES. However, this has been modified to make it more appropriate to a shallow
semantic representation (Briscoe et al, 2002) which might form the input to further semantic processing such as robust minimal recursion semantics construction or to ‘event’
extraction. The scheme is similar to the F-structure of LFG, but is used by RASP purely
as an output representation with no effect on the underlying space of possible derivations.
The main changes over the EAGLES-derived scheme (which is described on the web at are that most
GRs are now factored into simple binary lexical relations, and coordination is not distributed across conjuncts but rather the coordinator is treated as (semantic) head – this
avoids problems of interaction with scope in examples like: Every man smiled or laughed
i.e. (conj or smiled) (conj or laughed) (ncsubj or man ) – which does not entail that
every man smiled and laughed.
The new scheme also adds several relations, notably text adjunct (ta) which encodes
information conveyed by punctuation. The aim is that the GR scheme is a factored
representation of a directed graph which is ‘almost’ acyclic. Most cycles are between
adjacent nodes and involve disjoint relations in which head and dependent lexical items
are reversed. For example in the hidden chest we output a non-clausal modifier (ncmod)
relation between head (chest) and dependent (hidden) and a non-clausal surface subject
/ underlying object relation between head (hidden) and dependent (chest) i.e. (ncsubj
hidden chest obj). However, we also create cycles to capture unbounded dependencies
between nominal heads and (relative) clause postmodifiers, for example in ideas (that)
linguists (want to) promote, there is an underspecified obj (dobj/obj2/iobj) relation (obj
promote ideas) as well as (cmod (that) ideas promote). However, given that this is an
unbounded dependency it can involve non-adjacent nodes: (cmod (that) ideas want),
(xcomp want promote), (obj promote ideas).
All GRs are of the following form:
(GR-type optional-subtype head dependent optional-initial-GR)
In practice, only non-clausal subjects take the optional initial GR field and only some
modifier and complement relations take the optional subtype field. We also use a single
feature (passive head) which is required to facilitate recovery of the initial-GR field for nonclausal subjects. There are 17 fully-specified GR types, however, these are organized into
an inheritance hierarchy shown in Figure 1 which facilitates underspecification of relations
e.g. with some unbounded dependency constructions (see section 5.9) and also cross-parser
evaluation. In addition to the main GR types, there are subtypes of some GRs which
indicate the presence of some grammatical markers such as infinitive to, complementizer
that and so forth. The intention is to output a representation which conveys as much as
possible about the likely underlying predicate-argument structure of the input sentence.
To this end, certain aspects of the output are default, especially default subtype/initialGR values ( ) and relations in unbounded dependency constructions (arg mod, obj), and
others are frequency-based heuristics that may need overriding when more specific lexical
information is available, as in Kim promised Lee to go (ncsubj go Lee ) → (ncsubj go
Kim ) – see section 7.2 below.
Further information can be obtained by inspecting the grammar file and comments associated with rules. In the GDE, the GR output can be viewed from the parser command
line using ‘v(iew) sem(antics)’ as the mechanism used to construct GRs is a simple modification of that used to produce the compositional semantics of the ANLT full grammar
Xhhh hh
arg mod
det aux conj
h hh
h hhh
hh arg
!! Q
! subj dobj QQ
ncmod xmod cmod pmod !!! ((((
pcomp P
ncsubj xsubj csubj !!H
obj2 iobj
Figure 1: The GR hierarchy
(Grover et al, 1993). The following contains a brief description of each GR type, its
subtypes and intended use.
conj encodes relations between a coordinator and the head of a conjunct. There will be
as many such binary relations as there are conjuncts of a specific coordinator. It has no
additional fields.
Kim likes oranges, apples, and satsumas or clementines
(ncsubj likes Kim _) (dobj likes and)
(conj and oranges) (conj and apples) (conj and or)
(conj or satsumas) (conj or clementines)
aux encodes relations between main verbs as (semantic) head and auxiliary dependents.
It has no additional fields. There will be as many such binary relations as there are
auxiliaries. If a copular or main verb form of an auxiliary is present then it is the head of
any such aux relation. The head of aux can be ellip(tical) as in Kim will.
Kim has been sleeping
(ncsubj sleeping Kim _) (aux sleeping has) (aux sleeping been)
det encodes a binary relation between articles, quantifiers, partitives and other single
word forms which can begin NPs and the head of the NP. It has no additional fields.
Some men came
(det men Some) (ncsubj came men _)
ncmod encodes binary relations between non-clausal modifiers and their heads. There are
subtypes: default ( ), part(itive), prt(particle), poss(essive), num(ber), ta(text adjunct),
and ij(interjection). The default case covers most pre-/post-modification.
the old man in the barn slept
(ncmod _ man old) (ncmod _ man in) (dobj in barn)
Numbers are identified as special types of modifier where possible. Possessives are treated
as relations between head and dependent nouns:
the butcher’s shop
(ncmod poss shop butcher)
where the head can be ellip(tical). Partitive predeterminers are (ncmod part men all) all
the men). Verbal particles are (ncmod prt look up) (look up the word).
xmod encodes binary unsaturated predicative relations between modifiers (VPs, APs)
and heads. There are subtypes default ( ) and ‘to’, the latter is used when the modifier
is an infinitive VP (though the current grammar doesn’t always recover it):
who to talk to
(xmod to who talk) (iobj talk to) (dobj to who)
cmod encodes binary saturated relations between clausal (S) modifiers and heads. There
are subtypes default ( ) and complementizer ‘that’: although he came, Kim left
(cmod _ left although) (ccomp although came)
pmod encodes binary relations between PP modifiers with PP complements and heads:
he went, off into the darkness
(pmod went off) (pcomp off into) (dobj into darkness)
ncsubj encodes binary relations between non-clausal subjects (NPs, PPs) and their verbal
heads. There are no subtypes but four initial GR values: default/subj ( ), underlying
object (obj), raising subject (rais) and inverted (inv) which is used for locative (PP,
AdvP) inversion and quote inversion (said Kim):
the upset man
(ncsubj upset man obj) (passive upset)
All passives get the passive feature but the overwriting of the default ncsubj subtype is
often left to the GR inference stage (see section 7.2 below. (This is because of limitations
of the rule-to-rule translation mechanism.) This relation is also used for understood
subjects of unsaturated predicative complements and some modifiers – these assignments
are heuristic and based on the assumption that most transitive/intransitive verbs are
object/subject equi, as in Kim wants to go (ncsubj go Kim ). The raising subtype is
thus never inserted by the grammar.
xsubj encodes binary relations between unsaturated predicative subjects (VP, AP) and
verbal heads. This relation only has non-default subtype value inverted (inv) used for
extraposition examples like to go appears difficult:
leaving matters
(xsubj matters leaving _)
csubj is a binary relation between saturated clausal subjects (S/V2) and verbal heads.
The subtype slot is filled by the complementizer if the clause is finite and left empty for
non-finite ‘small clauses’ like her coming matters:
that he came matters
(csubj matters came that) (ncsubj came he)
dobj is a binary relation between verbal or prepositional head and the head of the NP to
its immediate right. Thus, it doesn’t distinguish between themes and goals with dative
alternation verbs:
She gave it to Kim
(dobj gave it) (ncsubj gave She _) (iobj gave to) (dobj to Kim)
obj2 is a binary relation between verbal heads and the head of the second NP in a double
object construction:
She gave Kim toys
(obj2 gave toys) (dobj gave Kim) (ncsubj gave She _)
The NP immediately to the right of a verb in a passive construction is sometimes recognized correctly as obj2 (e.g. when in a reduced relative) but not when in a main clause.
However, the presence of (passive given) in the GR set for Kim was given it can be used
to trigger an overwrite rule – see section 7.2 below.
iobj is a binary relation between a head and the preposition of a PP argument when the
PP complement is a NP:
Kim flew to Paris from Geneva
(ncsubj flew Kim _) (iobj flew to) (iobj flew from)
(dobj to Paris) (dobj from Geneva)
pcomp is a binary relation between a head and the preposition of a PP argument when
the PP complement is itself a PP:
Kim climbed through into the attic
(ncsubj climbed Kim _) (pcomp climbed through)
(pcomp through into) (dobj into attic)
xcomp is a binary relation between a head and an unsaturated VP complement. It has
subtypes: default ( ) and ‘to’, the latter indicating an infinitival complement. However,
the current grammar doesn’t always insert ‘to’ when the complement is infinitival:
Kim thought of leaving
(ncsubj thought Kim _) (xcomp _ thought of) (xcomp _ of leaving)
ccomp is a binary relation between a head and the head of a saturated clausal complement, either finite, subjunctive, headed by a wh-element or a non-finite ‘small clause’. It
has subtypes: default ( ) and ‘that’. The head of the dependent clause is usually the verb
but can be the subject of the ‘small clause’:
Kim asked about him playing rugby
(ncsubj asked Kim _) (ccomp _ asked about) (ccomp _ about him)
(ncsubj playing him _) (dobj playing rugby)
ta is a binary relation between a head and the head of a text adjunct delimited by some
punctuation (see section 4). It has subtypes: ‘quote’, ‘brack’(et), ‘dash’ ‘colon’, ‘comma’,
‘bal’(anced), ‘end’, ‘echo’, ‘tag’ (questions), ‘refl’(exive), ‘voc’(ative). The first five subtypes are self-explanatory. Balanced text adjuncts have matching delimiting punctuation
(e.g. commas or dashes at both boundaries). End text adjuncts usually occur sentencefinally and thus the matching punctuation mark at the right boundary is promoted to a
full stop. The remaining subtypes attempt to infer some of the semantic/discourse import
of a comma from information in the PoS tagset concerning nominal lexical types. They
are sometimes wrong in the current grammar, so could more conservatively all be mapped
to ‘comma’:
He made the discovery: Kim was the abbot; Lee was the host.
(ncsubj made He _)
(dobj made discovery)
(ta colon discovery was)
(ncsubj was Lee _)
(xcomp _ was host)
(det host the)
(ncsubj was Kim _)
(xcomp _ was abbot)
(det abbot the)
(det discovery the)
Underspecification is used by the current parser, especially in the analysis of unbounded
dependencies, so that obj, arg and arg mod may be output and also may be further
specified as described in section 7.2. mod, subj, comp and subj or dobj are provided
primarily for cross parser evaluation purposes.
Mapping from the EAGLES Evaluation Scheme GRs
The first release of RASP supported GR output in a format very similar to that proposed
by the EAGLES parser evaluation working group (see The following rules describe the mapping from the old scheme to the new scheme. The mapping
will not be perfect as there are more subtypes on some GRs in the new scheme (e.g. part,
num, ta on ncmod), new GRs (e.g. ta), and more information (especially control and
unbounded relations) extracted.
Lower case variables, x, y, . . . range over word| lemma-affix-tag tokens which represent
the heads/dependents in GRs and/or over subtype values of specific GRs. Where the
values of these variables need restricting, I use the convention that x=( A) matches any
word/lemma with PoS tag prefix A, x=( *A*) any word/lemma with PoS infix A, while
x=( ) indicates the literal (subtype default) value ‘underscore’. I also use x=(A|B) as a
disjunction operator over tag and token components.
1. ∀ x=( |poss|prt),y,z (ncmod x y z) ⇒ (ncmod x y z)
2. ∀ x=( I),y,z=( N) (ncmod x y z) ⇒ (ncmod
3. ∀ x=( I),y,z=( I) (ncmod x y z) ⇒ (pmod
y x) ∧ (dobj x z)
y x) ∧ (pcomp x z)
4. ∀ x=( |to),y,z (xmod x y z) ⇒ (xmod x y z)
5. ∀ x=( |that),y,z=( V) (cmod x y z) ⇒ (cmod x y z)
6. ∀ x=( I|C),y,z (cmod x y z) ⇒ (cmod
y x) ∧ (ccomp x z)
7. ∀ x=( *Q*),y,z (cmod x y z) ⇒ (cmod
y x) ∧ (arg x z)
8. ∀ x,y,z (ncsubj x y z) ⇒ (ncsubj x y z)
9. ∀ x,y (ncsubj x y obj) ⇒ (passive x)
10. ∀ x,y,z (xsubj x y z) ⇒ (xsubj x y z)
11. ∀ x,y,z (csubj x y z) ⇒ (csubj x y z)
12. ∀ x,y,z (dobj x y z) ⇒ (dobj x y)
13. ∀ x,y (obj2 x y) ⇒ (obj2 x y)
14. ∀ x,y,z( N) (iobj x y z) ⇒ (iobj y x) ∧ (dobj x z)
15. ∀ x,y,z( I) (iobj x y z) ⇒ (pcomp y x) ∧ (pcomp x z)
16. ∀ x=( |to),y,z (xcomp x y z) ⇒ (xcomp x y z)
17. ∀ x=( |that),y,z=( V) (ccomp x y z) ⇒ (ccomp x y z)
18. ∀ x=( I|C),y,z (ccomp x y z) ⇒ (ccomp
19. ∀ x,y,z (arg mod x y z) ⇒ (ncmod
y x) ∧ (ccomp
x z)
y x) ∧ (dobj x z)
20. ∀ x,y,z,. . . (conj x y z . . .) ⇒ (conj x y) ∧ (conj x z) ∧ (conj x . . .) . . .
21. ∀ x,y (detmod
22. ∀ x,y (aux
x y) ⇒ (det x y)
x y) ⇒ (aux x y)
GR Inference
The new GR extraction process relies on purely local information available within a local instantiated tree/PS rule. In some cases, it is possible to D(elete), C(orrect) or
A(dd) to the resulting GRs using rules which look at subsets of the GR output including L(emmas)(see section 7.1 for other notational conventions). The following is not an
exhaustive list of such rules. In particular, there are many more rules similar to those
beginning with 10) below which add or further specify missing GRs in unbounded dependency constructions. Such rules require a valency dictionary and could be used to
probabilistically extend GR sets if the valency dictionary included conditional probabilities of the form P (valency | verb).
1. Passive1: ∀ x,y (passive x) ∧ (ncsubj x y ) ⇒ C:(ncsubj x y obj)
2. Passive2: ∀ x,y (ncsubj x y obj) ∧ ¬ (passive x) ⇒ A:(passive x)
3. Passive3: ∀ x,y,z ((ncsubj x y obj) ∨ (passive x)) ∧ (dobj x z) ⇒ C:(obj2 x z)
4. Not-passive1: ∀ x,y,z (aux x have) ∧ ((ncsubj x y obj) ∨ (passive x)) ⇒ C:(ncsubj
x y ) ∧ D:(passive x)
5. Not-passive2: ∀ x,y,z (aux x have) ∧ ((ncsubj x y obj) ∨ (passive x)) ∧ (obj2 x
z) ⇒ C:(dobj x z) ∧ C:(ncsubj x y ) ∧ D:(passive x)
6. Predicative: ∀ x=( VB),y,z (xcomp
x y) ∧ (ncsubj x z ) ⇒ C:(ncsubj y z )
7. SRaising: ∀ x=(L V),y,z,w,w′ (dobj x y) ∧ (ncsubj x z ) ∧ ¬ (passive x) ∧ ¬
(xcomp w x w′ ) ⇒ C:(xcomp x y) ∧ C:(ncsubj y z ), L=become,seem,. . .
8. ORaising: ∀ x=(L V),y,z,w (dobj x y) ∧ (xcomp w x z) ⇒ D:(dobj x y) ∧ A:(ncsubj
z y ), L=believe,suppose,. . .
9. SControl: ∀ x=(L V),y,z,w,w′ ¬ (dobj x y) ∧ (xcomp w x z) ∧ (ncsubj x w′ ) ⇒
A:(ncsubj z w′ ), L=try,want,. . .
10. OControl: ∀ x=(L V),y,z,w (dobj x y) ∧ (xcomp w x z) ⇒ A:(ncsubj z y ),
L=want,expect,. . .
11. Promise: ∀ x=(L V),y,z,w,w′ (dobj x y) ∧ (xcomp w x z) ∧ (ncsubj x w′ ) ⇒
A:(ncsubj z w′ ), L=promise,. . .
12. Scomp2Rel1 ∀ x,y=( N),z=(L V),w, w′ (ccomp x y z) ∧ (ncsubj z w ) ∧ ¬ (dobj
z w′ ) ⇒ C:(cmod x y z) ∧ A:(dobj z y), L=seek,. . . (oblig. transitive)
13. Scomp2Rel2 ∀ x,y=( N),z,w=(L V),w′ , w′′ (ccomp x y z) ∧ (xcomp w′ z w) ∧ ¬
(dobj w w′′ ) ⇒ C:(cmod x y z) ∧ A:(dobj w y), L=seek,. . . (oblig. transitive)
14. Rel1-UB ∀ x,y=( N),z=(L V),w, w′ (cmod x y z) ∧ (ncsubj z w ) ∧ ¬ (dobj z w′ )
⇒ A:(dobj z y), L=seek,. . . (oblig. transitive)
15. Rel2-UB ∀ x,y=( N),z,w=(L V),w′ , w′′ (cmod x y z) ∧ (xcomp w′ z w) ∧ ¬ (dobj
w w′′ ) ⇒ A:(dobj w y), L=seek,. . . (oblig. transitive)
16. Rel-Inf-UB ∀ x,y=( N),z=(L V) (xmod x y z) ∧ (obj z y) ⇒ C:(dobj z y),
L=seek,. . . (oblig. transitive)
Other Output Representations
In this section, I give an example of RASP’s various outputs – further examples can be
found at For the input
sentence, I can can a can the system produces GRs:
(|ncsubj| |can_VV0| I_PPIS1 _)
(|aux| |can_VV0| |can_VM|)
(|dobj| |can_VV0| |can_NN1|)
(|det| |can_NN1| |a_AT1|)
In order to distinguish multiple occurrences of the same form in a sentence and so word
order can be reconstructed unambiguously from GR output, by default, the word sequence
in each sentence is numbered: 1-n in the output:
(|ncsubj| |can:3_VV0| I:1_PPIS1 _)
(|aux| |can:3_VV0| |can:2_VM|)
(|dobj| |can:3_VV0| |can:5_NN1|)
(|det| |can:5_NN1| |a:4_AT1|)
In addition to GR output descibed in section 7 above, there are other output formats
many of which can be selected from the command line using the RASP shell script
Weighted GRs are GRs with weights computed from the probabilities of the set of
derivations in the top n derivations which yield this GR. This output mode allows the
user to define the trade-off between precision and recall by setting a weight threshold
for further processing. However, the set of GRs output is no longer consistent so further
processing is required to select a consistent and complete set of GRs for the input (see
Carroll & Briscoe, 2002). The definition of consistency and completeness will be very
similar to that for LFG F-structures but is complicated by the fact that a few words like
complementizer that and infinitive marker to appear as GR subtype values as opposed to
as nodes in the graph. A further caveat is that when no complete parses can be found for
an input, weighted GRs are computed from a single sequence of subanalyses (and thus all
appear with weight 1). For instance, the weighted GRs for the example sentence are:
(|ncsubj| |can_VV0| I_ZZ1 _)
(|dobj| |can_VV0| |can_NN1|)
(|aux| |can_VV0| |can_VM|)
(|ncsubj| |can_VV0| I_PPIS1 _)
(|det| |can_NN1| |a_AT1|)
If multiple tags for can were not filtered out, then further weighted GRs would be present.
Grammar Rules as labels of the nodes in a derivation can be output primarily for
debugging purposes. Rule names mostly follow the convention ‘M/d1-f1-f2-fN d2 dN’
where M is the mother category, d1...dN are the daughter categories and f1...fN are salient
features of daughters (see the TSG file comments for more details on naming conventions).
It is a bad idea to base further processing on these names as they invariably change when
the grammar is modified. The labelled bracketing with grammar rule names as nodes for
the example sentence is:
(|S/np_vp| I_PPIS1
(|V1/modal_bse/--| |can_VM|
(|V1/v_np| |can_VV0| (|NP/det_n1| |a_AT1| (|N1/n| |can_NN1|)))))
(|End-punct3/-| ._.))
PTB-style (Penn Treebank) labels on the nodes of derivations can be selected if such
similarity is useful. However, the tree topology will be different (and more informative)
in RASP as it uses the X-bar scheme (see section 5).
Shallow Phrasal labels on the nodes of derivations emulate the Susanne (Sampson,
1995) treebank scheme though again not the tree topology.
Alias labels on the nodes of derivations use the alias declarations from the grammar file
located in $RASP/prob/data/tsgX. For instance,
ALIAS V1 = [V +, N -, BAR 1].
so nodes labelled V1 will be subsumed by this aliased feature set.
Other representations of node labels in derivations are available and can be selected from
within Lisp using the functions linked to the ‘view’ command in the GDE, including the
full featural representation of each node. However, in general we recommend GRs, and
it is possible to specify many tasks that are often taken to require trees in terms of GRs
(e.g. for anaphora resolution see Preiss & Briscoe, 2003).
Performance and Accuracy
Training on about 4K sentences paired with unlabelled brackets from the Susanne treebank, and testing on 560 sentences from the Penn Treebank reannotated with GRs (part of
the Parc DepBank700 reannotated from section 23 of the Wall Street Journal in the Penn
Treebank) the parser achieves a microaveraged F1-score of 74.9%. RASP took 153secs.
to load and parse these sentences on a 1.8GHz Opteron CPU. See $RASP/prob/greval
for more details of the evaluation scheme, data, etc.
Using RASP in tag thresholding mode yields little improvement in accuracy on this test
set but may be worthwhile for datasets with a large number of unknown words (for the PoS
tagger) at some cost to speed. Using the subcategorization option is likely to yield some
small improvements in accuracy at little cost to performance, so is recommended. Using
the inside-outside algorithm variant of weighted GR extraction will increase accuracy at
little cost to efficiency and is, therefore, recommended (see parser options in the RASP
shell scripts).
Improving parse selection accuracy by lexicalizing the current model, building a model
over GRs, and/or using discriminative techniques over the parse forest is a current area
of research (Sloane, 2005; Watson et al, 2005).
Tokenization / Sentence Boundary Detection
Sentences which are not already split and marked using the hat symbol will be split by
the RASP sentence boundary detector (see $RASP/sentence) and then passed to the
tokenizer (see $RASP/tokenise). Idiosyncracies of input can lead to failure at this point
with dramatic effects on parser performance. Unusual sequences of punctuation marks,
such as ‘(A):-’ or ‘-----’, and archaic abbreviations, such as ’twas or ’tis will not be
treated correctly by the current modules.
Therefore it is worth looking carefully at the output of this stage and either preprocesssing
with global substitutions in a batch editor to make such sequences compatible with the
current modules (and tagger lexicon) or modifying the modules directly (e.g. ----- →
--; or ’twas → it was).
GR Output
Modifying the syntactic part of the grammar rules is not possible as the public release does
not yet allow you to retrain the probabilistic parser. However, altering the GR output
from (sets of) rules is possible and may be desirable if, for example, it is preferable to
have a less informative set of GRs which are guaranteed to yield directed acyclic graphs.
The description of the semantic formalism used to generate GRs is given in Carroll et al
(1991) and Grover et al (1993) in more detail. However, to find cycles in the GRs it is
only necessary to find PS rules which specify more than one GR output with identical
daughter variables (usually reversed as head/dependent), for example:
PSRULE N1/n_ppart : N1[POSS -, MOD -] --> H0[NTYPE NORM] V1[VFORM
PPART] : 1 : (xmod _ 1 2) : (ncsubj 2 1 obj) : (passive 2).
PSRULE NP/np-whpro_inf : N2[WH +, POSS -, MOD +] --> H2[WH +, NTYPE PRO]
V1[VFORM INF, FIN -] : 1 : (xmod to 1 2) : (obj 2 1).
Removing the second of these GRs from the output specification is usually most appropriate and will remove cycles, at some cost to informativeness.
Grammar Tuning
Though it is not possible to add rules to the grammar or retrain the grammar, it is
possible to remove rules by customizing the RASP shell script with a list of unwanted rule
names (see $RASP/scripts/README for details). This can be useful because the overall
accuracy of the parser may be improved by deletion of rules that will rarely apply correctly
within a particular text type or genre, but which may nevertheless occur incorrectly in
significant numbers of highly ranked parses. Removing such rules will not cause the parser
to fail outright, as the fragmentary parse mechanism will be invoked in cases where the
only possible parse required the deleted rule.
Many of the -r rules in the grammar are there to cover quite specific constructions which
occur rarely in many types of text. For example, text which does not contain a high
proportion of questions or elliptical constructions with auxiliaries (as in fictional dialogue)
may be parsed more accurately without the four rules with names ending in ‘ gap-r’. By
examining incorrect but highly ranked parses on specific data, it should be possible to
identify candidate rules for removal.
Further Processing
Many people are tempted to use trees labelled with grammar rule names as the output they
will work with (as it appears informative and is the default tree output representation).
This output is very useful for debugging and understanding the grammar. However, it
changes almost every time the grammar is modified. Therefore, it is far better to work
with a relatively stable output scheme. The GRs are the most stable output available
as a command line option in the distribution. Derivations containing category aliases, or
Penn Treebank style output are more stable than those containing rule names but may
still change if modifications to the grammar alter tree topology. These can be selected as
a parameter to the RASP shell script.
Abdalla, R.M. (2005) Syntactic and lexical variants of cue phrases, M.Phil Dissertation,
University of Cambridge Computer Laboratory.
Abney, S. (1996) ‘Part-of-speech tagging and partial parsing’ in Church, K. et al (eds.),
Corpus-based Methods in Language and Speech, Kluwer, Dordrecht.
Eneko Agirre, Aitziber Atutxa, Koldo Gojenola, Kepa Sarasola (2004) ‘Exploring portability of syntactic information from English to Basque’, Proceedings of the LREC’04,
Alina Andreevskaia, Zhuoyan Li and Sabine Bergler (2005) ‘Can Shallow Predicate Argument Structures Determine Entailment?’, Proceedings of the Pascal Challenge Workshop on Recognizing Entailments, Southampton.
Baldwin, T. (2005) ‘Bootstrapping Deep Lexical Resources: Resources for Courses’, Proceedings of the ACL SIGLEX Workshop on Deep Lexical Acquisition, Ann Arbor,
pp. 67–76.
Baldwin, T. and Bond, F. (2003) ‘Learning the Countability of English Nouns from Corpus
Data’, Proceedings of the 41st Annual Meeting of the Association for Computational
Linguistics, Sapporo, Japan, pp. 463–470.
Bannard, C. (2005) ‘Learning about the Meaning of Verb Particle Constructions from
Corpora’, Journal of Computer Speech and Language, vol.19.4,
Bennett, P. and Carbonell, J. (2005) ‘Detecting action items in e-mail’, Proceedings of the
28th Int. ACM SIGIR R&D in Information Retrieval, Salvador, Brazil.
Briscoe, E.J. (1994) Parsing (with) Punctuation, etc., Xerox European Research Laboratory, Grenoble, MLTT-TR-002 // ejb/
Briscoe, E.J. and J. Carroll (1995) ‘Developing and Evaluating a Probabilistic LR Parser
of Part-of-Speech and Punctuation Labels’, Proceedings of the ACL/SIGPARSE 4th
Int. Workshop on Parsing Technologies (IWPT95), Prague / Karlovy Vary, Czech
Republic, pp. 48–58.
Briscoe, E.J. and J. Carroll (1997) ‘Automatic extraction of subcategorisation from corpora’, Proceedings of the 5th ACL Conf. on Applied Nat. Lg. Proc., Morgan Kaufmann, pp. 356–363.
Briscoe, E.J. and J. Carroll (2002) ‘Robust accurate statistical annotation of general text’,
Proceedings of the 3rd Int. Conf. on Language Resources and Evaluation (LREC’02),
Las Palmas, Gran Canaria, pp. 1499–1504.
Briscoe, E., J. Carroll, J. Graham and A. Copestake (2002) ‘Relational evaluation schemes’,
Proceedings of the LREC’02 Workshop, Beyond PARSEVAL: Towards Improved Evaluation Measures for Parsing Systems, Gran Canaria.
Briscoe, E.J., C. Grover, B.K. Boguraev and J. Carroll (1987) ‘A formalism and environment for the development of a large grammar of English’, Proceedings of the 10th Int.
Joint Conf. on Artificial Intelligence, Milan, Italy, pp. 703–708.
Briscoe, E.J. and N. Waegner (1992) ‘Robust stochastic parsing using the inside-outside
algorithm’, Proceedings of the AAAI Workshop on Probabilistically-Based NLP Techniques , San Jose, Ca, pp. 39–53.
Buttery, P. and Korhonen A. (2005) ‘Large-scale analysis of verb subcategorization differences between child directed speech and adult speech’, Proceedings of the Verb
Workshop 2005, Interdisciplinary Workshop on the Identification and Representation
of Verb Features and Verb Classes, Saarland University.
Callmeier, U., A Eisele, U Schafer, and M Siegel (2004) ‘The DeepThought Core Architecture Framework’, Proceedings of the LREC’04, Lisbon.
Carroll, J. and E.J. Briscoe (2002) ‘High precision extraction of grammatical relations’,
Proceedings of the 19th Int. Conf. on Computational Linguistics (COLING’02),
Taipei, Taiwan, pp. 134–140.
Carroll, J., E. Briscoe and C. Grover (1991) A development environment for large natural
language grammars, Cambridge University Computer Laboratory, TR-233
Church, K. (1988) ‘A stochastic parts program and noun phrase parser for unrestricted
text’, Proceedings of the 2nd Conference on Applied Natural Language Processing,
Austin, Texas, pp. 136–143.
Clark, S. and Weir, D. (2002) ‘Class-Based Probability Estimation using a Semantic
Hierarchy’, Computational Linguistics, vol.28.2, 187–206.
Copestake, A. (2003) Report on the Design of RMRS, Deep Thought EU project Deliverable D1.1a
Curran, J. (2003) From distributional to semantic similarity, Phd Dissertation, Dept. of
Informatics, University of Edinburgh.
De Souza, S. (2005) Simple discourse topic summarisation, M.Phil Dissertation, University of Cambridge Computer Laboratory.
Elworthy, D. (1993) Part-of-speech tagging and phrasal tagging, Acquilex-II Working Paper 10, Cambridge University Computer Laboratory ($RASP/tag/documentation).
Elworthy, D. (1994) ‘Does Baum-Welch re-estimation help taggers?’, Proceedings of the
4th ACL Conference on Applied NLP, Stuttgart, Germany, pp. 53–58.
Gazdar, G., Klein, E, Pullum, G. and Sag, I. (1985) Generalized Phrase Structure Grammar, Blackwell, Oxford.
Grover, C., E.J. Briscoe and J. Carroll (1993) The Alvey Natural Language Tools Grammar
(4th Release), Cambridge University Computer Laboratory, TR-284
Grover, C., Lascarides, A. and Lapata, M. (2005) ‘A comparison of parsing technologies
for the biomedical domain’, Natural Language Engineering, vol.11.1, 27–65.
Jurafsky, D. and Martin, J. (2000) Speech and Language Processing, Prentice-Hall.
Kaplan, R., S. Riezler, T. H. King, J. T. Maxwell III, A. Vasserman and R. Crouch (2004)
‘Speed and accuracy in shallow and deep stochastic parsing’, Proceedings of the Human
Language Technology Conference and the 4th Annual Meeting of the North American
Chapter of the Association for Computational Linguistics (HLT-NAACL’04), Boston,
Khan, S. (2006) ‘Negation in sentiment analysis: employing contextual valence shifters to
enhance classification accuracy’, Proceedings of the ACL-Coling’06, submitted.
Korhonen, A. (2002) Subcategorization Acquisition, Techical Report UCAM-CL-TR-530,
Computer Laboratory, University of Cambridge.
Korhonen, Anna, Yuval Krymolowski and Zvika Marx (2003) ‘Clustering Polysemic Subcategorization Frame Distributions Semantically’, Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics, Sapporo, Japan, pp. 64–71.
Leidner, Jochen L., Johan Bos, Tiphaine Dalmas, James R. Curran, Stephen Clark, Colin
J. Bannard, Bonnie Webber, Mark Steedman (2003) ‘QED: The Edinburgh TREC2003 Question Answering System’, Proceedings of the TREC’03 QA Track, NIST.
McCarthy, D., and J Carroll (2003) ‘Disambiguating nouns, verbs, and adjectives using
automatically acquired selectional preferences’, Computational Linguistics, vol.29.4,
McCarthy, D., B Keller, and J Carroll (2003) ‘Detecting a continuum of compositionality
in phrasal verbs’, Proceedings of the ACL-2003 Workshop on Multiword Expressions,
Sapporo, Japan.
McCarthy, D., R. Koeling, J. Weeds and J. Carroll (2004) ‘Finding predominant word
senses in untagged text’, Proceedings of the 42nd Annual Meeting of the Association
for Computational Linguistics, Barcelona, Spain, pp. 280–287.
Medlock, B. (2006) Anonymisation of informal text, LREC’06, in press.
Nielsen, L. (2004) ‘Verb Phrase Ellipsis detection using Automatically Parsed Text’, Proceedings of the COLING’04, Geneva.
Nissim, M., and Markert, K. (2003) ‘Syntactic Features and Word Similarity for Supervised Metonymy Resolution’, Proceedings of the 41st Ann. Mtg of Assoc. for Comp.
Linguistics, Sapporo, Japan.
Nunberg, G. (1990) The linguistics of punctuation, CSLI Lecture Notes 18, Stanford, CA.
Preiss, J. (2003) ‘Using Grammatical Relations to Compare Parsers’, Proceedings of the
Eur. Assoc. Comp. Linguistic, Budapest, pp. 291–298.
Preiss, J. (2004) ‘Probabilistic Word Sense Disambiguation’, Journal of Computer Speech
and Language, vol.18.3, 319–337.
Preiss, J. and E.J. Briscoe (2003) ‘Intermediate parsing for anaphora resolution? Implementing the Lappin and Leass non-coreference filters’, Proceedings of the Proc. of the
ACL03 Workshop on Anaphora Resolution, Morgan Kaufmann, pp. 1–6.
Pustejovsky, James, Patrick Hanks, and Anna Rumshisky (2004) ‘Automated Induction
of Sense in Context’, Proceedings of the Coling’04, Geneva.
Sagae,K., A Lavie, B MacWhinney (2005) ‘Automatic Measurement of Syntactic Development in Child Language’, Proceedings of the 43rd Annual Meeting of the ACL, Ann
Arbor, pp. 197–204.
Sampson, G. (1995) English for the Computer, Oxford University Press, Oxford, UK.
Sporleder, C. and Lascarides, A. (2005) ‘Exploiting Linguistic Cues to Classify Rhetorical Relations’, Proceedings of the Recent Advances in Natural Language Processing
(RANLP-05), Borovets, Bulgaria.
Tsang V. and S. Stevenson (2004) ‘Using selectional profile distance to detect verb alternations’, Proceedings of the HLT/NAACL 2004 Workshop on Computational Lexical
Semantics, Boston.
Vlachos, A., Gasperin, C., Lewin, I., Briscoe, E. J. (2006) ‘Bootstrapping the recognition
and anaphoric linking of named entities in Drosophila articles’, Proceedings of the
Pacific Symposium in Biocomputing, Hawaii.
Wang, D. (2005) Classifying text using linguistic features, M.Phil Dissertation, University
of Cambridge Computer Laboratory.
Watson, R., Preiss, J. and Briscoe, E.J. (2003) ‘The Contribution of Domain-independent
Robust Pronominal Anaphora Resolution to Open-Domain Question-Answering’, Proceedings of the Int. Symposium on Reference Resolution and its Applicati on to
Question-Answering and Summarisation, Venice.
Watson, R., J. Carroll and E. Briscoe (2005) ‘Efficient extraction of grammatical relations’, Proceedings of the 9th Int. Workshop on Parsing Technologies (IWPT’05),
Vancouver, Ca..
Weeds, J., B Keller, D Weir, I Wakeman, J Rimmer (2004a) ‘Natural language expression
of user policies in pervasive computing environments’, Proceedings of the OntoLex
2004 (LREC Workshop on Ontologies and Lexicons), Lisbon.
Weeds, J., D Weir, and D McCarthy (2004b) ‘Characterising Measures of Lexical Distributional Similarity’, Proceedings of the Coling’04, Geneva.
Weeds J., D Weir, and B Keller (2005) ‘The Distributional Similarity of Sub-Parses’,
Proceedings of the ACL SIGLEX Workshop on Deep Lexical Acquisition, Ann Arbor,
pp. 7–13.
Yallop, J., Korhonen, A. and Briscoe, E.J. (2005) ‘Automatic acquisition of adjectival
subcategorisation from corpora’, Proceedings of the 43rd Assoc. for Comp. Ling.,
Morgan Kaufmann, pp. 614–621.
Yamada, I. and Baldwin, T. (2004) ‘Automatic Discovery of Telic and Agentive Roles
from Corpus Data’, Proceedings of the 18th Pacific-Asia Conf. on NLP, Tokyo, Japan,
pp. 115–126.
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF