TOWARDS A DYNAMIC PRAGMATICS A DISSERTATION SUBMITTED TO THE DEPARTMENT OF LINGUISTICS

TOWARDS A DYNAMIC PRAGMATICS A DISSERTATION SUBMITTED TO THE DEPARTMENT OF LINGUISTICS
TOWARDS A DYNAMIC PRAGMATICS
A DISSERTATION
SUBMITTED TO THE DEPARTMENT OF LINGUISTICS
AND THE COMMITTEE ON GRADUATE STUDIES
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
Sven Lauer
August 2013
Abstract
This dissertation explores the interplay of conventional and interactional factors in
the interpretation of natural language utterances. It develops a formal framework,
dynamic pragmatics, in which pragmatic inferences arise as contextual entailments
in a dynamic system in which information states are updated with information
about the occurrence of utterance events (in contrast to dynamic semantics, where
information states are updated with the content of linguistic expressions). In this
way, the framework is able to faithfully model Gricean pragmatic inference as
interlocutors’ reasoning about each other’s utterance choices.
Linguistic utterances are analyzed as having essential effects of two distinct
types: Epistemic effects (i.e., effects on the information states of the interlocutors)
and normative effects (i.e., effects on the interlocutors’ commitments). The latter
effects are carried by extra-compositional, normative conventions of use that mediate the form–force mapping; the former arise largely due to the interlocutors’
presumptions about each other’s beliefs, preferences, and method of determining
which (utterance) actions are best (i.e., practical reasoning).
The framework of dynamic pragmatics allows us to consistently take a thoroughly Gricean perspective on language use, and allows us to explore how the
interpretation of an utterance arises through the interplay of sentential force, content, and context. At the same time, the framework of dynamic pragmatics sheds a
new light on the nature of conversational implicature, and language use in general.
iv
Acknowledgements
There is nothing I can do about it: This paragraph is going to fall short. There is
no way to express how much I owe Cleo Condoravdi, and how grateful I am to
her, for being my teacher, advisor, mentor, and collaborator. Whenever one leaves
a place, there are people and things one leaves behind that one rather would not. I
have left many places, but I feel I have never left behind anything so valuable as I
will when I leave Stanford and will no longer able to work closely with Cleo. I will
miss her guidance, her understanding, her generosity, her insight . . . my words fail
me, so I have to say it simply. Thank you, Cleo.
While Cleo’s influence on this dissertation cannot be overstated, Chris Potts
also shaped it in many crucial ways. I am indebted to him for his help, advice,
encouragement, guidance, enthusiasm and expertise, all of which he has always
given freely and generously. He has taught me a lot, and his help was essential
in making this dissertation, and its completion, a possibility. And I am grateful
to Paul Kiparsky for many valuable comments, for his useful skepticism, and for
many helpful discussions. I feel very lucky to have had Cleo, Chris and Paul as my
dissertation committee.
I am also grateful to my former teachers, who have all influenced my thinking about the topics of this dissertation: Paul Dekker, Robert van Rooij, Jeroen
Groenendijk, Frank Veltman, Peter Bosch, Carla Umbach and Graham Katz.
The ideas that form the basis for this dissertation have gestated in my mind for
a long time, and so there are many colleagues and friends who have influenced
them over the years. It would be futile to try and compile a complete list, but a
few of them deserve special mention. In the following pages, Michael Franke will
v
recognize many themes familiar from long after-yoga dinner conversations during
our joint time in Amsterdam. In many ways, I first started to really think about
imperatives and clause types when I sat in Magdalena Kaufmann’s ESSLLI course
and when reading her dissertation thereafter. Her work was a crucial inspiration
for Cleo’s and my work on these topics. When Anna Chernilovskaya stopped by in
Stanford, she was busy finding answers to questions about exclamatives that I had
not even thought to ask, and our collaboration was both fruitful and stimulating.
And finally, there is Alex Djalali, who has listened to me ranting and ranted back,
who has been my office mate and fellow sufferer, and who I will always regard
as my friend and brother. I hereby grant him permission to try and make me
uncomfortable by hugging me whenever he pleases.
The Stanford linguistics department has been a wonderful place to write this
dissertation, and I am thankful to everyone there. Alex is not the only colleague
who became a dear friend. There are also Seung Kyung Kim, Laura Whitton, Jessica
Spencer and Natalia Silveira. Getting to know these people alone was worth the
struggle of graduate school, the degree is an extra bonus.
This dissertation, and my time at Stanford, would not be the same without the
stimulating environment of the rest of the semantics crowd, broadly construed,
both former and current, permanent and visiting (in no particular order): Lauri
Karttunen, Annie Zaenen, Eve Clark, Ivan Sag, Stanley Peters, Scott Grimm, Nola
Stephens, Marie-Catherine de Marneffe, Chigusa Kurumada, Marisa Tice, Eric
Acton, Tania Rojas-Esponda, Dasha Popova, James Collins, Lelia Glass, Bonnie
Krejci, Tyler Schnoebelen, Mason Chua, Adrian Brasoveanu, Uli Sauerland, Patricia
Amaral, Theres Grueter, Dan Lassiter, Tine Breban, Lucas Champollion, Fabio
del Prete, and everyone else who ever attended a Meaningful Lunch. This list
deliberately leaves out one person, because she needs to be set apart. I cannot
thank Beth Levin enough for all her help and for all that she has taught me.
Finally, there is my family. I would not have made it to Stanford, or anywhere,
without them: My uncle Dietmar, my aunt Sylvia and the rest of the Holeins. My
grandparents Hubert and Elli. My mother Hanne, and my brothers Jan and Lars,
who always have my back. I love you very, very much.
vi
Contents
Abstract
iv
Acknowledgements
v
1
Introduction
1
1.1
Going beyond conversational implicatures . . . . . . . . . . . . . . .
1
1.2
The centrality of clause typing . . . . . . . . . . . . . . . . . . . . . . .
4
1.3
Two kinds of Griceanism . . . . . . . . . . . . . . . . . . . . . . . . . .
5
1.4
Utterance choice and dynamic pragmatics . . . . . . . . . . . . . . . .
7
1.5
Overview of the dissertation . . . . . . . . . . . . . . . . . . . . . . . .
8
2
The basic system
11
2.1
Dynamic pragmatics: The very idea . . . . . . . . . . . . . . . . . . . 11
2.2
Communicating contents . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3
The basic system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3.1
Languages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3.2
Models for time and belief . . . . . . . . . . . . . . . . . . . . . 17
2.3.3
Events . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.4
Constraining belief change . . . . . . . . . . . . . . . . . . . . 23
2.3.5
The dynamic perspective . . . . . . . . . . . . . . . . . . . . . 26
2.4
Communicating contents in the basic system . . . . . . . . . . . . . . 29
2.5
Communicating with and without intentions . . . . . . . . . . . . . . 31
2.6
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
vii
3
4
Clause types
35
3.1
Introducing clause types . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.2
Denotation type is not a sufficient guide to function . . . . . . . . . . 39
3.3
A case for extra-compositional constraints on use . . . . . . . . . . . . 42
3.4
Illocutionary acts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.5
Sincerity conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.6
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
The sentential force of declaratives
53
4.1
Mapping the territory . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.2
Declaratives as governed by Lewis-conventions . . . . . . . . . . . . 60
4.3
Declaratives as expressing beliefs . . . . . . . . . . . . . . . . . . . . . 64
4.3.1
Bach and Harnish’s (1979) on communicative speech acts . . . 67
4.3.2
R-intentions à la Bach & Harnish . . . . . . . . . . . . . . . . . 68
4.3.3
Expressing a belief as R-intending . . . . . . . . . . . . . . . . 73
4.3.4
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.4
Counts-as rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.5
Normative theories of clause typing . . . . . . . . . . . . . . . . . . . 79
4.6
4.7
4.5.1
Normative preconditions on utterances . . . . . . . . . . . . . 80
4.5.2
Normative effects of utterances . . . . . . . . . . . . . . . . . . 83
4.5.3
Commitment to be held responsible for consequences . . . . . 87
4.5.4
Commitments to act as though one has a belief . . . . . . . . . 89
Commitments to act according to an attitude . . . . . . . . . . . . . . 90
4.6.1
Declaratives with incompatible contents . . . . . . . . . . . . . 92
4.6.2
Declaratives express beliefs . . . . . . . . . . . . . . . . . . . . 92
4.6.3
Other normative consequences of declaratives . . . . . . . . . 93
The enduring commitments of loose talk: a case for the commitmentto-belief view . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.7.1
Loose talk—the basic facts . . . . . . . . . . . . . . . . . . . . . 97
4.7.2
Loose talk in multi-sentence discourses . . . . . . . . . . . . . 99
4.7.3
Loose talk in the commitment-to-belief account . . . . . . . . 101
viii
4.7.4
5
Action Choice and Commitment
5.1
5.2
5.3
6
Communicative reasons for retraction . . . . . . . . . . . . . . 102
105
Modeling action choice . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.1.1
Actions and agents . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.1.2
Beliefs about action choices . . . . . . . . . . . . . . . . . . . . 110
Modeling Preferences . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
5.2.1
Preference structures . . . . . . . . . . . . . . . . . . . . . . . . 113
5.2.2
A working definition for Opt . . . . . . . . . . . . . . . . . . . 117
5.2.3
Reasoning about preferences . . . . . . . . . . . . . . . . . . . 120
5.2.4
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
Modeling commitment . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
5.3.1
Commitments exclude possible future states of the world . . 124
5.3.2
Commitments to beliefs and preferences . . . . . . . . . . . . 127
5.4
Modeling conventions of use . . . . . . . . . . . . . . . . . . . . . . . 130
5.5
Communicating contents, again . . . . . . . . . . . . . . . . . . . . . . 131
5.6
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Commitments to preferences
6.1
136
Imperatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
6.1.1
Imperatives as creating commitments to preferences . . . . . 139
6.1.2
Deriving directive uses . . . . . . . . . . . . . . . . . . . . . . . 143
6.1.3
Deriving advice uses . . . . . . . . . . . . . . . . . . . . . . . . 148
6.1.4
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
6.2
Performative uses of desideratives . . . . . . . . . . . . . . . . . . . . 156
6.3
Interrogatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
6.4
Denotation types and the form–force mapping . . . . . . . . . . . . . 164
6.4.1
Denotations for imperatives . . . . . . . . . . . . . . . . . . . . 165
6.4.2
Denotations for interrogatives . . . . . . . . . . . . . . . . . . 168
6.4.3
The (possible) indeterminacy of denotation types . . . . . . . 170
6.4.4
One convention for declaratives and imperatives? . . . . . . . 172
6.4.5
The form of the clause-typing conventions . . . . . . . . . . . 173
ix
6.4.6
6.5
7
Representing force in the compositional system . . . . . . . . . . . . . 177
Explicit performatives
181
7.1
Saying makes it so . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
7.2
Explicit performatives as ‘self-verifying’ . . . . . . . . . . . . . . . . . 184
7.3
Truth conditions for performative verbs . . . . . . . . . . . . . . . . . 187
7.4
Deriving self-verification . . . . . . . . . . . . . . . . . . . . . . . . . . 189
7.5
7.6
7.7
8
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
7.4.1
Introducing illocutionary verbs into Sen . . . . . . . . . . . . . 190
7.4.2
Performatives with promise and order are self-verifying . . . 191
7.4.3
Performatives with claim are self-verifying . . . . . . . . . . . 192
7.4.4
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
Further predictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
7.5.1
Reportative uses of performative predicates . . . . . . . . . . 194
7.5.2
Performatives and logical operators . . . . . . . . . . . . . . . 195
7.5.3
‘Illocutionary’ vs. ‘perlocutionary’ verbs . . . . . . . . . . . . 199
Comparison to existing approaches . . . . . . . . . . . . . . . . . . . . 201
7.6.1
Searle (1989): Explicit performatives as declarations . . . . . . 202
7.6.2
Bach and Harnish (1979, 1992) . . . . . . . . . . . . . . . . . . 207
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
Exclamatives and expressives
8.1
8.2
8.3
213
Exclamatives as an expressive clause type . . . . . . . . . . . . . . . . 215
8.1.1
The two implications of exclamatives . . . . . . . . . . . . . . 215
8.1.2
The nature of the expressive implication . . . . . . . . . . . . . 217
More expressive meanings . . . . . . . . . . . . . . . . . . . . . . . . . 224
8.2.1
Non-sentential expressions . . . . . . . . . . . . . . . . . . . . 224
8.2.2
Non-at-issue meanings . . . . . . . . . . . . . . . . . . . . . . . 224
8.2.3
Expressive vs. prescriptive conventions . . . . . . . . . . . . . 228
8.2.4
Determining the correct convention type . . . . . . . . . . . . 229
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
x
9
Conversational implicatures
9.1
9.2
9.3
232
Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
9.1.1
Maxims and preferences . . . . . . . . . . . . . . . . . . . . . . 235
9.1.2
Alternative utterances . . . . . . . . . . . . . . . . . . . . . . . 237
9.1.3
Two types of preferences and a visual representation of Opt . 241
Some classical implicatures . . . . . . . . . . . . . . . . . . . . . . . . 243
9.2.1
A ‘relevance’ implicature . . . . . . . . . . . . . . . . . . . . . 245
9.2.2
A scalar implicature . . . . . . . . . . . . . . . . . . . . . . . . 247
9.2.3
The ‘epistemic step’ . . . . . . . . . . . . . . . . . . . . . . . . 250
9.2.4
Intended implicatures . . . . . . . . . . . . . . . . . . . . . . . 252
9.2.5
Unintended implicatures . . . . . . . . . . . . . . . . . . . . . 254
Need a Reason: Mandatory Gricean implicatures . . . . . . . . . . . . 256
9.3.1
Optionality and cancelability . . . . . . . . . . . . . . . . . . . 257
9.3.2
The ignorance implicature of disjunction . . . . . . . . . . . . 262
9.3.3
The NaR implicature of disjunction in dynamic pragmatics . . 266
9.3.4
Generalizating NaR . . . . . . . . . . . . . . . . . . . . . . . . 271
9.3.5
More NaR implicatures? . . . . . . . . . . . . . . . . . . . . . . 274
10 Outlook
281
10.1 The role of intentions in pragmatic theory . . . . . . . . . . . . . . . . 281
10.2 Ambiguity and underspecification . . . . . . . . . . . . . . . . . . . . 283
10.3 A question of commitment . . . . . . . . . . . . . . . . . . . . . . . . . 285
10.4 Belief revision, salience, and awareness . . . . . . . . . . . . . . . . . 286
10.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
A The basic system
289
Bibliography
293
xi
List of Figures
2.1
A branching-time model (from Condoravdi 2002) . . . . . . . . . . . 18
9.1
Decision in worlds vp`rel where the speaker believes p and takes it to
be ‘relevant’ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
9.2
Decision in worlds vp`
rel
where the speaker believes p and does not
take it to be ‘relevant’ . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
9.3
Decision in worlds vp where the speaker believes p . . . . . . . . . . 248
9.4
Decision in worlds ve where the speaker believes e, but not p . . . . 248
9.5
Decision in worlds v
9.6
Decision in worlds vp`
e
where the speaker believes neither e nor p . . 248
where the speaker believes e and p, on
rel
the assumption that p is not relevant . . . . . . . . . . . . . . . . . . . 249
9.7
Decision in worlds vi e^
p
where the speaker knows that e is true
and p is false, and believes the implicature will not be drawn . . . . . 253
9.8
Decision in worlds vi
p^e
where the speaker knows that e is true
and p is false, and believes the implicature will be drawn . . . . . . . 253
9.9
Decision in worlds v
p^ q
9.10 Decision in worlds vp^
9.11 Decision in worlds v
where the speaker knows neither p nor q 268
q
where the speaker knows p but not q . . . 268
p^q
where the speaker knows q but not p . . . 268
9.12 Decision in worlds vp^q where the speaker knows both p and q . . . 268
9.13 Decision in worlds where the speaker believes p and has a preference
against revealing that . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
9.14 Decision in worlds where the speaker believes p and does not have
a preference against revealing that . . . . . . . . . . . . . . . . . . . . 270
xii
Chapter 1
Introduction
This dissertation is about pragmatics, in a broadly Gricean sense. It is not a
dissertation about conversational implicatures, at least not in the classical sense of
the term. Implicatures will play a role (Chapter 9 is dedicated to the them), but they
are not the central topic. A large part of the dissertation will focus on a question
that may seem quite un-Gricean, due to its focus on linguistic convention: What
kind of linguistic convention makes it so that sentences of different types—such
as declaratives, interrogatives and imperatives—are used in different ways, and
support different kinds of pragmatic inferences?
This question, however, will be addressed from a very Gricean angle. Pragmatic inference is construed as language users’ reasoning about utterance events.
Or, more precisely, as language users’ reasoning about how utterance events are
chosen. A central aim of this dissertation is to show that consistently taking such a
perspective is fruitful, indeed, necessary if we want to understand language use.
The dissertation develops a formal framework, Dynamic Pragmatics, that enables
us to consistently take such a Gricean perspective.
1.1
Going beyond conversational implicatures
Of course, in the theory of Grice (1975), conversational implicatures are (intended)
inferences about utterance choices. But implicatures, at least as they are usually
1
CHAPTER 1. INTRODUCTION
2
understood, are only a special case of such inferences. And limiting oneself to
the study of implicature excludes a large number of interesting phenomena from
consideration.
Firstly, the classical cases of implicatures are almost always strengthening inferences.1 A speaker utters a declarative sentence, and communicates its truthconditions. Implicatures are added to this communicated meaning.
But not all inferences about utterance choice are of this kind. There are also
inferences weaken the conveyed content. This is what happens in what Lasersohn
(1999) calls Loose Talk, when a speaker utters a sentence like (1.1) in order to
convey that Mary arrived around three o’clock.2
(1.1)
Mary arrived at three.
In other cases, what is conveyed by an utterance is neither a logical strengthening
nor a weakening of the semantic meaning of the sentence. This is what happens
if the audience has reason to doubt the speaker’s honesty. If such a speaker utters
(1.1), the hearer will not come to believe its semantic content, but he will still draw
inferences about why the speaker said what he said (in the way that he said it, at
the time that he said it).
Secondly, the Gricean theory of implicature, as originally introduced and usually understood, presupposes a theory of how sentences of different types get
associated with their force. The theory starts from the idea that a speaker uttered
a declarative sentence in order to convey information, that is, in order to make the
hearer believe that the truth-conditional content of the sentence is true. But the
force of a given sentence, of a given type, in any given context, depends in part on
the inferences hearers draw about the speaker’s utterance choice. Taking a broader
perspective on pragmatic reasoning allows us to model such cases.
Thirdly, in construing pragmatic theory as the study of conversational implicature, somewhat paradoxically, one runs a considerable risk of misunderstanding
1
There is one notable exception, viz., instances where the quality maxim is flouted, as in And
I’m a monkey’s uncle.
2
Loose talk will play a crucial role in Chapter 4, in particular Section 4.7.
CHAPTER 1. INTRODUCTION
3
conversational implicature itself. That is because it is tempting to think of implicatures not as instance of reasoning about language use, but rather to view them as
just another kind of implication an utterance may have. That is, it is easy to slip into
moving from the perspective articulated in (1.2) into the perspective articulated in
(1.3), or even the one articulated in (1.4).
(1.2)
The ‘inside’ perspective
Grice taught us that is useful and necessary to think of language use as
a species of purposive human behavior. Implicatures are inferences that
arise due to interlocutors being aware of this. Their properties follow from
the way we derive them.
(1.3)
The ‘outside’ perspective
There is a process/module that generates implicatures. Grice gave a theory
of how this process/module works and taught us that they have certain
properties (optionality, cancelability, non-detachability . . . ).
(1.4)
The ‘just another implication’ perspective
Conversational implicatures are implications just like at-issue entailments,
presuppositions, conventional implicatures etc., which happen to have
certain properties, like optionality, cancelability, non-detachability . . .
I don’t mean to deny that the ‘outside perspective’ is useful. It is a convenient
abbreviated way to think of pragmatic inference, and it allows us to set aside the
details of pragmatic theory when we are working on other issues.
However, I maintain that if we are taking the ‘outside’ perspective, we have to
keep in mind that we have taken a conceptual shortcut, and that the inferences we
take for granted arise in a complex manner from factors that influence language
use in general.
And I think it is hardly ever prudent to take the ‘just another implication’
perspective. This perspective obscures the difference between semantic facts (i.e.,
facts about the grammar of the language) and pragmatic facts (i.e., facts about
language use). This is problematic because the reasoning giving rise to pragmatic
CHAPTER 1. INTRODUCTION
4
facts is operative on every occasion of language use, not just in cases where an
identifiable implicature arises. It is easy to overlook this fact if one does not
carefully distinguish between semantic and pragmatic implications. As I will argue,
this has led to consequential misunderstandings about the nature of conversational
implicature, both about what Gricean theory predicts them to be like and about
how they behave empirically.
1.2
The centrality of clause typing
Clauses of different types have different uses, and they license different kinds of
pragmatic inferences. A declarative like (1.5a) is typically used to convey information, an interrogative like (1.5b) is typically used to request information, while an
imperative like (1.5c) is typically used to get the addressee to do something.
(1.5)
a.
You will come tomorrow.
b.
Will you come tomorrow?
c.
Come tomorrow!
(declarative)
(interrogative)
(imperative)
There must be something conventional about this association of a certain class of
expressions with their use, and a large part of this dissertation will be concerned
with the question what this conventional association is like. This focus on the issue
of clause-typing is motivated by two considerations.
Firstly, there is surprisingly little work on the conventional link between clauses
of different types and their use, at least in the formally-oriented semantics literature.
There was some early work on this issue in philosophy (e.g., Lewis (1975)), but soon,
it seems, researchers abandoned the issue. In part this was likely due to the fact
that the dominant paradigm in speech act theory (i.e., that of Searle (1969, et seq.))
made the project of assigning a conventionally-specified use to sentences based on
their type seem hopeless: Sentences of any given type can be used to perform acts
of almost any given Searlean illocutionary type. Perhaps because of this, work in
linguistics has often focussed on declarative sentences, and tried to approach other
CHAPTER 1. INTRODUCTION
5
sentence types mainly by studying their embedded occurrences. Only in the last
decade or so has there been renewed, sustained interest in understanding the uses
of various clause types in the formal semantics literature.3 Developing a systematic
framework for studying the conventionally specified use of sentences of different
types is thus a very timely project.
Secondly, it is this conventional connection between sentences and their use that
connects semantic content, as it is studied in linguistics, with language use, and inferences about utterance choice. As I just pointed out, the classical Gricean account
of implicatures starts from the assumption that a speaker utters a (declarative) sentence in order to convey information, and thus, as I argue in Chapter 3, presupposes
an answer to the question of how sentences are conventionally associated with a
certain use. And if we want to take a Gricean perspective more generally—if we
want to investigate how interlocutors reason about each other’s action choices—we
need to know how the contents we study in linguistic semantics relate to the use
of sentences. An understanding of clause typing thus is central to developing a
formal framework that lets us take a pragmatic perspective in general.
1.3
Two kinds of Griceanism
I have said that this dissertation takes a Gricean perspective on language use. This
is correct insofar as it construes pragmatic inference as interlocutors’ reasoning
about each other’s utterance choices.
In another way, however, this dissertation takes a rather un-Gricean perspective.
It does not involve any central appeal to the notion of intention and it does not
construe communication and pragmatic inference as being essentially intentionrecognition. This is un-Gricean, as Grice famously developed an account of speaker
meaning that analyzed the concept in terms of intention (Grice 1957, Grice 1969,
Grice 1982). To mean something, for Grice, was to have an intention; to understand
3
e.g., Portner (2005, 2007), Schwager (2006), Davis (2009, 2011) Kaufmann (2012), Condoravdi
and Lauer (2012) on imperatives; Zanuttini and Portner (2003), Rett (2008, 2011), Castroviejo Miró
(2008) on exclamatives; Gunlogson (2003, 2008), Groenendijk and Roelofsen (2009), Davis (2009,
2011), Farkas and Roelofsen (forthcoming) on declaratives and interrogatives.
CHAPTER 1. INTRODUCTION
6
what someone means was to recognize the corresponding intention.
Now, I do not doubt that when Grice (1975) presented his theory of conversational implicature, and made clear that, for him, a speaker implicates something if
he means it (but does not say it), he had in mind just this kind of speaker-meaningas-intention. But it seems to me that the basic idea of his account of pragmatic
reasoning as reasoning about utterance choices is quite independent from the idea
that intentions and their recognition are central to language use.
I do not want to make a grand claim that intentions are irrelevant to understanding language use (and, undoubtedly, in certain circumstances they are), but I will
not start from the assumption that intentions are central. Throughout this dissertation, I will occasionally raise the question whether anything crucial has been left
unsaid because I have not appealed to intentions, and the conclusion will generally
be that it has not. This does not establish the negative claim that intentions are not
central to pragmatic reasoning, but it raises the question whether intentions and
intention recognition is central to the kind of questions this dissertation aims to
answer.
Griceans of a more linguistic bent (e.g., Levinson (2000)) are sometimes accused
by Gricean philosophers (e.g., Bach (2012)) of confusing epistemological issues
with ontological or metaphysical ones. For example, they are accused of confusing
the question of how an addressee can infer the existence of an intended inference
with the issue of how it is determined (in the metaphysical sense) whether there is
an intended inference. I am sure this charge is sometimes warranted, but I am not
certain it is always the linguists who are confused. Instead, in many cases, it seems
to me that linguistic pragmaticists simply talk about, and are mainly interested
in, the ‘epistemological’ issue of how inferences are derived on the part of the
interlocutors and have nothing to say about metaphysical or ontological issues. If
a philosopher reads such a linguist’s work on the assumption that the linguist is
talking about ontological or metaphysical issues, it is not surprising that he will
think the linguist is confused.
In large part, this dissertation takes the ‘epistemological’ perspective, at least as
far as pragmatic inference is concerned. I am interested in how interlocutors reason
CHAPTER 1. INTRODUCTION
7
about each other’s utterances, what they learn from each other’s utterances, and in
how far what they learn is based on their linguistic knowledge as opposed to their
knowledge about how other people generally behave. This is likely part of the
reason why intentions do not seem to play such a great role in the questions I am
investigating, while philosophers like Grice and Bach take them to be so central.
They are investigating questions different from the ones I am concerned with here.
1.4
Utterance choice and dynamic pragmatics
The framework of Dynamic Pragmatics developed in this thesis aims to faithfully treat pragmatic inference as interlocutors’ reasoning about utterance choice.
Utterance choice is construed as an instance of action choice in general. As such,
this dissertation is of a piece with recent game-theoretical approaches to pragmatics
(Parikh 2001, van Rooij 2004, Benz and van Rooij 2007, Jäger 2007, Franke 2009, Jäger
and Ebert 2009, Franke 2011, Jäger 2012, a.o.). It differs from these approaches in
two ways. Firstly, it largely abstracts away from the question what the correct ‘decision procedure’ is that we should assume agents are using when deciding which
(utterance) action to perform, while these approaches generally make very specific
assumptions about the decision procedure. Secondly, in this dissertation, I will
mainly be concerned with the question what conventional constraints on use there
are, while game-theoretic treatments usually ignore this question.4 I see the framework developed here as largely complementary to these game-theoretic approaches,
rather than as an alternative to them.
Besides being strongly inspired, as the game-theoretic treatments are, by Grice’s
work, the framework developed here also owes much to the work of Stalnaker
(1978, et seq.) and that of other authors building on his insights. Stalnaker’s conception has sometimes been referred to as ‘dynamic pragmatics’ (e.g., by Schlenker
(2010, p. 390)), though I am not aware that Stalnaker has used the label himself. As
4
Most game-theoretic models of pragmatics employ a semantic interpretation function that is
assumed to be given by convention, but they do not invoke any conventional constraints on use
proper.
CHAPTER 1. INTRODUCTION
8
in Stalnaker’s work, a crucial role is played by interlocutors public beliefs (though
I construe this notion somewhat differently, cf. Chapter 4) and how they get updated in the course of a conversation. The conception offered here differs in that it
assumes that there also are public preferences, and in that it emphasizes the normative
character of both notions. Finally, it explicitly models action choice, and reasoning
about action choice, which doesn’t usually play a big role in Stalnaker’s writings
on language.5
1.5
Overview of the dissertation
The following chapters develop a formal framework for pragmatic reasoning that
integrates an articulated theory of clause typing and applies the framework to
a number of phenomena. The first five chapters alternate between introducing
the formal set-up and conceptual and empirical considerations that motivate it,
while the chapters thereafter apply the framework to specific phenomena. The
progression is as follows.
Chapter 2 introduces the basic idea of a dynamic pragmatics as the term is understood here, and sets up the basics of the formal framework. The exposition
uses as its running example a very simple pragmatic inference, viz., the inference to the truth of a declarative utterance.
Chapter 3 introduces some of the basic questions raised by the existence of different clause types. It argues that the association between sentences of different
types and their uses must be conventional in nature, and lays out some basic
assumptions about clause typing that are made in this dissertation.
Chapter 4 focuses on declarative sentences. It explores various hypotheses about
what their conventionally-specified use is, and what kind of convention specifies it. The chapter concludes with an informal characterization of my own
5
Of course, it does play a big role in his writings on the epistemic foundations of game-theory
(Stalnaker 1994, Stalnaker 1996), which also have shaped, though more indirectly, some of the ideas
in this dissertation.
CHAPTER 1. INTRODUCTION
9
proposal, building on work in Condoravdi and Lauer (2011, 2012) and Lauer
(2012). Utterances of declaratives are argued to commit their speaker to a
belief in the truth of the uttered sentence.
Chapter 5 extends the framework introduced in Chapter 2 with the tools necessary
to formally implement the theory of declaratives proposed in Chapter 4.
Action choice, preferences and commitment. The chapter concludes with
reconstructing the account of the pragmatic inference in Chapter 2.
Chapter 6 moves beyond declarative sentences. The main focus is on imperatives, which are claimed to commit their speaker to a preference instead of a
belief. The main focus of the chapter is to demonstrate that the framework of
dynamic pragmatics allows us to show how the varied uses of imperatives
arise in context from an interaction of semantic content, sentential force, and
interactional reasoning.
Chapter 7 takes up the issue of explicitly performative utterances, which have played
a great role in speech act theory. It integrates the analysis of Condoravdi
and Lauer (2011) into the current framework, and shows how the central
properties of these sentences arise straightforwardly from an interaction of
the proposed lexical meanings with the pragmatic system.
Chapter 8 offers some preliminary considerations on the question whether all
conventional constraints on use should be understood in terms of commitments, as the previous chapters have proposed for declaratives, imperatives
and interrogatives. I argue that this is not the case, and that we should recognize different kinds of conventions, of the kind proposed by Lewis (1969),
in particular for exclamative sentences and various ‘expressive’ items.
Chapter 9 deals with conversational implicatures. It shows how some standard
cases can be treated in the framework of dynamic pragmatics, mainly to illustrate how various ‘optimization-based’ theories of implicatures (including the
game-theoretic ones mentioned above) fit into the current set-up. The chapter
CHAPTER 1. INTRODUCTION
10
then goes on to show how the framework of dynamic pragmatics lets us appreciate a significant, and surprising, prediction that such optimization-based
theories (including Grice’s own) make. Gricean pragmatic inferences can be
mandatory, i.e., neither cancelable nor optional. This has significant consequences, both because optionality and cancelability have often been used as
tests for implicature-hood, and because it potentially extends the domain of
Gricean pragmatics in a significant way. Phenomena that usually have been
taken to be outside the reach of Gricean explanations may be amenable to a
Gricean treatment after all.
Chapter 2
The basic system
In this chapter, I will work towards a model of a very simple pragmatic inference—
so simple, indeed, that at first glance it may not seem like a pragmatic inference at
all. Suppose Ad is on the phone with Sp, and Sp utters (2.1).
(2.1)
It is raining in Chicago.
From observing Sp’s utterance in (2.1), Ad can learn, if the context is right, that it
is raining in Chicago. This is the inference this chapter is dedicated to modeling.
In doing so, I will set up the foundations of the framework of dynamic pragmatics
that will be developed and applied in the rest of this dissertation.
2.1
Dynamic pragmatics: The very idea
The general idea of a Dynamic Pragmatics is to appropriate some of the tools
of Dynamic Semantics to model information change. The basic idea of dynamic
semantics is that utterances change information states, which are taken to either
represent what the addressee of an utterance knows, what the speaker knows,
what is Common Ground between the interlocutors, or something of the kind. The
various incarnations of dynamic semantics (Kamp 1981, Heim 1983, Groenendijk
and Stokhof 1991, Groenendijk, Stokhof and Veltman 1995, Veltman 1996, Beaver
11
CHAPTER 2. THE BASIC SYSTEM
12
2001, among many others) thus start by specifying a formal representation of
such information states (sets of possible worlds, files, discourse representation
structures, assignment functions, etc.), and then analyze the meaning of linguistic
expression as Update Potentials—that is, update operations on information states
(usually represented as functions or relations over information states). We can
schematize the idea as in (2.2):
(2.2)
new info state =
old info state + meaning of the uttered expression
For present purposes, we can identify the meaning of an expression with its informational content, and so we can instantiate this general schema for the sentence in
(2.1) as in (2.3).
(2.3)
new info state =
old info state + the information that it is raining in Chicago
In dynamic pragmatics, we also conceive of utterances as changing information
states, but the information we add is of a different kind: It is the information
that a certain utterance has happened. The information states that we update are
intended to model the beliefs of the addressee (or more generally, any interlocutor).
Schematically, for (2.1):
(2.4)
new info state =
old info state + the information that Sp uttered It is raining in Chicago.
I immediately introduce two idealizing assumptions. Firstly, I will assume throughout that utterance observation is perfect: When a speaker utters a sentence, his
interlocutors will perceive it, and not doubt their perceptual capacities, i.e., they
will come to believe that the utterance happened. Secondly, I abstract away from
all issues of ambiguity and semantic underspecification, and essentially assume
that interlocutors observe utterances of disambiguated logical forms.
CHAPTER 2. THE BASIC SYSTEM
13
Both of these assumptions are, of course, rarely warranted in everyday discourse, but making them will keep our framework simpler. In the concluding
Chapter 10, I will return to this assumption and indicate how it could be lifted. In
the meantime, we will see that there is plenty of complexity to deal with even if
we ignore the possibility that hearers mishear, or have false beliefs about, or are
uncertain about, the value of contextual parameters.
Information states will be modeled, as in Veltman (1996)’s update semantics,
as sets of possible worlds. That is, the information state of agent Ad is the set of
possible worlds that are compatible with what Ad believes. Then, we can model
information gain as the removal of worlds from an information state. Learning that
p is modeled by removing all those worlds from the information state that do not
make p true. With this, our schema becomes:
(2.5)
new info state =
old info state X tw | Sp uttered It is raining in Chicago in wu
Given this general possible-worlds setup, it makes sense to identify the informational content of It is raining in Chicago with a set of possible worlds, as well.
Then we can contrast the dynamic pragmatics schema in (2.5) with the dynamic
semantic schema in (2.6) (this is just the update that happens in Veltman (1996)’s
system for non-modal sentences):
(2.6)
new info state =
old info state X tw | It is raining in Chicago in wu
We can think of (2.5) as a ‘cautious update’ and (2.6) as a ‘credulous update’: With
(2.6), we model something that happens if an addressee observes an utterance of
It is raining in Chicago and believes it to be true. With (2.5), by contrast, we
model what happens whenever an addressee observes an utterance of It is raining
in Chicago: The addressee automatically adds to his information state is what he
can be certain about, viz., that the utterance he just observed happened.
CHAPTER 2. THE BASIC SYSTEM
14
Much of this chapter is dedicated to further spelling this out formally. In order
to do so, we must specify, for example, what it is for an utterance to occur at a
world. Before doing so (in Section 2.3), I want to briefly preview how to derive the
inference to the truth of the content.
2.2
Communicating contents
We ultimately want to model how it is that a speaker comes to believe in the truth
of the content of an observed utterance. One way to do this would be to simply
assume that two updates happen: The one in (2.5) and the one in (2.6). This is the
idea in much work building the conception of assertion in Stalnaker (1974, 1978,
1994, 2002): When an utterance like It is raining in Chicago happens, the update
in (2.5) is performed automatically, and at the same time, the speaker proposes to
also perform the update in (2.6).
This is not what is intended here: The only update that happens when a speaker
makes an utterance is the one in (2.5). But this update, will, if the context is right,
also inform the addressee that the uttered sentence is true. That is, in certain
contexts, the content of an utterance will be a contextual entailment of the fact that
the utterance was made. In which contexts? The typical case is one in which the
addressee takes the speaker to be both honest and well-informed with respect to
the topic of his utterance.
If the addressee takes the speaker to be honest, (2.7) will be true in all worlds in
his information state (where ñ is material implication):
(2.7)
(Sp utters It is raining in Chicago) ñ (Sp believes it is raining in Chicago)
And if the addressee takes the speaker to be well-informed, all worlds in his
information state will also make true statement in (2.8):
(2.8)
(S believes it is raining in Chicago) ñ (It is raining in Chicago)
But if both (2.7) and (2.8) is true in a given world, so is (2.9):
CHAPTER 2. THE BASIC SYSTEM
(2.9)
15
(S utters It is raining in Chicago) ñ (It is raining in Chicago)
This means that the worlds in the addressee’s information state fall into one of two
classes:
1. Worlds in which the speaker utters It is raining in Chicago and it is raining
in Chicago.
2. Worlds in which the speaker does not utter It is raining in Chicago.
The addressee observes the speaker’s utterance of It is raining in Chicago. That
is, we remove all worlds from the information state in which the speaker has not
made this utterance.
But that means that we are left with an information state that only contains
worlds in the first class—and these are all worlds in which it is raining in Chicago.
That is, in virtue of his prior belief that the speaker is honest and knowledgeable
about rain in Chicago, learning that the speaker uttered It is raining in Chicago
also means, automatically, learning that it is raining in Chicago. If the addressee’s
pre-utterance belief state is of the right kind, then performing the update in (2.5)
will also eliminate all worlds in which it is not raining in Chicago.
2.3
The basic system
In the present section, I will set up the basic ingredients for framework of dynamic
pragmatics. I will only introduce the ingredients necessary to model the simple
inference to the truth of the content that I just summarized informally. Chapter 5
will extend this basic system into one in which more pragmatic phenomena can be
treated.
CHAPTER 2. THE BASIC SYSTEM
2.3.1
16
Languages
The object language Prop
To keep things maximally simple, the system I set up here will model pragmatic
inference in a community of speakers of classical propositional logic. This has
the advantage that we will not have to worry whether the semantics given for
the language is adequate to any particular group of speakers, as, by assumption,
our fictional speech community uses the language with its classical semantics. In
Chapter 3 and Chapter 7, I will very minimally extend the object language, but its
basic propositional character will remain the same.
The formal definition of the language Prop and its semantics is given in Appendix A. The semantics is an entirely standard intensional one: Proposition letters are interpreted as a set of possible worlds, conjunction is interpreted as set
intersection, negation as set complement, and so forth.
The pragmatic language PProp
Given an object language L, we define a corresponding Pragmatic Language PL .
The pragmatic language PProp , by necessity, much richer than our Prop itself. We
want to be able to talk about agents (speakers, addressees) and other individuals,
and their beliefs. Further, we want to be able to talk about time and we want to be
able to quantify over individuals and times.
To distinguish formulas of the object language and the pragmatic language,
I will use indexed versions of ϕ as meta-variables for object language formulas,
and indexed versions of φ and ψ as meta-variables for formulas of the pragmatic
language.
I will use i, i1 , i2 etc. to refer to agents (interlocutors) and t, t1 etc. to refer to times.
The pragmatic language is a multimodal language whose modalities are indexed
to individuals and times. Thus, to represent the fact that i believes φ at time t, we
write:
CHAPTER 2. THE BASIC SYSTEM
(2.10)
17
i,t φ
‘At time t, agent i believes that φ.’
PProp also shares its propositional constants with with Prop, and it has the same
propositional connectives. As a result, Prop is a sublanguage of PProp , and we can
write things like:
(2.11)
i,t p p ^ q)
‘At t, i believes that p is false and q is true.’
And we have the usual first-order quantification over individuals and times, so we
can represent:
(2.12)
a.
Dx : x,t p
‘There is an agent that believes, at t, that p.’
b.
Dt : i,t p
‘There is a time at which i believes that p.’
c.
Dx : x,t D y : y,t p
‘There is an agent that believes that there is another agent who believes that p’
2.3.2
Models for time and belief
In Section 2.2, we conceived of information states as sets of possible worlds. In
order to talk about beliefs and time, we will conceive of the worlds in an information
state as being drawn from a background model that has some additional structure.
I will introduce the background models first and then, in Section 2.3.5, show how
we can define dynamic belief states in terms of them. The result will be that we
take a dynamic perspective on the static background models.
CHAPTER 2. THE BASIC SYSTEM
18
Figure 2.1: A branching-time model (from Condoravdi 2002)
Forward-branching time
The formulas of PProp are interpreted on ‘forward-branching’ models of the T ˆ Wtype (Thomason 1984). Models like this have been fruitfully employed in the
linguistics literature to capture the interaction of tense, modality and conditionals
by Condoravdi (2002), Kaufmann (2005) and Kaufmann and Schwager (2009),
among others. Forward-branching models capture the intuition that the past of a
world is fixed, but its future is not. At every given time, there are multiple ways the
world could develop in the future. The structure of the models can be visualized
as in Fig. 2.1.
This forward-branching structure is constructed in the following way. The set
T of times is linearly ordered. Worlds are taken to determine complete courses
of affairs at all times. Frames determine a time-indexed equivalence relation «t
between worlds. w1 «t w2 intuitively means ‘w1 and w2 share their history at least
up to time t’. For a given world w, the equivalence class tv P W | w «t vu can be
thought of the set of possible futures of w at time t—that is, at t, all worlds in this set
are ‘the same’, the worlds in the equivalence class constitute all the ways how the
world could evolve after t.
CHAPTER 2. THE BASIC SYSTEM
19
To obtain the frame in Fig. 2.1, we let
(2.13)
a.
W “ tw1 , w2 , w3 , w4 , w5 u
b.
T “ tt1 , t2 , t3 u
c.
t1 ă t2 ă t3
d.
w1 «t1 w2 «t1 w3 «t1 w4 «t1 w5
w2 «t2 w3 «t2 w4
«t3 “ H
Thomason’s definition of T ˆ W-frames is in Definition 4 in Appendix A. Throughout this dissertation, I will assume that time is discrete, and in fact that T is just
the set of natural numbers with its usual ordering. This is sufficient for present
purposes and has the advantage that we can refer to ‘the time just after t’ as t ` 1.
I do not assume, however, that all branches have a common root—i.e., that «0 is
total, as Fig. 2.1 might suggest.
Beliefs
In order to interpret the belief-modalities i,t , we extend the basic TˆW-models with
accessibility relations of the kind familiar from the relational models introduced by
Kanger (1957), Hintikka (1961) and Kripke (1963). For given i, t, Ri,t is a relation on
W that is transitive, Euclidean and serial.1 Modal operators are interpreted using
this relation in the usual way, resulting in an SD45-logic for i,t . In particular this
ensures the usual introspection properties:
(2.14)
a.
Positive Introspection
w i,t pφq ñ w i,t pi,t φq
b.
Negative Introspection
w
1
i,t pφq ñ w i,t p i,t φq
A relation on a set X is transitive if xRy and yRz implies xRz, it is Euclidean if xRy and xRz imply
yRz and it is serial for all x P X there is a y such that xRy.
CHAPTER 2. THE BASIC SYSTEM
20
Variables and constants are interpreted rigidly, i.e., independently from the world.
In order to ensure that beliefs respect the structure of the branching-time model,
we impose two constraints:2
(2.15)
Historicity constraint
If w1 «t w2 , then w1 Ri,t v iff w2 Ri,t v
(2.16)
No fore-belief constraint
If v1 «t v2 , then wRi,t v1 iff wRi,t v2
(2.15) requires that the belief relations respect «, i.e., if two worlds have not divided
at t, an agent must have the same beliefs in both worlds. (2.16) captures the idea
that, if two worlds are undivided at t, then it is ‘objectively unsettled’ which one of
the two will become actual—in that case, an agent should either take both worlds
to be possible, or neither. This is a very intuitive constraint as long as we think of
the objective uncertainty to be due to ‘external factors’. In Chapter 5, we will see
that we have to weaken it in the context of action choice.
We also introduce an operator for common belief at a time Ct , which is interpreted using the transitive closure of the individual belief relations at t:
(2.17)
Ct φ
‘At t, it is common belief that φ.’
As usual, this means that φ is common belief iff everyone believes that φ, everyone believes that everyone believes that φ, everyone believes that everyone that
everyone believes that φ, and so on.
2.3.3
Events
The final ingredient in the basic system is the notion of an event. We introduce
these in two steps. First, we define a set of event types, which are model-theoretic
2
The name ‘historicity’ for (2.15) is taken from Kaufmann (2005). He calls (2.16) ‘No foreknowledge’. The full constraints on accessibility relations are summarized in Definition 9 in Appendix A.
CHAPTER 2. THE BASIC SYSTEM
21
entities standing for classes of events with fixed participants. Then, we define a
function that specifies which event type is instantiated in a given world, at a given
time, and introduce event predicates into PProp that allow us to talk about events in
the language.
Event types
I assume a stock of event classes E. Throughout this dissertation, we will be largely
using only one such class, utter. Event types are composed of such an event class,
together with a set of arguments (participants). utter takes three arguments, two
individuals and a formula of the object language, forming event types such as:
(2.18)
utterpi1 , i2 , ϕq
‘an event of i1 uttering ϕ towards i2 ’
Note that even though these event types have the form of atomic predications, they
are not formulas of PProp . Instead, they are model-theoretic entities. The individual
‘arguments’ are not constants or variables referring to entities, but the entities
themselves. This means that the formulas of Prop are also part of the models for
PProp , about which I will have more to say below.
Event instantiation
Models for PProp contain a partial function Hap (for ‘happens’), that specifies which
event types are instantiated at a world and time. For a given event type Epa1 , . . . , an q,
Hapw pt1 , t2 q “ Epa1 , . . . , an q intuitively stands for ‘Epa1 , . . . , an q happens between t1
and t2 ’.
Assuming Hap is a function is a useful simplification. To simplify even further,
I restrict attention to models in which Hapw pt, t1 q is only defined if t1 “ t ` 1—
this means that events only happen ‘between’ two adjacent time-steps (i.e., events
are instantaneous), and there is at most one event happening at any given time.3
3
These assumptions that could be lifted, but they will make the subsequent definitions run much
more smoothly. Of course, in a general setting, we want to allow for events to overlap temporally,
CHAPTER 2. THE BASIC SYSTEM
22
Further, I impose the following two more constraints on Hap. The first is a minimal
requirement, ensuring that Hap respects the temporal structure of our models, the
second is a useful simplification.
(2.19)
Historicity of Hap constraint
If w1 «t w2 then for all t1 , t2 ď t : Hapw1 pt1 , t2 q “ Hapw2 pt1 , t2 q
(2.20)
Determinism constraint
If w1 «t w2 and Hapw1 pt, t ` 1q “ Hapw2 pt, t ` 1q then w1 «t`1 w2 .
(2.19) ensures that if an event type is instantiated in one world, but not the other,
the two worlds diverge. (2.20) requires the converse, and is arguably conceptually
dubious.
According to (2.20), worlds only divide if different event types are instantiated in
them—which means that the type of an event fully determines the effect the event
has in a given world. This is likely too strong (if not for conceptual reasons, then
because the event types that we will be using are too ‘coarse’), but in the present
context, the assumption does no harm, and it makes it a little easier to intuitively
grasp the shape of our models. Given an initial specification of «0 (for the first time
step 0), « is fully determined by Hap.
Event predicates
We have introduced event types in our models, as well as a function that specifies
when an event happens. But so far, we cannot yet talk about events in the language
PProp . In order to do so, we assume the language contains event predicate letters,
which are standard predicate letters, indexed with event constants or variables.
For what is to follow, we again need only one such predicate letter, utter. Instead
of interpreting it via an interpretation function that is then suitably constrained to
give utter its intended interpretation, I state the satisfaction conditions of utterformulas directly:
and allow for non-instantaneous events so as to reason about what happens while an event is
underway, but neither will be necessary for what is to follow.
CHAPTER 2. THE BASIC SYSTEM
(2.21)
23
w uttere pa, b, pϕqq iff Hapw pIpeqq “ utterpIpaq, Ipbq, ϕq
Note that one of the arguments of utter contains a formula of Prop. Implicitly, this
definition treats p¨q as an operator that takes a Prop-formula (which, as indicated
above, are part of our models) into a name that refers to this formula.4 Much like
with parentheses, I will often omit the p¨q operator from example formulas, unless
it aids readability.
Note that the symbol utter is overloaded—it figures both in event types and in
event predications. Similarly, it will frequently be useful to sloppily use a single
symbol to refer to an individual in a model and a constant of the language that
refers to this individual—e.g., I will often use Sp to refer to the speaker-as-agent in
the model, but also use Sp as a constant that implicitly is assumed to refer to the
model-theoretic entity Sp. I trust no confusion will arise from this.
Finally, I define two more versions of utter which take fewer arguments. The
other arguments are existentially closed: uttere pi, pϕqq is true if e is an utterance of
ϕ by i towards an arbitrary agent, uttere piq is true if e is an utterances by i of an
arbitrary sentence towards an arbitrary agent.
2.3.4
Constraining belief change
So far, I have not said anything substantial about how beliefs evolve over time. I
specified some general constraints requiring that the basic structure of the underlying frame is respected, but nothing prevents an agent from having wildly different
beliefs at adjacent times t and t1 . We remedy this in the present section.
Ensuring utterance uptake
It is here where we encode, in terms of a constraint on admissible models, the
idealizing assumption that agents perfectly observe each other’s utterances.
4
Potts (2005, Appendix A.3) makes essentially the same move—his logic LU , which in many
ways can be seen as similar to P also contains constants that refer to sentences of his ‘object’ language
LCI .
CHAPTER 2. THE BASIC SYSTEM
(2.22)
24
PAL Constraint
For all i, w, w1 , t :
if there is v such that wRi,t v and Hapw pt, t ` 1q “ Hapv pt, t ` 1q, then
wRi,t`1 w1 iff wRi,t w1 and Hapw pt, t ` 1q “ Hapw1 pt, t ` 1q
This formulation is quite complex, because we have to take into account that an
agent can have false beliefs, and hence can observe events that he took to be
impossible. If we excluded that possibility (or did not require the belief relations
to be serial), we could impose the simpler constraint in (2.23):
(2.23)
Ri,t`1 “ thw, vi P Ri,t | Hapw pt, t ` 1q “ Hapv pt, t ` 1q u
What (2.23) says is that when an agent observes an event, this eliminates all links
from his belief relation that point to worlds in which the event does not happen.
(2.22) imposes the same requirement, except that it leaves unconstrained what
happens if an agent observes an event he antecedently believed would not occur.
In full generality, we would want to impose additional constraints to cover this
case, which ensure that the necessary revision of beliefs is minimal. How to achieve
this is a difficult topic, and it is orthogonal to the concerns in this dissertation,
and so I leave belief revision unconstrained, except for assuming that the result of
revision ensures that the agent believes that the event he just observed happened.
(2.22) does two things: (i) it ensures that beliefs only change in light of events5
and (ii) in ensures that all agents have perfect knowledge about all events that
occur. Given that this is globally the case throughout our models, whenever
an event happens, it will thereafter be common knowledge that it did, and events
happening is the only way in which knowledge changes. As a consequence, all
acquired knowledge is shared knowledge. This is surely inadequate in general—
people can and do learn things privately, and pass on this information later, and
that needs to ultimately be reflected in a theory of pragmatics.
5
With one exception: If an agent was certain that an event would occur in the next time step, but
none occurs, this will trigger belief revision.
CHAPTER 2. THE BASIC SYSTEM
25
However, for present purposes, this conception will be sufficient. Recall that
the events we are mainly interested in are utterance events happening in the course
of a conversation. If we assume that all agents are party to the conversation,
assuming that they have common knowledge of all utterance events is not so crazy.
Essentially, the idealizing assumption we are making with (2.22) is that all events
are local to the interlocutors, that all interlocutors (correctly) observe all events and
that they also observe each other observing the events.
The connection to Public Announcement Logic (PAL)
Modeling information gain as the removal of links from an accessibility relation
is not a novel idea at all. It goes back at least to Landman (1986). Semantics for
Public Announcement Logic (PAL)6 often make use of the idea. Indeed, putting
the constraint in (2.22) on our models means that we can construe our timelines
(modulo belief revision) as sequences of PAL models where the model at each time
step is derived from the previous one by the ‘public announcement’ that the event
happened.
In PAL, updates can be made with arbitrary formulas, including those that contain knowledge (or belief) operators.7 In the present system, ‘updates’ only happen
with event formulas. And the conceptualization is quite different. In PAL, ‘announcements’ are classically construed as utterances that provide the information
that their content is true. This makes PAL quite limited in modeling conversation
in general, as it essentially presupposes that all utterances are honest and based on
perfect information.
But in our present system, we do not have ‘public announcements’ of the content
of utterances, but rather of the fact that an utterance happened—that is we update
with utterpa, ϕq, not with ϕ. Instead of ‘public announcement’, we should instead
6
More precisely: Those semantics that are not stated as a special case of a more general framework, such as Dynamic Epistemic Logic (DEL), that is, cf. van Ditmarsch, van der Hoek and Kooi
(2008).
7
That is why in PAL, it is not generally ensured that φ is commonly known after a public
announcement, i.e., rφsCt φ is not a theorem of PAL—it is invalidated in cases like φ “ p ^ p (‘You
don’t know yet that p’)—but we disallow such updates, as we update only with event expressions,
not with arbitrary formulas.
CHAPTER 2. THE BASIC SYSTEM
26
perhaps speak of ‘joint observation’. But the logic remains essentially the same.
Such a conception of PAL is not without precedence in the literature:
“So far, all announcements were made by an agent that was also modelled in the system. We can also imagine an announcement as a ‘public
event’ that does not involve an agent. Such an event publicly ‘reveals’
the truth of the announced formula. Therefore, announcements have
also been called ‘revelations’—announcements by the divine agent, that
are obviously true without questioning.”
(van Ditmarsch et al. 2008, p.76)
Except that in the present case, where ‘announcements’ only concern publiclyobservable events, we need not invoke a divine announcer—we can replace him
by the agent’s ability to observe (and observe each other observing). The tight
relation between the way beliefs change in the current model and PAL suggests that
generalizations of PAL are useful resources for lifting the simplifying assumptions
about perfect uptake made here. I return to this issue in Chapter 10.
In summary, the PAL constraint ensures that, if an event happens, all agents
learn that it did, and this is the only information they acquire. Given that this is
common knowledge in our models, that also means that when an event happens,
it becomes common knowledge that it did.
2.3.5
The dynamic perspective
So far, the system I have introduced is not quite dynamic. It is a model of how
beliefs change, to be sure, but our models themselves are entirely static. The
overall T ˆ W model is fixed once and for all. It takes “the view ‘from above’,
viewing epistemic temporal models as a Grand Stage where events unfold” (van
Benthem and Pacuit 2006, p. 88).
By contrast, the ‘preview’ I gave in Section 2.2 was phrased in entirely dynamic
terms—taking, in van Benthem and Pacuit’s terms, the ‘view from below’. In the
CHAPTER 2. THE BASIC SYSTEM
27
present section, I introduce a way to take a ‘local’ perspective on the ‘global’ Grand
Stage model.
The Grand Stage model is useful to set up the system, to define constraints
regulating how beliefs should evolve, and to specify which non-epistemic consequences events have on the world (which we will do in Chapter 5). But the local,
‘dynamic’ perspective is easier to handle if we want to talk about what happens at
a particular time step. This is why, in the later parts of the thesis, when we apply
the model developed here and in Chapter 5, we will almost exclusively take the
dynamic perspective. For a fixed PProp -model, we define:
(2.24)
Bi,t,w :“ tv P W | wRi,t vu
‘the belief state of i at t, w’
Bi,t,w simply collects the worlds that are compatible with what i believes (in world
w, at time t). For such belief states, we then define the update operation r¨s:
(2.25)
Let ev be an event formula, then
Bi,t,w revs :“ tv P W | v P Bi,t`1,w and Hapv pt, t ` 1q “ ev u
‘the belief state that i would be in if event ev happened just after t’
If the (immediate) occurrence of ev is not ruled out8 in Bi,t,w , this is obviously
equivalent (given the PAL constraint above) to:
(2.26)
Bi,t,w revs “ tv P Bi,t,w | Hapv pt, t ` 1q “ ev u
(2.26) is essentially the update familiar from Veltman (1996)’s update semantics,
except that we don’t update with formulas of the language, but with event type
formulas. We define a notion of support for information states, as follows:9
8
In case ev was believed to be impossible, the right-hand side of (2.26) is empty, but the same is
not true for Bi,t,w revs, which in this case is derived by belief revision.
9
Veltman ultimately defines support in terms of vacuous update:
(i)
B φ iff Brφs “ B
CHAPTER 2. THE BASIC SYSTEM
(2.27)
28
For any information state B and PProp formula φ:
B φ iff for all v P B : v φ
With this, we obviously have:
(2.28)
Bi,t,w φ iff w i,t φ
We can thus think of the dynamic perspective as ‘zooming in’ to (an agent’s perspective at) a given world and time in the T ˆ W model. The benefit of having a
background model is that we immediately obtain results like the following:
(2.29)
For any two agents i, i1 , we have, at all t, w:
Bi,t,w rutterpa, b, ϕqs i1 ,t`1 uttert pa, b, ϕq
And the stronger:
(2.30)
Bi,t,w rutterpa, b, ϕqs Ct`1 uttert pa, b, ϕq
The benefit of the dynamic perspective, by contrast, is that it is much easier to
handle, and more perspicuous, than the sometimes unwieldy background model.
To state (2.29) in terms of the background model, for example, we would have to
say instead:
(2.31)
For all i, i1 , t, w: If Hapw pt, t ` 1q “ utterpa, b, ϕq, then
w i,t`1 i1 ,t`1 uttert pa, b, ϕq
(2.29) makes much easier to see, at a glance, what is going on. We thus can think
of the dynamic perspective as an interface to the global model. But it is also more
than just an interface, as it stresses the dynamic nature of what we are modeling.
In many cases when I use the dynamic perspective in the rest of this dissertation,
I will actually suppress all temporal parameters (and, when appropriate, also the
world parameters). Often, these will be irrelevant because throughout this thesis,
We do not have this option here, as we do not update with formulas, but with event expressions.
Luckily, since all formulas can be evaluated pointwise, we loose nothing by defining support in the
distributive manner done in (2.27).
CHAPTER 2. THE BASIC SYSTEM
29
I will be concerned with one-shot utterances. The basic set-up of the model has
obvious potential to be extended so as to allow reasoning about discourse planning,
involving multiple acts in succession, but we will not deal with such things here.
Even when the value of the temporal indexes are relevant, there intended values
will often be clear from the context. This suggests that, ultimately, we could
simplify the dynamic perspective further, by having it use a simplified version of
the full pragmatic language without time parameters. However, for the space of
this thesis, I shall simply leave these parameters implicit wherever possible. If we
do so, then (2.29) becomes the very readable (2.32):
(2.32)
For any two agents i, i1 , we have:
Bi rutterpa, b, ϕqs i1 putterpa, b, ϕqq
2.4
Communicating contents in the basic system
We have just seen a very general result about updates:
(2.33)
Bi,t,w rutterpa, b, ϕqs Ct`1 uttert pa, b, ϕq
(2.33) holds for all i, t, w in all models for PProp . Actually, this is about the strongest
thing we can say about the general case. But we can now make contextual assumptions that allow us to derive more interesting things. In particular, we can
straightforwardly capture the way in which an addressee comes to believe in the
truth of what a speaker says, if he takes the speaker to be well-informed and honest
(I drop the world parameter in what follows).
(2.34)
Contextual assumptions: Trusting addressee
a.
‘Honest speaker’
BAd,t uttert pSp, ϕq ñ Sp,t pϕq
b.
‘Informed speaker’
BAd,t Sp,t pϕq ñ ϕ
CHAPTER 2. THE BASIC SYSTEM
30
If (2.34a) is true, then so is (2.35):
BAd,t rutterpSp, Ad, ϕqs Sp,t pϕq
(2.35)
And if, in addition, (2.34b) is true, so is (2.36):
BAd,t rutterpSp, ϕqs ϕ
(2.36)
So we capture the fact that if a speaker is taken to be honest and knowledgeable,
then learning that he uttered ϕ amounts to learning that ϕ is true. Surely, this is
not a great achievement of our system, but we now have a very general system
in place for reasoning about utterance events and their (epistemic) consequences.
That it validates such pervasive inference is reassuring.
Actually, the system gives us a little more. Suppose the conditions in (2.35) are
believed to be true by the speaker about the addressee’s belief state.10 Then we also
have:
BSp,t rutterpSp, ϕqs Ad,t`1 ϕ
(2.37)
That is, if a speaker thinks his addressee takes him to be honest and knowledgeable,
he will also believe that if he utters ϕ, his addressee will come to believe ϕ, too.
That is, we can model the fact that the speaker is aware of the epistemic consequences
of his potential actions.
In Chapter 5, we will see how this gives us a way to reason about utterance
choice. Suppose a speaker has to decide whether to utter p, q, p _ q or p ^ q—then
he has to compare the expected consequences of doing each. That is, he has to
compare the information states in (2.38) with respect to how well they satisfy his
goals.
10
(i)
That is, we have
a.
b.
BSp,t Ad,t puttert pSp, ϕq ñ Sp,t ϕq
BSp,t Ad ppSp,t ϕq ñ ϕq
CHAPTER 2. THE BASIC SYSTEM
(2.38)
a.
BSp,t rutterpSp, pqs
b.
BSp,t rutterpSp, qqs
c.
BSp,t rutterpSp, p _ qqs
d.
BSp,t rutterpSp, p ^ qqs
31
In order to do this, of course, we need to be able to talk about goals and their
satisfaction. That is what Chapter 5 will be all about.
2.5
Communicating with and without intentions
An interesting fact about the way we just derived (2.36) is that we did not make any
references to the speaker’s intentions or desires. In particular, we did not assume
that the speaker intended or desired to communicate anything—we just assumed
that the addressee believed that the speaker is honest and well-informed.
One might question, then, whether what I have described above is truly
communication—it is an instance of information transfer (at least if we assume that
the speaker indeed believes what he said), but for communication we might want to
require that the speaker also intends that the addressee form the belief in question.11
Even this is not always taken to be sufficient for true communication (or at
least successful communication) to occur. Bach and Harnish (1979) maintain that
communication is successful only if the addressee also recognizes the speaker’s
intention to make him believe something.12 But on a certain level of description,
there is little difference between situations in which an addressee comes to correctly
recognize the intention of the speaker to communicate something, and those in
11
Though English communicate at least has one sense on which this is not required, as attested
by examples like (i).
(i)
In some false confession cases, details of the crime are inadvertently communicated to a
suspect by police during questioning.
(http://www.innocenceproject.org/fix/False-Confessions.php, last retrieved on August 21, 2013)
12
In fact, this is more the view of Grice (1957)—I am not sure what exactly the intention is
that Bach and Harnish (1979) claim the hearer must recognize for communication to succeed—see
Section 4.3.2 for discussion.
CHAPTER 2. THE BASIC SYSTEM
32
which he otherwise believes that the speaker would only utter sentences he believes
to be true.
Of course, this assumption may be justified, on a particular occasion, by the
following two beliefs:
(2.39)
a.
The speaker will utter ϕ only if he intends me to believe it.
b.
The speaker would not intend me to believe anything he does not
believe himself.
(2.39a) is true and believed by the addressee, then observing an utterance will
lead the addressee to recognize that the speaker intends to communicate ϕ—but
even this is not necessarily enough for a communicative intention à la Bach and
Harnish to succeed—for, they say, a communicative intention is also ‘intended to
be recognized as intended to be recognized’. So it is not enough if an addressee
recognizes the speaker’s intention to make him believe something—he must also
recognize that the speaker intended him to recognize the intention.
At this point we start to wonder whether all this intention recognition is relevant
in everyday linguistic practice. Suppose a speaker wants to pass on a certain piece
of information—say, that it is raining in Chicago—to someone else. He utters a
sentence—say, It is raining in Chicago—and the addressee takes him to be honest
and well-informed, and so he comes to believe that it is raining in Chicago. Does it
matter, to the speaker or the addressee, why the addressee was predisposed to accept
the speaker’s utterance as truth? Has communication failed, in an interesting sense,
if the addressee fails to also believe that the speaker intended him to recognize that
he intended him to recognize that he intended him to believe that it is raining in
Chicago?
2.6
Conclusion
In this chapter, I have introduced a number of the foundational notions necessary
for the framework of dynamic pragmatics: a formal system that allows us to talk
CHAPTER 2. THE BASIC SYSTEM
33
about beliefs and how they develop over time and about the occurrence of events
and how they influence agents’ beliefs. A particular subset of events, utterance
events are of particular importance, as one would expect from a system that is
intended for linguistic pragmatics. I have then illustrated the basic workings of the
system by showing how we can model a pragmatic inference that other systems—
in particular many systems of dynamic semantics—take for granted: I have shown
how the model accounts for the fact that an addressee who takes the speaker to be
honest and knowledgeable can come to believe in the informational content of an
utterance he observes.
Before moving on, I want to briefly stop and reflect on a particular feature of
the system as set up so far. The information that gets added to our agents’ belief
states in the course of a conversation is, in some sense, not particularly linguistic
in nature. It is simply the fact that an event happened. On a very basic level,
utterance events are on a par with non-linguistic and non-communicative events.
Thus, to use Stalnaker’s oft-cited example, basic gain in information by observing
an utterance event is on a par with the basic gain in information by observing that
a goat just walked in.
Let us dwell on this for a second. Our event types could include, for example, an
event donkey walks inpdq13 and if we introduce the corresponding predicates, whose
interpretation is suitably specified, we could derive facts as such as the following:
(2.40)
Bi,t,w rdonkey walks inpdqs Dx : donkeypxq
‘If i observes a donkey walking in, he comes to believe that a donkey
exists.’
This may give one pause. Is the system of dynamic pragmatics supposed to model
every aspect of our cognitive experience of the world? Does a full specification of
the model require an account of every kind of cognitive effect of an observation?
This seems crazy, and it would be.
13
This is a particularly weird representation for such an event, of course—but this is part of my
point here: The current system, as a system of dynamic pragmatics, is not intended to give plausible
representations of arbitrary events. It is interested in giving plausible representations of utterance
events.
CHAPTER 2. THE BASIC SYSTEM
34
On a very abstract level, the system of dynamic pragmatics indeed can be seen
as a model of any kind of belief formation in response to events. I think this is
appropriate—it embodies the very Gricean assumption that pragmatic reasoning
is continuous, at least in principle, with other kinds of reasoning.
But all that means is that the system of dynamic pragmatics, as set up so far,
could, in principle, be extended to model non-linguistic reasoning. A non-linguist
could appropriate the model, adding his own event types and predicates and
suitably constraining their interpretation to model reasoning about non-linguistic
events. But in doing so, he would have to extend the system, making assumptions that are appropriate to the domain he is modeling. Take the example in
(2.40): In assuming that there is an event type such as donkey walks inpdq (which
gets perfectly observed), we have to assume that our agents have the concept of a
donkey, and that they can (perfectly) distinguish donkeys from non-donkeys, that
their concept of donkeys is compatible with donkeys changing location, that they
have a concept of ‘coming in’, and so forth. These assumptions may be reasonable (as idealizations) for certain purposes, but they are completely non-linguistic
assumptions, which is why I have not made them.14
The assumptions that I have made instead—in assuming that there is an event
type utter, and that there is a corresponding event predicate, and the way it relates to
the event type—are thoroughly linguistic. I have assumed that there are such things
as utterances of sentences, and that agents recognize them as such. And the idealization I have made in assuming that agents observe utterances of disambiguated
logical forms amounts to the assumption of (idealized) linguistic competence: I
assume not only that agents recognize utterances as such, but also that they have
the appropriate competence to analyze the phonological, morphological, syntactic
and semantic structure of the utterance so as to assign truth conditions to it.
Much of the rest of this dissertation will be concerned with the question of what
other linguistic knowledge we need to assume our agents to have beyond this.
14
Of course, a linguist might sometimes make some of these assumptions, for example when
studying the semantics of prepositions like in.
Chapter 3
Clause types
3.1
Introducing clause types
For the sake of simplicity, the object language of the system PProp is just that of
propositional logic. This choice conveniently ignores much of the complexity of
actual natural languages, allowing us to focus on pragmatic concerns. There are
many pragmatic phenomena, of course, that crucially involve the richer resources
of natural languages, but to get a handle on some fundamental issues, it will be
useful to focus on issues that arise even for very simple languages.
There is one way, however, in which natural languages are richer than Prop
that we need to take into consideration: Sentences in natural languages come in
various types, such as declaratives, interrogatives and imperatives, and the type of
a sentence greatly influences how the sentence can be used, and what pragmatic
inferences its use supports.1
(3.1)
a.
You will come tomorrow.
b.
Will you come tomorrow?
c.
Come tomorrow!
(declarative)
(interrogative)
(imperative)
1
At present, I limit the discussion to what Sadock and Zwicky (1985) call the ‘major sentence
types’, declaratives, interrogatives and imperatives. In Chapter 8, I will extend the discussion to
include exclamatives.
35
CHAPTER 3. CLAUSE TYPES
36
On some level of description, all the sentences in (3.1) involve the same content,
which can be characterized as the proposition that is true iff the addressee will come
tomorrow. However, utterances of these sentences employ this proposition in very
different ways: An utterance of (3.1a) intuitively claims that this proposition is true,
while an utterance of (3.1b) inquires whether the proposition is true and an utterance
of (3.1c) attempts to induce the addressee to make the proposition true. A central
question for a pragmatic theory to answer is how this difference comes about, and
more generally how content, sentence type and context interact in language use.
To this end, we define a minimal extension of the object language Prop, as in
(3.2).
(3.2)
Sen :“ L$ Y L? Y L! , where
ˇ
(
a. L$ :“ $ φ ˇ φ P Prop
ˇ
(
b. L? :“ ?φ ˇ φ P Prop
ˇ
(
c. L! :“ !φ ˇ φ P Prop
The question at hand now is how utterances of sentences in L$ differ from utterances of sentences in L? , and how both differ from utterances of sentences in
L! . The answer to this question should tell us why sentences in L$ are typically
used for informing or claiming (and in which contexts they can be used for other
purposes), while sentences in L? are typically used to request information (and
in which contexts they can be used for other purposes), while sentences in L! are
typically used in an attempt to get the addressee to do something (and in which
contexts they can be used for other purposes).
At the end of Chapter 2, when showing how information can be conveyed
by means of a utterance (which was implicitly taken to be the utterance of a
declarative), I used the assumption that the addressee believes that the speaker
would only utter a declarative ϕ if he believed ϕ to be true. The question about L$
we are facing now is how to justify this assumption.
That this assumption needs justification, even in cases of perfect cooperation,
may not be obvious: Isn’t it simply more cooperative to speak the truth than to
CHAPTER 3. CLAUSE TYPES
37
speak falsity? Grice (1975) certainly thought so: His Maxim of Quality purportedly
is motivated by his Cooperative Principle:
(3.3)
Maxim of Quality
(Grice 1975, p. 46)
Try to make your contribution one that is true.
(3.4)
1.
Do not say what you believe to be false.
2.
Do not say that for which you lack adequate evidence.
Cooperative Principle
(Grice 1975, p. 45)
Make your conversational contribution such as is required, at the stage at
which it occurs, by the accepted purpose or direction of the talk exchange
in which you are engaged.
Of course (3.3) does not follow from (3.4) in any direct fashion. And it should not,
for (3.3) only is intuitively valid for declaratives, but not for interrogatives and
imperatives. So, the maxim should really be formulated as in (3.5).
(3.5)
Maxim of Quality for Declaratives
Try to utter only those declaratives that are true.
1.
Do not utter declaratives that you believe to be false.
2.
Do not utter declaratives for whose truth you lack adequate evidence.
This new version makes it more obvious that the maxim should not follow from assumptions about cooperative behavior alone. Why should general considerations
about what constitutes cooperative behavior, by themselves, say anything about
sentences of a syntactically individuated class of sentences? There must be some
linguistic property of declaratives, that, together with general considerations of
cooperative behavior, allows us to derive a maxim like (3.5) (in contexts where the
maxim is appropriate).
Something that plausibly could be derived from general considerations about
cooperative behavior is a maxim like the one in (3.6).2
2
A formulation of the quality maxim that is quite close to (3.6) has bee used by Joshi (1982) in
the context of question answering to account for clarification follow-ups.
CHAPTER 3. CLAUSE TYPES
(3.6)
38
Maxim of Quality for Communicated Contents
Do not intentionally make your addressee believe anything that is not true.
1.
Do not intentionally make your addressee believe something that you
believe to be false.
2.
Do not intentionally make your addressee believe something for which
you lack adequate evidence.
If we want to derive (3.5) from (3.6), we need to explain why declaratives generally
have the effect of making the addressee believe their contents—but that is just what
we wanted the Quality maxim to help us explain!
Put differently: Grice’s Quality maxim only sounds like a plausible principle of
cooperative behavior if we assume that declaratives usually have the effect (or are
usually intended to have the effect) of making the addressee believe their content—
but then, this maxim cannot be the reason for this being the typical function of
declaratives.
We could derive similar maxims for interrogatives and imperatives—if we have
some independent means of deriving their (typical) functions:
(3.7)
Maxim of ‘Quality’ for Interrogatives3
Try to utter only those interrogatives that you want your addressee to
answer truthfully.
(3.8)
Maxim of ‘Quality’ for Imperatives
Try to utter only those imperatives that you want your addressee to fulfill.
But again, as maxims of conversation, these principles are not the explanans that
tells us why interrogatives and imperatives have the uses they do. Rather, they
3
(i)
If we were to adopt such a maxim, we might also want to add the following sub-maxim:
Try to utter only those interrogatives that you don’t know the answer to.
Though ((i)) is not appropriate for all utterances of interrogatives—rhetorical questions and exam questions are counter-examples. Similarly (3.7) arguably is not obeyed by certain kinds of combativelyused interrogatives; see Section 6.3.
CHAPTER 3. CLAUSE TYPES
39
are an explanandum: They should follow from general principles of cooperative
behavior, together with what is independently known about interrogatives and
imperatives.
Clearly, it is not just general considerations about cooperative behavior that
allow us to explain why sentences of different types have the uses they have. This
is fairly obvious for non-declaratives, but considering the question in the setting of
multiple clause types makes us realize that it is true for declaratives just as well.
3.2
Denotation type is not a sufficient guide to function
I have introduced the syntax of Sen, but I have not yet specified the semantics of
expressions of the language. In formal semantics, it is general practice to assign
different kinds of denotations to sentences of different types. For declaratives and
interrogatives, this can be motivated on semantic grounds alone, without much
reference to pragmatics: Both declaratives and interrogatives can be embedded in
larger sentences, and the semantics of the larger sentences puts certain constraints
on what the semantics of the embedded sentences can be. So, while it may be
suitable to assume that declarative sentences denote intensional propositions (sets
of possible worlds), we cannot assume that interrogatives denote such propositions: It is hard to see what proposition a wh-question like who will come to the
party could denote so that we can correctly derive the intuitive truth-conditions of
sentences like John wondered who will come to the party and John decided who
will come to the party and We argued about who will come to the party.4
This difference in logical type, however, is not sufficient to explain the differences in use. Some of these difference may be traced to differences in denotation
4
Recent work in the Inquisitive Semantics paradigm (Groenendijk and Roelofsen 2009, Ciardelli
and Roelofsen 2011, a. o., building on the work of Hamblin (1958)) has argued that we can (and
should) assign a uniform semantic type to declaratives and interrogatives—crucially, this value
must be more complex than a flat set of worlds, i.e., more complex than a proposition in the usual
intensional sense. An inquisitive proposition is a set of sets of worlds, which classically is assumed to
be the semantic value only of interrogatives.
CHAPTER 3. CLAUSE TYPES
40
type, but not all. For example, one might reason as follows: If interrogatives denote, say, sets of sets of possible worlds, while declaratives denote simply sets of
possible worlds, maybe this is enough to explain why sentences are used in the
way they are. Indeed, even independently from considerations of formal semantics, there is a strong intuition that the notion of truth does not properly apply to
interrogatives and imperatives. Doesn’t that suffice to explain why the Quality
maxim applies only to declaratives?
Maybe it does,5 but then we only have explained why interrogatives are not
(outside of special contexts) used to make claims or provide information; we have
not yet explained why declaratives typically are. We might venture on, though,
and argue as follows: The denotation type of declaratives simply is something that
is uniquely suited to conveying information and making claims, at least in standard
contexts. After all, one might think, declaratives (or, rather, their denotations), are
something that can be true or false at a world—so of course they are used to claim
that their contents are true, or to express the belief that their contents are true.
We can illustrate this possible line of thinking by means of an analogy: On a
given day, I know that my roommate has visited the German bakery in Palo Alto.
Returning home late at night from a session of dissertation writing, I find a loaf of
delicious bread sitting on the desk in my room. Considering the situation, what I
know about the uses breads can be put to, and about my roommate, I understand
that the bread is a gift, and that he intends me to eat and enjoy it.
Similarly, the story would go, when a speaker utters a declarative, he puts in
front of the addressee a certain semantic object (say, a function of type hs, ti), and
given what the addressee knows about the purposes that this type of semantic
object can be put to, he can reason to the conclusion that the speaker intends to
provide him with some factual information.
But this cannot work, at least if we maintain the assumption that matrix sentences have the same denotation as the sentence would have when embedded.
The reason is this: Given what we know about possible embedding environments,
5
Though note that a set of possible worlds, in itself, has no special relation to the notion of
truth—it acquires such a relation only when paired with a claim or judgement (to use Frege’s term)
that the actual world is contained in the set.
CHAPTER 3. CLAUSE TYPES
41
the semantic value of declaratives (and interrogatives) must be very versatile.
Whatever an embedded declarative denotes, it must be something that can be:
believed, disbelieved, preferred, dispreferred, surprised about, hypothesized, inferred, doubted, and many other things, given that declaratives can be embedded
under predicates such as believe, disbelieve, prefer (want), disprefer, surprised,
hypothesize, infer, doubt and so forth.6 Similar considerations apply to interrogatives: Whatever their denotation type is, it must be something that can be: asked,
known, agreed on, found out, pondered, decided, and many other things in virtue
of the fact that there is a multitude of question-embedding predicates.
To drive this point home: Assume, for example, that declaratives denote sets
of possible worlds. In itself, such a set is no more suited to expressing a belief
that the actual world is a member of the set than for many other purposes. For
example, such a set is equally well-suited for expressing the belief, or conveying
the information, that the actual world is not in this set, or that the speaker desires
the actual world to be in this set, or that he desires it not to be in this set, and many
other things. And the problem is not that sets of possible worlds are the wrong kind
of denotatum: As long was we assume that sentences have the same denotation in
matrix and embedded uses, this problem necessarily arises: A semantic object that
somehow is uniquely suited for expressing beliefs will not do as the denotation of
many embedded declaratives and a semantic object that is suited as a denotation
for all embedded declaratives will not be uniquely suited for expressing beliefs.
Thus, while the study of embedded sentences can tell us something about what
kind of denotation a sentence should have, the nature of this denotation will not
be enough to sufficiently constrain the uses such a sentence can be put to when it is
not embedded. We need something extra in order to explain why matrix sentences
6
I speak here of the denotation being ‘something that can be believed’, etc., which may suggest
that this kind of consideration only applies if we assume ‘static’ objects in the semantics, rather than
the kind of ‘dynamic’ denotations familiar from dynamic semantics. But this is not so: Even if we
take declaratives to denote, say, update relations between information states, in order to account
for the semantics of declarative-embedding predicates, we need to assume that these relations hold
between agents’ belief states, their desire states, and so forth. So a dynamic denotatum, too, must
be very versatile. Too versatile to make providing information or expressing a belief the ‘obvious’
or only thing a speaker could be trying to do with his declarative utterance.
CHAPTER 3. CLAUSE TYPES
42
of a particular type are used the way they are.
3.3
A case for extra-compositional constraints on use
So we need something extra—over and above the denotation sentences of various
types have when they are embedded—in order to explain the constraints on use
that we observe. One way to go would be to put this additional information into
the denotation of matrix sentences—thus giving up the idea that matrix sentences
have the same denotation as their embedded counterparts.
Perhaps the most straightforward way of doing this is to assume that matrix
sentences uniformly contain an operator that scopes over the whole sentence—call
it a ‘force’ operator. We then could hypothesize a different such operator for each
sentence type. This would amount to the claim that natural languages are in fact
structured like Sen, with its three operators $, ? and !, and every matrix sentence
is headed by such an operator. The denotational semantics of these operators then
would be responsible for specifying, or at least constraining, the possible uses of
the sentence.
This kind of analysis has been proposed by Krifka (2001b, to appear), who takes
the force operators to denote speech acts (in particular, Illocutionary Acts in the
sense of Austin (1962), a feature I discuss in Section 3.4).
There is, however, another option, and this is the one I will pursue in the
following: We can specify the context changes brought about by an utterance
directly, as a ‘convention of use’ that is followed by the speech community. On
this construal, we can maintain the assumption that matrix sentences have the
same denotation as embedded sentences. The constraints on use are captured
extra-compositionally, in the conventions of use.
One reason to do things this way is that ultimately, regardless of the denotation
one assigns to matrix sentences, be it static or dynamic, the denotation, on its own,
cannot solve the basic problem. What is asked for is a link between the denotation
on the one hand, and use on the other. Since the denotation is one of the two things
to be linked, it cannot be the link itself. If we assume that matrix sentences denote
CHAPTER 3. CLAUSE TYPES
43
speech acts, for example, we still will need a convention, rule or principle like (3.9).
(3.9)
Performance principle (hypothetical)
If a speaker utters a sentence that denotes a speech act a, a is thereby
performed.
Similarly, if sentences denote some kind of update operation on contexts, we still
would need a convention, rule or principle like (3.10).
(3.10)
Context-change principle (hypothetical)
If a speaker utters a sentence that denotes a context change operation, this
operation is thereby applied to the context.
In either way, we will have to assume something extra to go from what the sentence
denotes to the effect of its use. But once we grant that something is needed in
addition to a compositionally determined denotation—some kind of principle, rule,
or convention that is, by necessity, external to the system of semantic composition—
this additional component may well be the proper place to represent conventional
constraints on use. The principles in (3.9) and (3.10) don’t represent any constraint
of use by themselves—they just enable compositionally-encoded constraints to
have an effect. The alternative is to have multiple principles—one per sentence
type—that specify the dynamic effect of the utterance directly. These conventions
could make reference to the morpho-syntactic characteristics of sentence types
directly (as in (3.11)), or reference a denotation type that is particular to the sentence
type (as in (3.12)).7
(3.11)
a.
Declarative principle (hypothetical)
If a speaker utters a sentence that is morpho-syntactically declarative,
this has the following effect: . . .
b.
7
Interrogative principle (hypothetical)
The conception of clause-typing developed in this dissertation—which builds on Condoravdi
and Lauer (2011, 2012)—is neutral with respect to which of these two (or a number of other) options
is chosen, as I extensively demonstrate in Section 6.4.
CHAPTER 3. CLAUSE TYPES
44
If a speaker utters a sentence that is morpho-syntactically interrogative, this has the following effect: . . . .
(3.12)
c.
...
a.
Declarative principle (hypothetical)
If a speaker utters a sentence that has a denotation of type hs, ti, this
has the following effect: . . .
b.
Interrogative principle (hypothetical)
If a speaker utters a sentence that has a denotation of type hhs, ti , ti,
this has the following effect: . . . .
c.
...
At first glance, it may seem that a general principle such as (3.9) is more
uniform—but this uniformity comes at a price: We have to complicate the system of semantic composition, we have to give up the assumption that embedded
sentences and matrix sentences have the same denotation, and we have to assume
that every sentence contains a silent operator that is otherwise unmotivated. Moreover, just as we need one principle per sentence type in (3.12), we would need one
silent operator per sentence type in the compositional system—and then, in addition, a general principle that relates denotations to use. So, on balance, putting
information about force into the compositional system is less parsimonious than
keeping it outside of the system. At the same time, having extra-compositional
principles regulate the dynamic effects of sentences seems intuitively quite appropriate. After all, what we want to model is constraints on possible uses. What
better place to put these constraints than in the principle(s) that connect denotation
and use?
I shall hence proceed on the assumption that the form–force mapping is mediated by extra-compositional principles or conventions of use. Once the basic
concepts are in place, I will return (in Section 6.5) to the question whether a representation of force in the compositional system is necessary.
CHAPTER 3. CLAUSE TYPES
3.4
45
Illocutionary acts
I have just argued that, if it can be done, it is more parsimonious and conceptually
attractive to represent conventional constraints on use outside of the system of
semantic composition. Before we proceed, in the next chapters, to the question of
what the nature of these constraints should be, it is useful to set aside a popular
conception of modeling the effects of utterances and the constraints on use that
derive from these effects, viz., that of speech act theory in the sense of Austin (1962)
and Searle (1969, et seq.).
As mentioned above, Krifka (2001b, to appear) takes matrix sentences to denote
high-level Illocutionary Acts such as Assert and Order, etc. By analogy, one
could think that the extra-compositional conventions I envision map sentences to
the same kind of act:
(3.13)
a.
Declarative illocutionary principle (hypothetical)
If a speaker utters a sentence that is morpho-syntactically declarative
(or: that has a denotation of type hs, ti), he thereby Asserts the content
of the sentence
b.
Interrogative illocutionary principle (hypothetical)
If a speaker utters a sentence that is morpho-syntactically interrogative (or: that has a denotation of type hhs, ti , ti), he thereby Questions
the content of the sentence
c.
Imperative illocutionary principle (hypothetical)
If a speaker utters a sentence that is morpho-syntactically imperative
(or: that has a denotation of type X), he thereby Commands the
addressee to fulfill the content of the imperative.
It is well-known, however, that this cannot work: Sentences of a given type are
not always used to perform the same high-level illocutionary act, even when they
are used entirely literally and sincerely. This is particularly easy to illustrate in
the case of imperatives: As Schmerling (1982) pointed out, imperatives can be
used to perform a wide variety of illocutionary acts (Kaufmann (2012, based on
CHAPTER 3. CLAUSE TYPES
46
the author’s dissertation, Schwager (2006)) calls this the Problem of Functional
Inhomogeneity):
(3.14)
a.
Stand at attention!
(Command)
b.
Don’t touch the hot plate!
(Warning)
c.
Hand me the salt, please.
(Request)
d.
Do the right thing!
e.
Take these pills for a week.
(Advice)
f.
Please, lend me the money!
(Plea)
g.
Get well soon!
h.
Drop dead!
i.
Please, don’t rain!
j.
Okay, go out and play.
(Permission)
k.
Okay then, sue me, since it’s come to that.
(Concession)
l.
Have a cookie(, if you like).
(Exhortation)
(Well-wish)
(Curse)
(Absent Wish)
(Offer)
A conceptually unattractive reaction to this problem would be to disjunctively list
the possible illocutionary forces in the ‘imperative principle’:
(3.15)
Imperative illocutionary principle (hypothetical)
If a speaker utters a sentence that is morpho-syntactically imperative
(or: that has a denotation of type X), he thereby Commands or Requests
or Wishes or Advises or . . . the addressee to fulfill the content of the
imperative.
Similarly (and just as unattractively), in a Krifka-style analysis, we could assume
that imperatives (and sentences of all other types) are multiply ambiguous, with a
large range of possible force operators (all of them silent) that all can occur on the
matrix level in imperatives.
Instead, the approach pursued here is to assume that imperatives (and sentences
of other types) have a uniform force on all their (sincere) uses. This force, then,
must be quite general. It is only in context that this force gets strengthened to
CHAPTER 3. CLAUSE TYPES
47
something as specific as Order. Borrowing a term from Chierchia and McConnelGinet (1990), we can call the (general, underspecified) force assigned to sentences
by linguistic convention their Sentential Force, to distinguish it from the concept
of illocutionary force. The theory I will lay out in the following will not make any
essential use of the concept of illocutionary force. Nor will it involve a concept
that is anything like it, at least to the extent that illocutionary force is understood
in terms of high-level acts such as Promise, Order or Assert.
The reason for this is that I do not see any advantage in a two-step procedure that
first maps utterances to illocutionary acts and then relates these types to their use in
context. Such a two-step procedure would of course be advantageous if either the
mapping from sentence types to illocutionary acts or the mapping from illocutionary acts to purposes of interlocutors would be homogeneous (ideally, functional).
And it would be almost mandated if both mappings were homogeneous.
But neither mapping is, in fact, homogeneous. We have already seen that
imperatives do not generally correspond to a single illocutionary type (cf. (3.14)),
and it is no more homogenous for other sentence types. And illocutionary acts
of any type can be performed for all kinds of reasons (in Austin’s terms, with all
kinds of perlocutionary intentions), and give rise to all kinds of inferences about
the speaker’s preferences, beliefs and choices. It is hence far from obvious that
employing a level of description that uses high-level illocutionary acts as the basic
units of organization is helpful.
Further, not only is it dubious that employing a level of illocutionary description
is theoretically helpful, it also seems that which exact kind of illocutionary act has
been performed is often irrelevant in linguistic practice: Suppose A and B are in a
car, B is driving, A has a map. The following interaction takes place:
(3.16)
A: Take a left here.
B: [turns left]
What is the illocutionary force of A’s utterance? Is it an order? A request? A
suggestion? All three? Neither? More importantly: Does it matter? A wanted B
to turn left. B observed A’s utterance, concluded that A wanted him to turn left,
CHAPTER 3. CLAUSE TYPES
48
knew that their current goals are aligned, and turned left. B did not have to decide
what the illocutionary force of A’s statement is, nor is it clear that A intended any
particular illocutionary force. He intended to get B to turn left, and thought the
imperative was a good means to get him to do this, and it was, that is all. It is
not clear that we gain anything in describing this instance of communication by
making reference to illocutionary acts.8
I hence reject the idea (implicit in Searle’s work, and quite explicit in authors
such as Bach and Harnish (1979)) that what interlocutors do in conversation, first
and foremost, is to try and recognize each other’s ‘illocutionary intentions’, i.e., that
it is somehow a fundamental part of interpretation to identify which illocutionary
force a speaker has intended. Something like this may be crucial in certain, quite
special contexts (e.g., when a boss says to his employee You look awful, take the
rest of the week off—in this case, it may be crucial whether this counts as an order
or merely a piece of advice), but it is not central to communication in general.
Giving up the idea that it is paramount to explain how utterances of sentences
of a given type relate to illocutionary types (and how illocutionary types relate
to pragmatic inference) opens up the possibility of a simpler, and quite attractive,
understanding of the role sentence types play in language use. Uttering a sentence
of a given type has certain consequences, determined by linguistic conventions.
These consequences, together with facts about the contents and inferences about
language users’ goals, explain patterns of language use.
This is not to say that the acts that Austin and Searle singled out as illocutionary
acts are somehow suspect and that labels for these acts are not occasionally useful.
They are quite helpful as descriptive labels for uses-in-context. The labels on
the examples in (3.14) are a perspicuous way to make the point that different
imperative utterances have quite different effects in context. Secondly, there are
natural language predicates that correspond, at least roughly, to illocutionary types:
8
Of course, one could retreat and say that A’s utterance had some weak or underspecified
illocutionary force. But this just is the point: It is not clear that we need anything other than such
weak or underspecified forces—forces that are general enough so that we can assume they stand in
a one-to-one relation with sentence types. That is, we need sentential forces, but it is not clear that
we need high-level illocutionary forces of the kind discussed by Searle and Austin.
CHAPTER 3. CLAUSE TYPES
49
order, promise, assert, and so forth. Ultimately, we will want to specify the lexical
meanings of these verbs, and in so doing, we may well need to appeal to some of the
notions that have played a role in philosophical speech act theory. But the fact that
there are such verbs, in itself, is no reason to assume that the concepts involved
in their lexical semantics play a particularly central role in how communication
works.9 And the existence of such verbs (which may specify quite particular
contextual effects of the utterance) is sufficient to explain why these verbs are
highly useful for descriptively categorizing sets of uses-in-context.
There is one property of ‘illocutionary verbs’ (i.e., verbs that roughly correspond
to the illocutionary acts studied by speech act theorists) that sets them apart: They
can be used in ‘explicitly performative sentences’, such as (3.17).
(3.17)
I (hereby) promise to come.
In Chapter 7 I will show how the semantics of such verbs can be specified without
making reference to corresponding illocutionary acts, in a way that explains the
peculiar behavior of sentences like (3.17).
3.5
Sincerity conditions
There is one concept from speech act theory that one might take to be central in
our understanding of the form–force mapping, even if one rejects the centrality of
illocutionary notions: The concept of sincerity conditions (or ‘felicity conditions’
in general). Indeed, it might appear that talking in terms of sincerity conditions
could enable the linguist interested in the semantics of clause types to leave aside
any complex considerations about what kind of effect an utterance of a sentence
has in context.
One may grant that there are, or seem to be, properly linguistic conventions
involved in the form–force mapping, and that linguists should investigate them.
9
After all, there are also various verbs whose lexical semantics refers to the volume and related
acoustic properties of an utterance, such as whisper, shout, etc. From this, we do not conclude that
‘volume’ is a central notion for our explanation of language use.
CHAPTER 3. CLAUSE TYPES
50
But maybe we just need to find a way, any way, to state the content of these
conventions, and then move on to studying linguistic facts that are unadulterated
by facts about use.
More concretely, could we not just capture the conventional effects of clauses of
different types on the model of truth conditions, by stating sincerity conditions for
utterances, leaving the concept of sincerity unanalyzed?
We certainly could do that. We could say that a complete description of a
language requires the characterization of a function FORCE which takes utterance
events into the set tSincere, Insincereu. For English, a partial characterization of
this function could look as follows:
(3.18)
FORCE is a function such that:
a.
if u is an utterance of a declarative, FORCEpuq “ Sincere iff the
speaker of u believes the content of u.
b.
if u is an utterance of an interrogative, FORCEpuq “ Sincere iff the
speaker of u wants the addressee to tell him which of the propositions
in the content of u is true.
c.
if u is an utterance of an imperative, FORCEpuq “ Sincere iff . . .
We would then insist that the job of the linguist is done with the specification of
FORCE. Again in analogy with truth conditions, we would say that just as it is not
the job of a linguist to explain the notion of truth, it is not the job of a linguist to
explain the notion of sincerity. All a linguist needs to do is to explicate properly
linguistic constraints on sincerity.
Let us grant for a moment that this kind of architecture is the right one, and
that the putative clean separation of labor between linguistics and other fields
responsible for pragmatics (such as psychology and/or philosophy) is practicable
and desirable. Then we still would need an answer, however preliminary and
subject to revision, to the question what the conventional effects of utterances of
sentences are, and what role these effects play in the inferences speaker-hearers
draw about each other’s utterance choices. But this will require, in turn, an answer,
CHAPTER 3. CLAUSE TYPES
51
however preliminary and subject to revision, to how the concepts Sincere and
Insincere relate to language use; why it is a fact, if it is a fact, that speakers
generally strive to make utterances that are Sincere, and under what conditions
this generalization does and does not apply.
That is, even if we grant that, in some sense, a theory answering such questions
is outside the domain of linguistics proper, linguists still need such a theory—if
only for methodological reasons. It is well-known, for example, that there is no
pre-theoretical way to distinguish conversational implicatures from entailments—
but then, without a theory of how pragmatic inference works (and how it works
for sentences of all types), it is unclear how the data (actual utterances and native
speaker’s intuitions about possible utterances) relates to our linguistic theories.
Perhaps more importantly, the assumption that there is a function like FORCE
which maps utterances (or utterances and contexts) into tSincere, Insincereu is
not as theoretically harmless as it may seem. Sincerity (or felicity) conditions are
not just any specification of constraints on use, but rather a quite particular one.
Assuming that a specification of sincerity (or felicity) conditions is the correct way
to account for clause-typing amounts to the assumptions that what matters are
preconditions, i.e., rules that say when under what circumstances a sentence can be
sincerely uttered. But this is far from the only logical possibility. Indeed, as I will
argue in the following chapters, it is not in such preconditions, but rather in their
consequences (that is, postconditions) that utterances of sentences of different types
differ.
3.6
Summary
Sentences of different types are used in different ways, and support different kinds
of pragmatic inferences. In this chapter, I have pointed out that this cannot be
explained either by general principles of cooperative behavior, nor by the fact that
they have denotations of different kinds, at least if these denotations are anything
like the denotations these sentences have when embedded in other sentences.
I have further argued that it may be possible (and, if possible, desirable) to
CHAPTER 3. CLAUSE TYPES
52
capture conventional constraints on use outside of the system of semantic composition, and have cast doubt on the usefulness of the concept of illocutionary acts
in understanding language use. The latter two points are preliminary, and will
have to be revisited later on. In the following chapters, I will develop a theory
of clause-typing that assumes extra-compositional conventions of use, and does
not make any direct use of illocutionary concepts. Insofar as it is successful, it
will demonstrate that we can do without appeal to illocutionary acts. Once this
theory is in place, we can reconsider the question whether there is a reason to
represent force within the system of semantic composition, a point I will return to
in Section 6.5.
Chapter 4
The sentential force of declaratives
In the last chapter, I raised some basic questions concerning clause typing, arguing
that each clause type must be conventionally associated with its own use. In
so doing, however, I only spoke in the most general terms about what such a
conventionally-specified use is. In this chapter, I want to remedy this, focussing on
the case of declaratives.
In our idealized system for pragmatic reasoning, PSen , declaratives correspond
to the sentences of the sublanguage L$ , defined as:
(4.1)
ˇ
(
L$ :“ $ φ ˇ φ P Prop
Since we are assuming that matrix sentences have the same denotation as embedded
sentences have, we take each sentence of Sen to denote the set of possible worlds
that the corresponding Prop sentence denotes:
(4.2)
0
8Sen
0 8Prop
$φ
:“ φ
The question is now: What needs to be added to PSen to enable the system to
model the use of declarative sentences? This chapter explores—in an informal
manner—a number of hypotheses that answer this question. I will settle on a
particular proposal, which then is integrated into the system of dynamic pragmatics
in Chapter 5.
53
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
54
Before I examine some of the possibilities, it will be useful to have a set of explanatory targets—things about declaratives that we want to explain. I summarize
these in the rest of this section.
Declaratives are well-suited to express beliefs, and thereby to inform. This
property is instrumental in explaining why declaratives are typically used to convey
information: When a speaker utters a declarative, by and large, the audience is
licensed to infer that he takes the content of the declarative to be true, and if the
audience further takes the speaker to be well-informed, they can further conclude
that the content of the declarative is indeed true. In Chapter 2, I showed how
to derive such an inference in the obvious way: If the addressee believes that the
speaker would utter a declarative with content p only if he believes p to be true, then
observing an utterance of a declarative with content p will lead him to conclude
that the speaker does, indeed, believe p to be true. But in Section 3.1, I pointed out
that the contextual assumption driving this inference needs justification. And the
justification must be particular to declaratives, as sentences of other types are not
typically used to express beliefs.
Declaratives make claims, and thereby are subject to reprobation.
When a
speaker utters a declarative, he opens himself up to the possibility of reprobation. A speaker can be criticized, or be said to be ‘in the wrong’ if it transpires that
he uttered a declarative he knew to be false.
One may well wonder if this is something a linguistic theory of language use
should explain. Isn’t an injunction against lying rather something that should be
explained by a theory of ethics or the like? It is, and that means that this property
should not be explained by a linguistic theory of language use alone, but rather
only in tandem with a suitable theory of moral behavior. But the property also
cannot be explained by a theory of moral behavior alone. The theory of language
use must provide a ‘hook’ on which an injunction against lying can be hung.
The comparison with imperatives is useful. Whatever their precise semantics
is, intuitively, the content of an imperative should determine at least fulfillment
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
55
conditions, and thereby determine a set of truth conditions: The conditions that
obtain if the imperative is fulfilled. But when a speaker utters an imperative whose
truth-conditions are not fulfilled, he does not open himself up for reprobation, and
the injunction against lying does not apply. By contrast, this injunction does apply
to the truth conditions associated with declaratives. But it seems odd to assume
that a theory of moral behavior makes reference to morpho-syntactic distinctions.
The division of labor should be roughly as follows:
(4.3)
Theorem of a complete theory of language and language use:
A speaker who utters a declarative with content p under conditions C makes
a claim that p.
(4.4)
Theorem of a complete theory of moral behavior:
It is morally objectionable to make a claim that p if one does not
know/believe/have good reason to believe that p (except if conditions C1
obtain).1
(4.5)
From (4.3) and (4.4), it follows that:
It is morally objectionable to utter a declarative with content p under conditions C if one does not know/believe/have good reason to believe that p
(except if conditions C1 obtain).
(4.5) should no more be a theorem of a theory of moral behavior alone than it
should be a theorem of a theory of language and language use alone.
It is important to note that this property is, in principle, independent from the
fact that declaratives are well-suited to express beliefs. Not every explanation of
why declaratives have the latter property also explains why they have the former
property. In many typical cases, the two properties may be argued to be tightly
connected. Suppose an utterance of a declarative is taken to be honest (i.e., based
on a belief that its content is true), but actually is not. In this case, we could explain
1
The set of conditions C1 is a qualification necessary because I do not want to claim that the
correct theory of moral behavior categorically proscribes lying in all situations.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
56
the possibility for moral judgement in terms of norms on cooperative behavior—
i.e., by an injunction to not make other people believe something that you take to
be false. But a speaker is subject to reprobation for (knowingly) false statements
even in contexts in which he is not presumed to be honest—if we are in a situation
in which I take it to be likely that you would say things that are not true, that does
not give you license to utter untruths with impunity. A speaker cannot (at least
usually) argue that his declarative utterance was not a lie on the grounds that he
knew in advance that the audience took him to be untrustworthy.
Utterances of inconsistent declaratives require retraction.
Speakers are gener-
ally expected to maintain consistency between the contents of the declaratives they
utter. This does not mean that they cannot acknowledge that they now take the
content of a previous declarative to be false, and assert the opposite. But if a speaker
wants to (sincerely) utter a declarative that is inconsistent with his previous ones,
he has to acknowledge this fact, and generally will be required to indicate which
previous declarative he wishes to retract if there are multiple such candidates.
Appreciating this property is somewhat complicated by the fact that speakers
are not always aware of the inconsistency of two propositions. A typical instance
is a failure to be logically omniscient. Suppose a group of novice students of set
theory do not know that the Axiom of Choice and Zorn’s lemma are equivalent,
given the other axioms of ZF. About a given extension E of ZF, one of them first
utters (4.6), and a short while later, he utters (4.7).
(4.6)
E entails the Axiom of Choice.
(4.7)
E does not entail Zorn’s lemma.
His second utterance may appear to be perfectly fine, without any acknowledgement of contradiction. Pre-theoretically, we do not know whether that is because
the consistency requirement does not apply in this case (because the consistency requirement really is ‘Utterances that are known to be inconsistent require retraction’)
or because, even though the consistency requirement applies, the interlocutors are
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
57
unaware of what it mandates. But it is clear that if the interlocutors are or become
mutually aware that (4.6) and (4.7) are inconsistent, one of the two utterances must
be retracted.
Similar issues are raised by the fact that speakers and listeners do not generally
have perfect recall: They do not always remember all declarative utterances that
were made. So if (4.6) was made hours prior to (4.7), the inconsistency may not
be recognized simply in virtue of the fact that none of the discourse participants
remembers that (4.6) was uttered. In what follows, I will ignore issues of imperfect
recall and failures of logical omniscience. Indeed the idealized representation of
beliefs in the system set up in Chapter 2 takes agents to be logically omniscient
and to have perfect recall. I will maintain this representation throughout the
dissertation, as the drawbacks associated with these idealization are not at issue.
Again, I want to note that while this property often is closely connected to the
fact that declaratives are good vehicles for expressing beliefs, this does not mean
that any explanation of the latter property explains the need for retraction. One
reason is that beliefs change over time. It is perfectly reasonable to believe, at time
t, that p and to no longer believe p at a later time t1 . But then, expressing (at t) the
belief that p and expressing (at t1 ) the belief that
p is a perfectly sensible thing to
do, as well. So the fact that declaratives tend to express the belief that their contents
are true is not, in itself, an explanation why subsequent utterances of declaratives
need to be consistent, or be made consistent by retraction.
Summary The properties we wish to explain are summarized in (4.8).
(4.8)
a.
Declaratives are well-suited to express beliefs, and thereby to inform.
b.
Declarative utterance make claims, which are subject to reprobation.
c.
Utterances of multiple declaratives with inconsistent contents require
retraction.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
4.1
58
Mapping the territory
It will be helpful to have a rough understanding of the territory we are about
to enter—what is the space of plausible hypotheses about sentential force? That
space may well be infinite, but I will restrict attention to a number of prominent,
viable hypotheses from the literature, which have been proposed either directly as
hypotheses about declaratives or as hypotheses about the high-level speech act of
assertion.
The accounts I will consider fall into two broad classes: On the one hand, there
are accounts that take the specification of force to be a descriptive fact about speakers
of the language, a fact about how speakers actually do behave. On the other hand,
there are accounts that take the specification of force to be a normative fact, a fact
about how speakers ought to behave.2
The first kind of account generally links declaratives to belief directly: Either
it is simply taken to be a descriptive generalization that speakers tend to utter
only those declaratives that they believe to be true (as in Lewis’s (1975) influential
account, discussed in Section 4.2) or it is said that declaratives ‘express beliefs’.
To be interesting, such a claim must be combined with an account of what it is to
express a belief, and much of the discussion in Section 4.3 will center around this
question. Ultimately, I will argue that such expressive theories are equivalent to,
or need to combined with, either a Lewis-style account or a normative account in
order to explain how sentential force is determined.
The second kind of account—which takes sentential force to be essentially
normative—comes in two sub-types: Those that take sentential force to be determined by normative preconditions (discussed in Section 4.5.1) and those that
take it to be determined by normative consequences of utterances (discussed in Section 4.5.2). The former accounts specify ‘rules’ for when a declarative utterance can
be made (in analogy to a law that says that one can become president of the USA
2
This division roughly corresponds to Harnish’s (2005) distinction between Gricean and Austinian theories—though the former include only such theories that make essential reference to the
speaker’s intentions, which arguably is not a necessary feature of the first type of theory I discuss.
One prominent such theory, that of Bach and Harnish (1979), discussed in Section 4.3.1, however,
does put intentions center stage.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
59
only if one is a natural-born citizen), while the latter type of account specifies ‘obligations’ that arise from making a declarative utterance (in analogy to the action of
signing a rental contract, creating the obligation to pay a certain amount of money
in regular intervals).
Existing normative accounts, especially of the second subtype, relate to the
notion of belief only in a rather indirect way—the normative consequences of
declarative utterances do not in themselves involve the notion of belief. They only
make it so that an agent who wants to abide by these obligations generally has the
incentive to utter declaratives whose contents he believes to be true.
The account that I will ultimately propose (explicated in Section 4.6) can be
viewed as a hybrid of ‘expressive’ and ‘normative consequences’ accounts. It
directly appeals to the notion of belief in that it takes the conventional force of
declaratives to be to create a public belief of the speaker, where what a public belief
is is explicated in terms of its normative consequences.
This account will, in many ways, be similar to the popular account of assertion
of Stalnaker (1978, et seq.). Stalnaker takes an assertion of a proposition p to
be a proposal to add p to the conversational common ground of the discourse
participants. This conversational common ground is understood as the set of
propositions that the speakers mutually take each other to believe, or pretend
to do so. In this sense, it is a notion of ‘public belief’. In the recent linguistics
literature, various authors have followed Gunlogson (2003) in constructing the
common ground of a set of discourse participants from their individual ‘public
beliefs’ (Gunlogson 2008, Davis 2009, Farkas and Bruce 2010, Farkas and Roelofsen
forthcoming, Davis 2011)—a proposition is in the common ground if it is a discourse
commitment of all interlocutors. This separation is necessary to talk about cases of
overt disagreement: If A asserts p and his addressee B fails to accept p as true, p is
a public belief of A, but not of B. This shows that we need a notion of individual
public belief.
It is worth noting that speaker presupposition, at least in the way Stalnaker understands it, is not this notion. While speaker presupposition (in contrast with the
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
60
common ground) is an individual notion, it is not suitable to model overt disagreement: “one presupposes that φ only if one presupposes that others presuppose it
as well” (Stalnaker 2002, p. 701). It is not quite clear what ‘individual public belief’
would amount to in Stalnaker’s terminology. The only viable candidate seems to
be that φ is an individual public belief of A iff it is common ground that A believes p.
But the relationship between (actual) belief and speaker presupposition/common
ground in Stalnaker’s model is not quite clear, so it is not clear what this would
amount to. Also, note that this definition entails that if C and D explicitly agree
that A, who is absent, believes p, then p would thereby be a ‘public belief’ of A—but
that seems wrong.
My account of the sentential force of declaratives will be founded on a notion of
individual public belief, which can be construed as a generalization of the notion
of a ‘discourse commitment’ of Gunlogson, Davis, Farkas, et al. Besides adopting
the separation of common belief into individual public beliefs, it departs from
Stalnaker’s conception in that it stresses the normative character of ‘public beliefs’.
I shall not argue here one way or another whether the conception I will propose is
an elaboration of Stalnaker’s, a variant of it, or a replacement for it. The conceptions
are similar in many ways, but, minimally, they differ in emphasis, and the things I
emphasize will be instrumental in what I want to explain.
4.2
Declaratives as governed by Lewis-conventions
Our aim is to add something to the denotation of declaratives, in order to explain
why declaratives are well-suited for informing, while interrogatives, etc. are (usually) not. Perhaps we do not need to add much. Perhaps it is just precedence that
enables addressees to make the right inferences. Speakers, by and large, use declaratives when they want to inform, and they use interrogatives, by and large, when
they want to seek information, and so on. Addressees are aware of this, so they will
assume, at least by default, that a speaker who utters a declarative (interrogative)
intends to convey that he takes the declarative to be true (intends to inquire which
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
61
of the answers to the interrogative are true).3 If we adopted this view, we could treat
declaratives in the way we treated the formulas of Prop in Chapter 2: We would
simply assume that agents generally believe each other to only utter declaratives
that they believe to be true (or that they desire the addressee to believe to be true),
and we could add similar contextual beliefs about interrogatives, imperatives and
so forth. These beliefs would be justified by the knowledge that speakers have
adhered to the relevant generalization in the past.
This is essentially the view of Lewis (1975, elaborating on Lewis (1969)). He
starts from the assumption that a (declarative) language L is simply a function
from linguistic forms to sets of possible worlds (or any other specification of truthconditions). Given this conception of a language, he asks what has to be the case
for a particular language to be in use in a population.
Lewis’ answer is that there has to be regularity in behavior in the population
in question, such that members of the population, in general, only utter sentences
of L if their L-truth-conditions obtain. At the same time, there is a regularity of
behavior in that population according to which its members, generally, when they
observe an utterance, assume that the L-truth-conditions of the sentence obtain.
Lewis requires that the regularity in question has a number of particular properties, the most important is that it be known to the members of the community, that
it be beneficial (most members prefer that it be in place to it not being in place, most
of the time), and that it be stable and self-sustaining: Given that most members of the
population obey the regularity, most of the time, most members of the population
have incentive to continue obeying the regularity, most of the time. Regularities
in behavior that have this property (and a number of others) are what Lewis calls
conventions. So Lewis says that a population uses a particular language L if there is
a Convention of Truthfulness and Trust with respect to L in the population:
“My proposal is that the convention whereby a population P uses a
3
Arguably, such precedential facts are also at play in the bread example in Section 3.2: While
it is true that German bread is well-suited for providing delicious sustenance, there are, of course,
many other uses it could in principle be put to. Similarly, the fact that it is not common practice
to store one’s possessions in someone else’s room is instrumental in my inference that the bread is
intended as a gift, and so forth.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
62
language L is a convention of truthfulness and trust in L. To be truthful
in L is to act in a certain way: to try and never utter any sentences of L
that are not true in L. To be trusting in L is to form beliefs in a certain
way: to impute truthfulness in L to others, and thus to tend to respond
to another’s utterance of any sentence of L by coming to believe that the
uttered sentence is true in L.”
(Lewis 1975, p. 7)
Let us consider briefly how this conception accounts for the three explanatory
targets set above, which I repeat in (4.9).
=(4.8)
(4.9)
a.
Declaratives are well-suited to express beliefs, and thereby to inform.
b.
Declarative utterance make claims, which are subject to reprobation.
c.
Utterances of multiple declaratives with inconsistent contents require
retraction.
Lewis’ account does well on (4.9a): If a Lewis convention of truthfulness is in effect,
it is clear why listeners (at least by default) would assume that speakers believe the
contents of their declaratives.
But it is not quite clear how well a Lewisean account does on (4.9b). Certainly,
the Lewisean could appeal to communicative reasons for reprobation, appealing
to a norm of cooperative/social/moral behavior like (4.10).
(4.10)
One should not: Make anyone believe p if p is not true / one does not
believe that p.
But this will not be enough: It does not explain, for example, why a speaker who
is criticized for uttering an untruth cannot defend himself by saying that he knew
in advance that he would not be believed. That is, ‘making someone believe that
p’ is not general enough as a ‘hook’ to hang an injunction against lying on. If the
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
63
Lewisean is to explain the possibility for reprobation in general, he would have to
stipulate something like (4.11).
(4.11)
One should not: Make a declarative utterance that, in general, would only
be made if one believed that p if one does not believe that p.
It is not quite clear how (4.11) could be motivated—except perhaps by observing
that any time someone violates the Lewis-convention of truthfulness, he thereby
threatens the very existence of the convention. Given that the convention is instrumental in ensuring successful communication in many cases, this would be
dangerous behavior. But this explanation makes direct reference to the convention
itself, and to its communicative function.4 But then, it seems that the same kinds
of considerations should apply to all Lewis conventions, and their functions, and
so we should hypothesize a general normative rule or convention that says:
(4.12)
One should not: Violate a Lewis convention (a regularity in behavior in a
population that is stable, beneficial, etc.).
But (4.12) pretty much turns Lewis conventions themselves into normative conventions. Lewis’ concept of convention-as-regularity would still be useful in explaining
how such conventions can arise without a pre-existing language (which was one
of his major goals), but it is not clear that the concept plays any essential role once
the Lewis convention has been fully established and hence ‘normativized’.
Finally, with respect to observation (4.9c), it seems that the Lewisean has to rely
on communicative reasons again: If a speaker uttered p at a previous time, and
was believed to speak truly, but now has changed his mind and wants to assert
p, his communicative goals may well be thwarted if he does not acknowledge
the falsity of his prior utterance. If he does not, his audience might assume that he
believes both p and p (i.e., that he has inconsistent beliefs), or that he forms beliefs
rather nilly-willy, and hence that his new belief in
4
p is not particularly likely to
And it seems that (4.11) should, as well: It does not seem right to assume that (4.11) applies to
any action that, in general, is only be performed if the speaker has a particular belief—we do not
usually fault agents for being exceptions to reliable generalizations that support belief ascriptions.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
64
accord with the facts. Acknowledging the falsity of his previous utterances may
serve to guard against audience inferences like these.5 So the Lewisean can explain
the need for retraction when it is motivated on communicative grounds—but only
then.
Summary An account of the force of declaratives in terms of Lewis-conventions
explains why declaratives are well-suited to express beliefs and to inform. It also
explains why a speaker is subject to reprobation for utterances of declaratives that
he knew to be false—but only if he had reason to think that his utterances will be
taken to be true. Similarly, it can explain why retraction of inconsistent declaratives
is necessary—but only if this can be motivated by communicative reasons. At first
glance, this may seem sufficient, but as I will argue in Section 4.7.4, it is not.
4.3
Declaratives as expressing beliefs
Starting from our first target for explanation—declaratives are well-suited for expressing beliefs—an obvious idea is to assume that utterances of declaratives simply are expressions of belief. Such an analysis would assume a rule, convention, or
principle that ensures that (4.13) is true.
(4.13)
Whenever a speaker utters a declarative with content p, he thereby expresses a belief that p.
As will become clear in the following subsections, I think that such an approach is
ultimately unhelpful in understanding declarative (or any other) sentential force. I
shall discuss it at some length because it is a popular and initially attractive way to
answer the question ‘What do declaratives do?’—they express beliefs! We shall see,
however, that this thesis, in the way it has been understood, either (a) is untenable,
or (b) is essentially Lewisean, or (c) is essentially a variant of the normative theories
to be discussed in Section 4.5.
5
Observe, though, that in many cases, only an acknowledgement of falsity is needed—not an
exhaustive explanation why the previous belief was justified, and why the new belief is justified.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
65
(4.13) immediately raises the question what ‘expressing’ a belief amounts to. I
will discuss three possible answers in the rest of this section (and a fourth in Section 4.4). The first two can be set aside quickly; the third requires more substantive
discussion.
Hypothesis I: To express a belief is to indicate that one has it.
This conception is
essentially the broad sense of ‘expressing’ I have used when I said that declaratives
are well-suited to express beliefs, meaning that they are well-suited to give one’s
audience a reason to think that one has the belief.
On this conception, saying that declaratives, by convention, express beliefs
is simply saying that utterances of declaratives are actions that, by convention,
are generally good reasons to think that the speaker believes the sentence to be
true. The obvious question is why that would be so, and the obvious answer is:
Because that is what speakers tend to do. On this conception, then, the thesis
that declaratives express the belief that their content is true is simply Lewis’ idea
of a convention of truthfulness, and everything I have said above applies to the
expressive thesis, as well.
Hypothesis II: To express a belief is to perform an action caused by it. Williams
(2002) and Owens (2006) analyze assertions as expressions of belief, with a rather
different conception of what ‘expressing’ means. According to their conception, to
say that an action is expressing a belief is to say that it is caused by a belief. Among
other things, that means that an action can be said to express a belief only if the
agent in question actually has the belief. Thus Owens writes:
[E]xpressing a belief in action is one way of acting on that belief and one
can’t act on a belief which one does not have. So, as I use the term, an
expression of belief may be poor or inadequate but it can’t be insincere:
you can’t express beliefs you don’t have.
(Owens 2006, p. 109)
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
66
This sense of ‘expressing’ is rather liberal: Any action that is motivated by a
certain belief counts as expressing that belief. As MacFarlane (2011, p. 81) puts
it: “In reaching for the umbrella as I head to the door, I express my belief that
rain is likely.”6 But it is not liberal enough. As an account of assertion, it is too
restrictive because it rules out the possibility of insincere assertions.7 And as a
claim about declaratives (‘Whenever a speaker utters a declarative, he performs an
action caused by a belief in the truth of its content’), it is simply false. The only way
to save such an account of declaratives would be to say, instead, that an utterance
of a declarative is only sincere if the utterance is caused by the appropriate belief.
But this just begs the question (see also Section 3.5). Without either a descriptive
generalization that speakers tend to speak sincerely (i.e., a Lewis-convention of
sincerity) or a normative rule that speakers ought to speak sincerely, this version
of the expressive thesis does not have any explanatory power.
I conclude that the Williams (2002)/Owens (2006) conception of ‘expressing a
belief’ is not useful in elucidating the conventions governing the use of declaratives,
for it is either hopeless or it ends up being equivalent to one of the alternative
conceptions explored in this chapter.
Hypothesis III: To express a belief is to have a reflexive intention.
The rest of this
section will be concerned with a quite elaborate, and somewhat popular, conception
of what it is to express a belief (or an attitude in general), viz., that of Bach and
Harnish (1979). Their account is somewhat exceptional in philosophical speech
act theory in that they explicitly relate their account of assertion in terms of belief
expression to sentence types, in particular with declaratives (though, as we will
6
I am not certain that this use of express is consistent with its use in everyday conversation—
such a taking of an umbrella certainly can be described, on occasion, as expressing the belief that it
is raining—especially if it is done deliberately with the intention to convey this belief—but it is not
clear to me that any taking of an umbrella that is motivated by a belief that rain is likely counts as
‘expressing’ that belief.
But we can, of course, take express to be a technical term here, or we could replace it by
another, such as display, without changing anything about the putative analyses of declaratives
and assertions.
7
This leads Williams (2002, p. 74) to define assertion disjunctively: “A asserts that p where A
utters a sentence S which means that p, in doing which either he expresses his belief that p, or he
intends the person addressed to take it that he believes that p.”
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
67
see below, this connections turns out to be an essentially Lewisean one). This is the
first reason I discuss their account at some length. The second is that their intricate
account poses some conceptual difficulties which make it hard to evaluate. Since
they are also the main proponents of a popular approach to explicitly performative
sentences, to be discussed in Chapter 7, it is useful to point out these conceptual
difficulties here.
4.3.1
Bach and Harnish’s (1979) on communicative speech acts
Unlike Austin and Searle, Bach and Harnish (1979) (B&H) sharply distinguish two
kinds of illocutionary acts, following Strawson (1964): Those that are essentially
conventional and those that are, by contrast ‘communicative’. The former category
contains Effectives such as bidding and bequeathing, as well as Verdictives, such
as finding guilty. The latter class contains most other illocutionary acts, in particular
assertions, promises, orders, etc. According to B&H, such communicative speech
acts uniformly ‘express’ attitudes, such as beliefs, desires, and intentions. Their
conception of ‘expressing’ is based on the notion of a reflexive intention, or Rintention, which is familiar from Grice’s (1957) definition of ‘non-natural meaning’:
(4.14)
To R-intend is “to produce some effect in an audience by means of the
recognition of this intention”
(Grice 1957)
In terms of this, B&H defines what it is to express an attitude:
(4.15)
Expressing: For S to express an attitude is for S to R-intend the hearer to
take S’s utterance as reason to think S has that attitude.
(Bach and Harnish 1979, p. 15)
To express a belief thus amounts to (4.16).
(4.16)
Expressing a belief that p: For S to express the belief that p is for S to Rintend the hearer to take S’s utterance as a reason to think that S believes
that p.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
68
Unfortunately it is unclear how (4.16) should be understood. What exactly does
it mean to ‘R-intend the hearer to take S’s utterance as a reason to think that S
believes that p’? We need to understand this in order to be able to evaluate the
thesis that declaratives express beliefs against our three explanatory targets:
=(4.8)
(4.17)
a.
Declaratives are well-suited to express beliefs, and thereby to inform.
b.
Declarative utterance make claims, which are subject to reprobation.
c.
Utterances of multiple declaratives with inconsistent contents require
retraction.
The most important question in this connection turns out to be: Is to R-intend that
Y also to intend that Y? That is, does someone who expresses a belief that p intend
his addressee to think that he has the belief that p? As we shall see, the answer to
this question is not clear.
4.3.2
R-intentions à la Bach & Harnish
For Grice, to mean (‘declaratively’) that p, roughly, is to intend to get H to form the
belief that p, by means of recognition of the intention itself. This makes quite plain
that to Grice-R-intend that Y is (in part) to intend that Y.
Things are not so clear for B&H, though. They find fault with Grice’s definition
insisting (following Searle (1969, p. 46–50)) that ‘to get H to believe that p’ (or their
equivalent ‘for H to take S’s utterance as a reason to think that S believes that p’)
is a mere perlocutionary effect that does not need to be satisfied for S’s illocutionary
intention to be fulfilled.
“Searle argues that the sorts of effects Grice mentions, such as beliefs,
intentions, and actions, are not produced by means of recognition of
the intention to produce them. For example, the hearer might recognize
that he is to believe something and yet refuse. These sorts of effects are
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
69
perlocutionary, and the speaker’s illocutionary act, whose identity he
is trying to communicate, can succeed without the intended perlocutionary effect (if there is one) being produced. So a reflexive intention
is involved in communication, just as Grice claimed, but the kinds of
intended effects he specified are not of the right sort. Getting the hearer
to recognize them does not constitute producing them. In section 1.6
we consider just what sort of reflexive intention is fulfilled merely by
being recognized.”
(Bach and Harnish 1979, p. 14)
Unfortunately, B&H never define what it is to ‘R-intend that Y’, but they make
various remarks about the notion:
[T]he intended effect of an act of communication is not just any effect
produced by means of recognition of the intention to produce a certain
effect, it is the recognition of that effect.
(Bach and Harnish 1979, p. 15, emphasis in original)
And, after introducing their definition of ‘expressing an attitude’ ((4.15) above),
they write:
Accordingly, the intended illocutionary effect (or simply illocutionary
intent) is for H to recognize that R-intention.
(Bach and Harnish 1979, p. 16)
So B&H stress that the intended effect is the recognition of the intention itself, and
that the R-intention is “fulfilled merely by being recognized”. But if the intention is
fulfilled by being recognized, it seems it cannot involve an intention for something
other than being recognized. But then, it seems to ‘R-intend that Y’ cannot involve
an intention to Y—for such an intention is fulfilled only if Y obtains, not just if the
intention is recognized. But then, it becomes difficult to see what an ‘R-intention
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
70
to Y’ really could be—what is the role of Y in the intention, if it is not the thing that
needs to occur for the R-intention to be fulfilled?
An intention that is fulfilled once it is recognized, is, it seems just an intention
to be recognized—adapting Searle (1969)’s helpful notation, in which every occurrence of the verb intend or the noun intention is indexed with a name for the
intention predicated, this amounts to the statement in (4.18):
(4.18)
S R-intends iff S intendsi1 that his intentioni1 is recognized.
But what this formulation makes obvious is that there is no place for an argument
of ‘R-intend’ in the definiens: According to this definition, there can only be one
R-intention, whose content is only its own recognition. We can contrast this with
Grice’s original formulation (somewhat expanded), in the same format:8
(4.19)
S R-intends Y iff S intendsi1 that
a.
Y;
b.
the intentioni1 is recognized;
c.
Y obtain, in part, because of (b).
But (4.19) will be unsuitable for B&H, it seems, because it involves an intention to
Y, and hence the intention described in (4.19) will not be fulfilled merely by being
recognized (this only satisfies part (b)), but requires further that Y obtain.
At the same time (4.18) will not be useful in communication, because (4.18)
8
Considering this version of Grice’s definition helps to foreground that the conceptual difficulty
I am seeing here is not the one raised by Sperber and Wilson (1986), which casts doubt on any notion
of a reflexive intention: Sperber and Wilson argue, in a nutshell, that reflexive intentions cannot be
mentally represented, due to their self-referential nature. Their argument involves the premiss that,
in order to mentally represent an intention like (4.19), the occurrence of ‘the intentioni1 ’ in (b) has to
be replaced by a representation of i1 itself, leading to an infinite regress. I think this argument can be
resisted, on the assumption that such a mental representation could involve ‘mental demonstrative’,
or something of the kind (essentially, a pointer) that refers to the overall intention i1 . Siebel (2003)
mentions such a possibility, but expresses concerns about such a mental demonstrative device. It
may not be trivial to posit the existence of such mental representations, but given modern computer
languages are rife with self-referential representations (functions that call themselves, objects whose
members are objects that point to the parent object, etc.), it does not seem too far-fetched to assume
that mental representations could have the same feature.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
71
carries only one bit of information—if understanding were the recognition of Rintentions as characterized in (4.18), then to understand someone would simply
be to say ‘Ah, he R-intends!’—not ‘He R-intends me to Y’, or anything like that.
So it seems (4.18) cannot be what B&H have in mind.9 But at the same time, if
their R-intentions really are entirely fulfilled just by being recognized, it seems that
(4.18) must be what they had in mind.
I think Siebel (2003) is getting at the same conceptual difficulty when he discusses what he calls a ‘mereological difficulty’ (which he attributes to Wayne Davis
(p.c.)):10
“Second, [R-intentions] give rise to a mereological difficulty. Let us abbreviate the content of a particular reflexive intention by ‘X’: S expresses
the attitude A iff S intends X. This intention is supposed to be identical
with the intention that an addressee H, by means of recognizing that S
intends X, take the utterance as a reason to think S has A. Intentions,
however, are individuated by their content. That is, intending p is identical with intending q only if ‘p’ and ‘q’ specify the same content. Hence,
the content given by ‘X’ should be identical with the content given by
‘H, by means of recognizing that S intends X, takes the utterance as a
reason to think S has A’. But how could they be identical? After all,
the former content seems to be a proper part of the latter because the
sentence specifying the latter contains ‘S intends X’ as a proper and
semantically relevant part.”
9
(4.18) has one advantage: It makes sense of B&H’s claim that R-intentions are “intended to be
recognized as intended to be recognized”. Siebel (2003, p. 357) notes this, wondering whether it is
supposed to follow from the definition of an R-intention. According to (4.18), R-intentions indeed
are intended to be recognized as intended to be recognized—and only that.
10
The only place I know of where Siebel’s ‘difficulty’ has been addressed is Witek (2009, p. 77).
But Witek proposes the following description of the content (again with Searlean indexing):
(i)
S intendsi1 that H, by means of recognizing this intentioni1 , takes Ss utterance as reason to
think that S has A.
But (i) is like Grice’s definition in that it is not an intention merely to be recognized, but rather an
intention that is fulfilled only if the ‘perlocutionary effect’ (H takes S’s utterances as a reason to
think that S has A) occurs.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
72
(Siebel 2003, p. 356)
But humbly calling the issue a ‘mereological difficulty’ obscures its importance:
One might think that all Siebel’s worry shows is that the part-whole structure
of intentions (or their contents) is quite special, in that it does not validate all
principles that other mereological structures satisfy. I think, to the extent that the
considerations above are correct, they actually show that B&H’s conception of Rintentions is incoherent. This would be quite devastating, as R-intentions are the
very foundations on which their whole theory is build.
Perhaps we can salvage their notion of an R-intention, if we interpret their
characterizations of what it takes to be an illocutionary R-intention somewhat
loosely. Here is a version that is, I believe, coherent, and that has some of the
properties B&H desire. I am not sure it is what they had in mind, but seeing as it
is the only coherent interpretation that I have been able to come up with, it is the
one that I am going to work with in what follows:
(4.20)
S R-intends that Y iff S intendsi1 that
a.
H recognizes that S intendsi2 that Y;
b.
H recognizes the overall intentioni1 ;
c.
(a) obtains, in part, in virtue of (b).
Now, the intention in (4.20) does not quite fit B&H’s description. The intention i1 is
not fulfilled just by being recognized: For i1 to be fulfilled, another intention has to
be recognized, as well, namely i2, the intention to Y. This intention itself, however,
is not part of the intention i1, it just occurs in its content.
If we now assume in addition that ‘recognize’ in the description of the intention
is factive—and it seems that B&H assume this—then S cannot intend i1 without
also having the intention i2. But then, H cannot ascribe i1 without also ascribing i2
and so recognizing that S has i1 will involve recognizing that he has i2, as well. So,
once i1 is recognized, i2 must have been recognized as well.
On this interpretation, R-intending that Y is not intending that Y—the Rintention can be fulfilled without Y being true, but it still is the case that one
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
73
cannot R-intend that Y without also intending that Y. Again, I am not sure that
this is what B&H have in mind,11 but it is the version I am going to work with in
the sequel.
4.3.3
Expressing a belief as R-intending
With a workable understanding of an R-intention, we can now consider the hypothesis that declaratives express beliefs, on my reconstruction of B&H’s definition
of an R-intention. We then can evaluate how well this hypothesis fares in helping
us meet the explanatory targets set out at the beginning of this chapter. I repeat
their definition of ‘expressing a belief’:
(4.21)
Expressing a belief that p: For S to express the belief that p is for S to Rintend the hearer to take S’s utterance as a reason to think that S believes
that p.
With this, the principle (4.13) becomes (4.22):
(4.22)
Whenever a speaker utters a declarative with content p, he thereby Rintends the hearer to take his utterance as reason to think S believes that
p.
This immediately raises the following worry: Declaratives are often uttered, even
sincerely, if the speaker does not, and cannot, intend that his audience form the
appropriate belief—e.g., if the speaker knows that he will be not be trusted, but
wants to ‘go on record’ with his statement.
But it is worth setting this worry aside for a moment to dwell on B&H’s conception. Recall again that B&H’s claim is that asserting is the expression of a belief.
11
One reason to doubt that this is exactly what B&H have in mind is that they appear to think
that intending that ‘H take S’s utterance as a reason to think he has the attitude’ is a content
that is somehow particularly suited for R-intentions. But the definition in (4.20) does not put
any requirements on Y—any Y could be R-intended according to this definition, as long as it is
something that S can intend.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
74
So maybe it is not that B&H’s conception of expressing works as a way to understand the thesis that (all) declaratives express beliefs, but maybe their conception
of how asserting works (when it does) with declaratives can provide a window into
understanding declaratives—for unlike other authors mainly concerned with the
illocutionary act of assertion, B&H do spell out how asserting relates to declarative
force—to an extent.
They relate illocutionary acts to the content of declaratives by the conception
of literalness of a speech act. They write $ p. . . p . . .q to represent ‘what is said by a
declarative’ (where p. . . p . . .q, roughly, characterizes truth conditions and $ is just
a representational device), and then state the following Compatibility Condition,
which defines the notion of L-compatibility (Bach and Harnish 1979, p. 34): An
utterance is L-compatible if . . .
(4.23)
If S is saying that $ p. . . p . . .q, S is expressing a belief that p.
And then they use this notion to define what a literal performance is (where F is
an illocutionary act type, ˚ a metavariable over symbols such as $):
(4.24)
Literal performance (Lit) S’s F-ing that P in saying that ˚p. . . p . . .q is literal
just in case:
i.
P “ p. . . p . . .q
ii.
F-ing is L-compatible with saying that ˚p. . . p . . .q
What this boils down to, in the case of declaratives, is essentially this:
(4.25)
An utterance of a declarative with content p is Literal if it expresses a
belief that p.
Spelling out Bach and Harnish’s conception of expressing:
(4.26)
An utterance of a declarative with content p is Literal if its speaker Rintends his audience to take his utterance as a reason to think that he
believes p to be true.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
75
But this is not essentially different from the hypothetical specification of sincerity
conditions in Section 3.5. And the same considerations apply: (4.26) does not tell us
anything about how declaratives are used, unless it is paired with an assumption
to the effect that speakers generally do, or generally should, strive to be literal.
B&H actually articulate such a principle:
(4.27)
Presumption of Literalness (PL): The mutual belief in the linguistic community CL that whenever any member S utters any e in L to any other
member H, if S could (under the circumstances) be speaking literally, then
S is speaking literally.
This is no different from saying:
(4.28)
There is a Lewis convention in CL to the effect that speakers speak literally
and that hearers take speakers to speak literally.
Bach and Harnish’s conception of declarative sentential force is hence essentially
Lewisean—except that the content of the convention is slightly different. Lewis’
original proposal said that speakers tend to utter declaratives only if they are true,
and hearers tend to expect them to, while (4.28) says that speakers utter declaratives
only if they R-intend to convey that they believe the content to be true.
This brings us back to the worry, set aside above, about whether it makes sense
to assume that any speaker who utters a declarative (without any overt indication
of insincerity) can be said to R-intend his audience to think he believes the semantic
content of what he says. In particular, it is easy to imagine contexts in which a
speaker knows that he will not be trusted (imagine, for example, a suspect being
questioned by the police, who he knows think he is guilty—such a suspect may
well insist he is innocent, even though he knows his interrogators will not believe
that he believes himself to be innocent. But if he knows this, he cannot intend to
change their mind.
It seems that what B&H must say about such a case is that the suspect is not
speaking literally. Perhaps because this seems an odd thing to say (at least on
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
76
our everyday understanding of the word literal), Bach and Harnish argue that
the speaker can still be taken to express a belief in such a case—revising their
characterization of expressing a belief:
Instead of saying that expressing an attitude is R-intending H to take
one’s utterance as reason to believe that one has that attitude, we can
say that it is R-intending H to take one’s utterance as sufficient reason,
unless there is mutually believed reason to the contrary, to believe that
one has that attitude.
(Bach and Harnish 1979, p. 291)
As MacFarlane (2011) discusses, it is not clear that this revised explication of ‘expressing’ can be made sense of, as it appears to require us to believe that an agent
can ‘intend p unless q’ if he knows that q is true. Suppose I concluded the discussion
of B&H’s account with the sentence in (4.29). What kind of intention would I claim
to have?
(4.29)
I intend to discuss this issue further, unless the present work is a dissertation in linguistics.
In summary, it seems that B&H’s account does not fare any better (or worse) then
Lewis’ original account on the explanatory targets at issue here:
=(4.8)
(4.30)
a.
Declaratives are well-suited to express beliefs, and thereby to inform.
b.
Declarative utterance make claims, which are subject to reprobation.
c.
Utterances of multiple declaratives with inconsistent contents require
retraction.
As on Lewis’ account, target (4.30a) is easily met, on the assumption that speakers
typically don’t intend to make their addressees believe things that they do not
believe themselves. Targets (4.30b) and (4.30c) can, again as on Lewis’ account
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
77
only be met in cases in which there are communicative reasons for reprobation or
retraction: We can explain why a speaker can be criticized for speaking falsely in
cases in which he has reason to think that he would be believed, but not otherwise.
And we can explain why a speaker would feel the need to retract if he otherwise
runs the risk of being thought of as holding inconsistent beliefs, but not otherwise.
As indicated before, while that may seem sufficient at first glance, we will see in
Section 4.7 that it is not.
4.3.4
Summary
The main result of the present section is this: while it may be adequate, at a certain
level of description, to say that declaratives express beliefs, saying so does not help
us in understanding how declaratives are used. The conceptions of ‘expressing’ I
discussed are either are equivalent to, or require an independent specification of, a
Lewis-convention or a set of normative rules governing the expression of beliefs.
I conclude that expressive accounts, at least from our current perspective, are not
an alternative to Lewis-style accounts and normative accounts of clause typing,
but rather are variants of such accounts. Consequently, I will largely ignore this
category of account in what is to follow.
4.4
Counts-as rules
The notion of constitutive rules have played a significant role in philosophical speech
act theory, especially in the work of Searle (1969). Constitutive rules are, in the
words of Searle, rules that “do not merely regulate, they create or define new
forms of behaviour. The rules of football or chess, for example [. . . ] create the very
possibility of playing such games.” For Searle, illocutionary acts are defined by
exactly such rules.
There is no general agreement on how constitutive rules should be understood,
or whether they exist at all (for various views on the matter and discussion, see
Ransdell (1971), Giddens (1984) and more recently Hindriks (2009)). In recent
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
78
discussion (e.g., MacFarlane (2011, Section 2)), constitutive rules are often taken to
have the same format as other, ‘regulative’ rules, i.e., they specify what one must
or must not (or should and should not) do. These kind of rules will be the topic of
Section 4.5.1.
Searle himself tends to talk about constitutive rules in terms what he calls ‘the
counts-as locution’. And indeed, if we look to games such as soccer, it seems we
have such rules:
(4.31)
Offside Rule
A player counts as being in offside if he is nearer to his opponents’ goal
line than both the ball and the second-last opponent (and is not in his own
half of the field of play).
Searle stresses that such counts-as rules are not mere definitions:
“‘offside’, ‘homerun’, ‘touchdown’, ‘checkmate’ are not mere labels for
[a] state of affairs [. . . ], but they introduce further consequences by way
of, e.g., penalties, points, and winning or losing.”
(Searle 1969, p. 36)
Counts-as rules are constitutive rules, then, because they are part of a larger set
of rules that specify normative consequences of something counting as something
else. Perhaps we could take this as a model, and simply say that ‘expressing a belief’
is something that is, in part, constituted by the rule governing declaratives:12
(4.32)
Declarative Rule (counts-as version)
When a speaker utters a declarative with content p, his utterance counts
as an expression of the belief that p.
We would not define ‘expressing a belief’, then, and would say that ‘expressing
a belief’ is not something that is independently defined—expressing a belief is
12
Of course, one can also read (4.32) as an empirical claim—but then we need to specify independently what ‘expressing a belief’ means, and (4.32) itself is of no explanatory value.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
79
simply the thing that one does when sincerely uttering a declarative.13
It should be clear, though, that no explanatory work is done by a rule like (4.32).
It will not tell us why speakers tend to utter declaratives only when they take their
contents to be true, and it will not tell us why speakers of declaratives are subject
to reprobation if they fail to believe their contents to be true, or why subsequent
utterances of declaratives must be consistent, or made consistent by means of
retraction. This work must be done by additional rules that put constraints on
expressions of belief. The same is true of the offside rule in (4.31), which does not
tell us what the consequences of being in offside are. It needs to be combined with
other rules of soccer that spell out the normative consequences of certain actions
performed in an offside position.
That is, on the ‘counts-as’ rule approach hypothesized above, the explanatory
work has to be done by normative rules like the ones discussed in the next section.
4.5
Normative theories of clause typing
Neither Lewis-conventions, nor the counts-as, or constitutive rules investigated in
the previous sections are, by themselves, normative. In order to explain apparentlynormative facts about language use, they have to be paired with normative constraints that are motivated independently.
There is an obvious alternative: Assume that the form–force mapping itself is
governed by properly normative conventions (or perhaps simply by ‘norms’—I
shall use the two terms interchangeably).
This assumption is, or at least can be, independent from the thesis that ‘meaning
is normative’ (as claimed, e.g., by Boghossian (1989)). Saying that what language
13
This is not quite Searle’s view, though. His essential rule for assertions in Searle (1969) is the
somewhat opaque:
(i)
[The act] counts as an undertaking to the effect that p represents an actual state of affairs.
(Searle 1969, p. 66)
In Searle (1989), he says that ‘a statement is a commitment to the truth of the [stated] proposition’—
indicating that anything that counts as such an undertaking gives rise to such a commitment.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
80
users do with a sentence S that has truth-conditions φ is constrained by norms does
not amount to saying that the relationship between S and φ itself is normative, or
determined by normative facts (nor does it deny this).
Normative theories of the form–force mapping come in two kinds: On the
one hand, there are theories that take the relevant norms to specify normative
preconditions on when a sentence can be used. On the other hand, there are theories
that specify normative consequences of utterances.
4.5.1
Normative preconditions on utterances
Accounts of this type assume normative conventions that specify when it is appropriate to utter a sentence of a particular type. If a speaker utters a sentence
in violation of these conventions, he is in the wrong. Stenius’ rule for declarative
mood can be seen as an early version of such a theory:
(4.33)
Produce a sentence in the indicative mood only if its sentence-radical is
true.
(Stenius 1967, R 3, p. 268)
We find several variants on this in the philosophical literature on assertion. Here
I shall draw mainly on Williamson (1996), who proposes to understand assertion
as governed by a normative rule and MacFarlane (2011), who provides a useful
survey, and critical assessment of, various families of philosophical theories of
assertion. Williamson (1996) himself defends the knowledge rule of assertion, which
is restated in (4.34) as a rule about declarative utterances:
(4.34)
Knowledge rule
One must: Utter a declarative with content p only if one knows that p.
Williamson also discusses a Truth rule which in terms of declarative sentences
amounts to Stenius’ rule above.
MacFarlane (2011) adds the Reasonable-to-
Believe Rule, variants of which are defended by Douven (2006) and Lackey (2007).
I again restate it in terms of declarative utterances:
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
(4.35)
81
Reasonable-to-Believe Rule
One must: Utter a declarative with content p only if it is reasonable to
believe that p.
It seems intuitively obvious that the Truth and Knowledge rules (on the assumption that knowledge requires truth) are too strong. Adopting them would mean
that a speaker who utters a declarative p is at fault if p is false, even if the speaker
had good reason to believe it is true. Such a speaker is in error, but it does not
seem right to say that he has violated a rule. To be sure, we sometimes, perhaps
often, criticize speakers for asserting something that is false, but in these cases, we
generally either allege that the speaker did not believe what he asserted (and hence
criticize him for asserting something he did not believe), or we claim that he should
not have believed what he asserted.
This does not necessarily mean that these rules are untenable: One could argue
that if a speaker honestly but erroneously believed he was uttering a true declarative
(or one that he knew to be true), then we are generally willing to forgive the rule
violation because he thought he was in compliance with the rule. I shall not argue
about the merits of this move, but for our purposes, this understanding makes the
Truth rule equivalent to the Reasonable-to-Believe rule, and the Knowledge rule
equivalent to the rule in (4.36):
(4.36)
One must: Utter a declarative with content p only if it is reasonable to
believe that one knows that p.
But then, I think we can safely focus on the Reasonable-to-Believe rule (as any
additional constraints on knowledge will not play a role in what is to follow). How
does the rule do with respect to our explanatory targets, repeated in (4.37)?
(4.37)
a.
Declaratives are well-suited to express beliefs, and thereby to inform.
b.
Declarative utterance make claims, which are subject to reprobation.
c.
Utterances of multiple declaratives with inconsistent contents require
retraction.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
82
Presumably, we can assume that a speaker who strives to obey the Reasonable-toBelieve rule will mostly only utter sentences that he believes to be true, explaining
why declaratives are well-suited to express beliefs. And quite obviously, (4.37b) is
directly accounted for by the Reasonable-to-Believe rule. But it is not clear that
(4.37c) is explained under this rule.
What is reasonable to believe surely can change over time, even within the
space of a single conversation—why, then, do incompatible declaratives require
retractions, if they are governed by a rule that says that the speaker should only
utter declaratives that he has good (enough) reason to believe? Should a declarative
utterance that contradicts a previous one not simply be understood as indicating
that a speaker has changed his mind? Why does he need to acknowledge this
inconsistency?
The Truth and Knowledge rules might be seen to fare better on this count,
implausible as they may be on other grounds: If the speaker changes his mind
about the truth of a previously-uttered declarative, this means he realizes that his
previous utterance constituted a rule violation. Perhaps it is simply necessary
to acknowledge this fact. This would be essentially equivalent to assuming an
independent normative rule for retraction, such as the following two hypothesized
by MacFarlane (2011):
(4.38)
Retraction rule 3:
One must: retract a previous assertion A when one knows that one performed A and that the content of A was untrue.
(4.39)
Retraction rule 4:
One must: retract a previous assertion A when one performed A and A
was untrue.
But these rules prove too much: It is not in general the case that one must retract all
utterances of declaratives (or all assertions) that one realizes were in error. Suppose
that you want talk to John, who takes part in a all-day conference. Misremembering
the conference schedule, I sincerely utter (4.40).
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
(4.40)
83
The last coffee break starts at 5pm.
In actual fact, the last coffee break starts at 4:30 and ends at 5. But my mistake is
harmless, because you have another appointment at that time anyway, and decide
to bother John during the lunch break at 12. If I realize my mistake the day after,
there may be no need to retract my assertion in (4.40). If you realize the mistake,
you might question me about why I informed you wrongly, and if my mistake
resulted in you or John being inconvenienced, I might apologize, but it is not the
case that I violate a rule if I don’t. In general, it seems that I need to retract only
those utterances of declaratives that are such that their truth or falsity are still
relevant—or, if they are no longer relevant, but I still have reason to talk about the
actual facts.
We could, of course, adapt the retraction rules to something like (4.41).
(4.41)
Retraction rule 4:
One must: retract a previous assertion A when one performed A and A
was untrue and A is still relevant; or if one has reason to assert something
that is incompatible with A.
But with this, it seems that the retraction rule has become an important part of the
specification of the effect of assertions/utterances of declaratives. The retraction
rule states a normative effect of uttering declaratives, and in doing so, it supplies
something that intuitively was missing in the statement of the Reasonable-ToBelieve rule. In general, it seems that rules about normative preconditions for
utterances need to be complemented by rules that govern their normative effects,
i.e., rules that specify what the speaker’s behavior should be like after the utterance,
not only what should be the case prior to or at the time of the utterance.
4.5.2
Normative effects of utterances
Given that we need to specify normative effects of utterances anyway, we may
consider the possibility that clause-typing conventions are about normative effects
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
84
in the first place. Maybe we then can dispense with the assumption of rules specifying normative preconditions at all, or we can assume that declarative utterances
are subject to both rules about their preconditions and rules about their normative
consequences.
MacFarlane (2011) summarizes the essential features of an account (of assertion)
in terms of normative consequences:
“[T]his approach defines assertion in terms of its “essential effect.” But
it regards this essential effect as the alteration of a normative status—the
acquisition of new commitments or obligations.
It is important to see how the commitment approach differs from “constitutive rules” approach we considered above. Both describe assertion
in essentially normative terms. But, while the constitutive rules approach looks at “upstream” norms—norms for making assertions—the
commitment approach looks at “downstream” norms—the normative
effects of making assertions.”
(MacFarlane 2011, P. 91)
What kinds of commitments could be the result of utterances of declaratives? In
authors like Searle (1969, p. 29) we find locutions like ‘commitment to the truth of
p’, indeed “a (very special kind of) commitment to the truth of a proposition”. But
this is not helpful. If our account of the normative effects of declarative utterances
is to have any bite, we must specify what ‘commitment to the truth’ amounts to.
MacFarlane (2005) usefully summarizes various proposals that have been made in
the context of analyses of Assertions:
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
(W)
85
Commitment to withdraw the assertion if and when it is shown
to have been untrue.
(J)
Commitment to justify the assertion (provide grounds for its
truth) if and when it is appropriately challenged.
(R)
Commitment to be held responsible if someone else acts on
or reasons from what is asserted, and it proves to have been
untrue.
(MacFarlane 2005, p. 318)
I shall discuss these in turn, before offering my own conception of the normative
effects of utterances of declaratives, which I have developed jointly with Cleo
Condoravdi (Condoravdi and Lauer 2011, 2012).
Commitment to withdraw
(W) is similar to the retraction rule which I argued to be too strong above, but it
is weaker: It only requires retraction if the assertion is shown to have been untrue.
This may well be weak enough to not require retraction of irrelevant falsehoods,
but it is also far too weak to account for anything but the need to retract in response
to overt demonstrations that one’s utterance was false.
Perhaps, if we take ‘shown to have been untrue’ as weak enough, if we allow
that the obligation to retract does not rely on conclusive demonstration, but only
on reasonable doubt, we might be able to explain why speakers tend to utter
declaratives they take to be true, and hence why declaratives are good for conveying
information, if we assume that speakers have a preference against retracting their
own utterances.
But in any case, it seems that we want (W) to be a result of our account of the
uses of declaratives, rather than an independent stipulation. There is no direct
connection between (W) and the notion of belief, and it only partially explains
the need for retraction (since it does not explain why a speaker who decides, for
himself, that he now would like to assert the opposite from what he asserted earlier
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
86
must retract). (W) could be an independent component of the commitments we
undertake when uttering declaratives, but it would be preferable if we can derive
it from more basic assumptions about declarative utterances.
Commitment to defend
(J) is repeated below:
(J)
Commitment to justify the assertion (provide grounds for its truth) if and
when it is appropriately challenged.
This view has been held (for assertions), for example, by Brandom (1983). It has
the potential to explain why declaratives are well-suited to expressing beliefs, even
in cases in which there is conflict of interest—any assertion of something that one
does not believe would make one liable to mount a futile defense. And presuming
that incompatible claims cannot be jointly defended, such a view might be seen as
predicting the need for retraction in case of assertions with incompatible contents.
But the idea that utterances of declaratives commit their speakers to defend the
content when challenged seems to presuppose too special a kind of conversational
context to be generally applicable. MacFarlane (2005) articulates this worry well:
“[T]his may be over-generalizing from seminar-room assertions to assertions in general. Suppose someone were to say: ‘You’ve given some
very good reasons to doubt the truth of what I asserted. I have nothing
to say in answer to your objections, yet I continue to stand by my claim.’
She would not be playing the game of assertion the way philosophers
play it, but perhaps philosophers do not get to set the rules here. We
would surely take her assertions less seriously than we would if she
were responsive to reasons. But would we cease treating her as an
asserter at all? That is not so clear.”
(MacFarlane 2005, p. 318)
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
87
It is true that in many situations—familiar not only to philosophers but to any
academic—we expect that speakers are willing and able to defend their assertions,
and we think less of them if they cannot or won’t. But it does seem too strong to
say whenever someone fails to defend a claim that is challenged (even if he gives
responses much less polite then the one imagined by MacFarlane—e.g., Get lost,
I won’t argue with you or I won’t stand being questioned by you or I believe it,
regardless of what you say, etc.) has thereby violated a commitment.
Note also that while this conception allows us to understand why speakers
would tend to retract assertions if they want to assert something that is incompatible
with them, the assertion of two inconsistent propositions would in itself not be a
violation of the commitment—but it seems that if someone makes two incompatible
assertions—at least when he does so knowingly—he has thereby, with his second
assertion, already violated the commitment he took on with the first. But this does
not follow from (J).
So (J), again, appears to only partially capture the commitments we take on
with utterances of declaratives, and it also appears to predict a commitment that is
often not intuitively present.
4.5.3
(R)
Commitment to be held responsible for consequences
Commitment to be held responsible if someone else acts on or reasons from
what is asserted, and it proves to have been untrue.
Such a view is evident, in particularly strong form, in von Savigny (1988)’s characterization of constatives:
“A conventional make-up which is fundamental for constatives is that
at x’s expense, y can rely on p obtaining; this means that x is liable
for compensating all damage which results for y from y’s erroneously
relying on the fact that p.”
(von Savigny 1988, p. 51)
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
88
In this form, the requirement simply seems to strong—and indeed, if any utterance
of a declarative incurred such a strong commitment, speakers would not utter them
as frequently.
There are two possible responses: On the one hand, we might say that speakers
actually take on this strong commitment, but it is rarely enforced. This seems
to be a non-starter, because there are many cases where a sincere apology is all
someone can ask for, even if one relied on someone’s false assertion and had to
suffer negative consequences because of it. On the other hand, we might broaden
the notion of ‘compensation for damages’ so that something like an apology can
serve as sufficient compensation. Or, perhaps, we simply can say that all one
commits to is taking partial responsibility. This is what MacFarlane (2005) seems to
have in mind:
“A plausible answer (though not the only one) is that part of what it is
to make an assertion is to accept partial responsibility for the accuracy
of what one says.”
(MacFarlane 2005, p. 319)
This version is not only weaker in that it only claims partial responsibility, it also
does not tie the responsibility to negative effects of the hearer’s action (though
perhaps such negative effects can be generally assumed if the hearer’s action was
based on a false belief).
But it still seems too strong. Suppose, for example, both you and I observe
the same evidence, which convinces us both (individually) that Mary was at some
party. In our discussion of the evidence I utter (4.42), and you concur. Later it turns
out that Mary in fact was not at the party at all, but rather she was studying. Given
that we reasoned towards the conclusion that Mary was at the party together, it
seems that you cannot hold me responsible for any harm that came your way due
to your assumption that Mary was at the party.
(4.42)
Mary was at the party.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
89
Perhaps (4.42) is not an assertion in the sense that philosophers like MacFarlane
(2011) are after, but it surely is a sincere utterance of a declarative. Perhaps we could
say that by overtly agreeing (instead of just accepting) my utterance in (4.42), you
absolved me from responsibility (cf. Gunlogson (2008)’s distinction of accepting
and confirming declarative utterances). Or maybe we could say that the speaker
is responsible only insofar as the addressee’s reliance on the truth of the content of
the utterance was due to the utterance.
But what, then, about the following case: I have heard (but you did not) from
someone who both you and I usually trust as an authority about Mary’s whereabouts (or about parties) that Mary was at the party; or perhaps I have other
evidence that you and I would take to be conclusive of Mary’s being at the party,
even though you have not observed this evidence. When you wonder where Mary
was yesterday, I utter (4.42). You accept my utterance, and get in a vicious fight
with Mary, because you believe that she violated her duties by partying. It turns
out that I was wrong. Am I liable, or to blame, for the adverse consequences to
your relationship with poor Mary? I don’t think I am. It seems that I would be
well in my rights to express sympathy for your situation, but maintain that I only
relayed information that I had excellent reason to think was true, and hence am
entirely blameless.
4.5.4
Commitments to act as though one has a belief
Both (J) and (R) seem, in a way, too parochial: They capture intuitions about the
obligations speakers take on in certain situations (philosophical discussion, and
business deals, say), but they don’t quite get at the heart of the matter: Declaratives
are uttered in many contexts in which these effects do not seem to be present, or
seem to be present only to a rather attenuated degree. Crucially, in such contexts,
declaratives are often still very useful for expressing beliefs and to inform, and in
these contexts, too, speakers are required to retract their utterances.
Condoravdi and Lauer (2011, 2011) propose a more general normative effect
for utterances of declaratives. The idea is somewhat similar to the notion that
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
90
declaratives ‘express’ beliefs: A declarative utterance with content p commits its
speaker to act as though he has the belief that p
(4.43)
Declarative Convention (normative)
A speaker who utters a declarative with content p thereby becomes committed to act as though he believes p to be true.
That is to say, if a speaker utters a declarative, he puts on himself an obligation to
make only action choices that cohere with the belief that p (whether he actually has
this belief or not). I will elaborate this proposal in the next section.
4.6
Commitments to act according to an attitude
The idea is this: Intentional actions are chosen by the agent performing them.
Choosing an action means selecting it from a set of possible alternative actions
(minimally: performing the action or not performing it), where this choice is (at
least partially) determined by a form of practical reasoning about how to best satisfy
one’s preferences, based on one’s beliefs.
I choose to grab my umbrella as I am leaving the house (rather than not grabbing
it), because I believe that it is raining, or may be raining later in the day while I am
out, and because I also have a preference against getting wet, and another belief
that having an umbrella will enable me to stay dry in the rain. We may say that
given this set of beliefs and preferences (together with some others, perhaps, such
as that there will be no other way to procure an umbrella, and no other practical
way to stay dry), practical reasoning tells me that it is better to grab my umbrella
than not to. Idealizing a bit, we can say that choosing an action means performing
the best action, or one of the best actions, from a set of alternatives, based on a
given set of beliefs and preferences. Acting as though one has a belief that p, then just
means choosing only those actions that are best with respect to a set of beliefs that
includes p.
This will mean that a single commitment to a belief, by itself, will not completely
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
91
determine most action choices. If I sincerely utter (4.44), and thereby commit myself
to act as though I believe that it will be raining, this is compatible with me not
grabbing my umbrella. I may lack a preference for not getting wet, for example,
or I may also believe that I will be in my office before it begins to rain, and that I
won’t have to leave it before it stops raining. And so on.
(4.44)
It will be raining this afternoon.
This is desirable: If I only said (4.44), and you criticized me for not grabbing
my umbrella, then informing you that I don’t mind being in the rain without an
umbrella, or that I don’t plan to be outside all afternoon, etc. is a sufficient response.
However, my commitment to the belief that it will be raining will constrain
my action choices, in particular together with other commitments I have. For I
may be committed (or commit myself later) to other beliefs, and to having certain
preferences. If I am committed to the belief that I will have to go outside in the
afternoon, and I am committed to a preference for not getting wet, justifying my
not taking the umbrella as being in accordance with my commitments will get
harder. So the accumulation of commitments in dialogue can ultimately enable
the audience to make rather reliable predictions about how I will act in the future
(at least to the extent they take me to honor my commitments). Perhaps such
predictions will never be quite certain—I might always be able add an unexpected
belief or preference to subvert an action choice that otherwise would be mandated
by my commitments—but the more commitments I incur, the more reliable my
audience’s predictions of behavior will become.
How does this conception—that declaratives commit the speaker to behave as
though he had a belief—fare with respect to our three explanatory targets? The
following subsections argue that it does well. Section 4.6.1 discusses (4.44c), Section 4.6.2 discusses (4.44a) and Section 4.6.3 discusses how various other normative
effects, includes (4.44b) can be accounted for.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
4.6.1
92
Declaratives with incompatible contents
It might be thought that the need for retraction in case of incompatible contents
directly follows from the definition of the basic effect of declaratives—surely, committing to a belief that
p is not a way to act as though one believes that p? I am
indeed on record with saying this (Lauer 2012, p. 395), but I now realize that it is
not quite correct.
In Section 4.7, I will argue that it can be optimal for a speaker to assert p even
if he does not believe it—but if that is the case, then committing to a belief that
p
can be a way to act as though one believed that p.
However, the proposed conception of declarative force does predict the need
for retraction. While it may be rational to commit to a belief that one does not have,
it is never rational to commit to two contradictory beliefs, which would amount to
a commitment to act irrationally. So we explain the need for retraction not because
committing to
p is never a way to act as though one believed that p, but rather
because committing to both p and p is never a sensible way to act (and so is not a
sensible way to act as though one believes that p).
4.6.2
Declaratives express beliefs
This conception also directly explains why utterances of declaratives are (usually)
reliable indicators of belief: While it may be, on occasion, advantageous for an
agent to commit to a belief he does not actually have, any such a commitment will
involve a risk. A speaker who does so risks committing himself to acting against
his best interests.
For suppose a speaker believes that p, but commits himself to
p. At a later
time, it turns out that, if p is true, that necessitates a certain action a, while if p is
false, an alternative action a1 would be optimal. The agent is committed to perform
a1 instead of a, even though his actual beliefs tell him that a would be better.
In such a case, of course, the speaker will not necessarily be forced to do a1 .
He might instead be able to rescind his commitment to
p. Or he simply chooses
to not act in accordance with his commitment, and accept the blame for doing
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
93
so. But, by and large, it seems plausible that agents have incentive to abide by
their commitments (if only to remain credible in the eyes of their peers) and also
a weaker incentive to not rescind commitments all too frequently (for the same
reason).
This means that, by and large, an audience has good reason to think that a
speaker will commit to a belief that p only if he in fact believes that p. There will
be exceptions—for example the speaker might not quite believe that p, but take it
very likely that p obtains. Or it might be that the speaker believes that q, which
is incompatible with p, but which necessitates all the same action choices as p in
future situations he takes to be plausible. But in general, a speaker who utters
declaratives that he believes to be true will be on the safe side, while a speaker who
does not will engage in risky business.
For now, this informal sketch will suffice. Chapter 5 will be devoted to extending
the system of dynamic pragmatics in such a way that we can model explicitly
how an addressee infers that a speaker believes in the truth of his declarative
utterances—based on the contextual assumption that the speaker prefers not to
commit himself to beliefs he does not have.
4.6.3
Other normative consequences of declaratives
There is also some reason to think that the present conception has a role in explaining why some of the other proposals for normative consequences appeared
promising—by explaining how, in special contexts, committing to act as though one
has a certain belief commits one to certain courses of actions that can be described
in other terms.
We have already seen that the present conception explains the need for retraction
in case a speaker wants to make a new claim that is incompatible with his previous
ones. This extends directly to cases in which the speaker accepts someone else’s
arguments as to why his previous utterances is false. This is what (W) mandated.
Here is what MacFarlane says about this principle:
“Everyone should be able to agree that assertoric commitment includes
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
94
at least (W). Imagine someone saying: ‘I concede that what I asserted
wasn’t true, but I stand by what I said anyway.’ We would have a
very difficult time taking such a person seriously as an asserter. If she
continued to manifest this kind of indifference to established truth, we
would stop regarding the noises coming out of her mouth as assertions.
We might continue to regard them as expressions of beliefs and other
attitudes (just as we might regard a dogs whining as an expression of a
desire for food). We might even find them useful sources of information.
But we would not regard them as commitments to truth.”
(MacFarlane 2005, p. 318)
On the present conception, it is clear why MacFarlane’s unfazed asserter would be
viewed as doing something wrong: by saying I concede that what I asserted wasn’t
true, but I stand by what I said anyway, not only does the speaker not rescind his
previous assertion, he doubles down on it, while at the same time admitting that it
was false, and thereby committing himself to its falsity. But that just means that he
commits to inconsistent beliefs.14
Similarly, reconsider (J):
(J)
Commitment to justify the assertion (provide grounds for its truth) if and
when it is appropriately challenged.
14
There still is an aspect of (W) as originally formulated that the present conception does not
directly cover. The present perspective predicts that it is not possible (or at least not rational)
to accept a conclusive argument that his prior assertion was false without withdrawing it. But
(W) is stronger. It demands withdrawal whenever it is conclusively shown that one’s previous
assertion was false, independent of whether one is willing to accept this publicly. Put differently,
(W) implicitly mandates such public acceptance, while on its own, the present perspective does not:
A speaker is well within his rights to not retract his previous assertion as long as he is willing to
stick to his guns and refuse to publicly accept that he has been proven wrong.
I don’t think this is problematic: All we need to assume is that, at least in certain contexts, such as
in rational debate, a speaker is independently required to accept such conclusive demonstrations of
the falsity of his utterances. And I think that this is intuitively correct: A speaker who steadfastly
refuses to back down from a previous assertion that has been conclusively proven wrong is not
wrong because he fails to honor the commitments he made with his assertion. He is wrong because
he fails to correctly respond to the proof that it was wrong.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
95
I concurred with MacFarlane that it seems that this kind of commitment only applies
in special cases, such as discussions among philosophers and other academics (or
more general, inquiry). In the present context, we could explain the existence
of the commitment in (J) by assuming that philosophers, etc., when engaged in
discussions on their field of inquiry, are subject to an independent commitment
to only believe (or commit to believe) things that they are able to defend against
reasonable challenges. Together with the doxastic commitment created by an
assertion, this would predict that a speaker, in such a context, is committed to utter
only declaratives he is able to defend.
We can make similar assumptions about other cases where there are requirements on truth or epistemological warrant that are more stringent than in ordinary
conversation. As an example, let us take witness testimony in a court of law: We
can say that in this situation, a speaker is obligated to commit only to propositions
he knows to be true (or perhaps: he is obligated to commit only to propositions that
are true). That is, the commitment-based view, just as the expressive view, provides us with a ‘hook’ onto which further requirements, or moral, legal and social
obligations can be hung. These requirements can pertain to what commitments an
agent is allowed to take on.
4.7
The enduring commitments of loose talk: a case
for the commitment-to-belief view
In the discussion of alternative views, I have occasionally remarked that, unlike
the commitment-to-belief view articulated in the previous section, the alternative
views by themselves predict the necessity of retraction only if this can be motivated
on communicative grounds. In this section, I want to reprise the arguments made
in Lauer (2012) to the effect that this provides an argument for the commitment-tobelief view.
In many cases (in particular typical instances of cooperative conversation which
tend to be the focus of researchers on clause typing and speech act theory) almost
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
96
any of the alternative accounts can predict the need for retraction, given appropriate
assumptions about cooperative behavior, essentially because all of these alternative
accounts can, given appropriate assumptions, explain why declaratives are good
for informing or expressing beliefs.
If a speaker now utters a declarative with content p that is incompatible with
his previous assertions, he will often have reason to indicate that he has changed
his mind (which he can do be retracting his previous assertion): He may want to
make sure that his audience knows he had good reason for his previous utterance,
he may want to alert them to the change in his doxastic state for other reasons. If
he did not make this explicit, his audience might think he has forgotten that he
made his previous utterance, and has forgotten the reasons why he made it, and
might hence fail to believe his current utterance. That is, if he previously communicated p, retracting this utterance will frequently be instrumental in successfully
communicating p.
I am not certain that such a communication-based account of retractions could
be spelled out in detail in such a way that it makes the correct predictions in all cases
in which p is communicated prior to an instance where p was communicated. But
I am also not certain that it could not be spelled out in this way. After all, we
generally expect that belief revision is not the norm. We generally expect that
speaker’s beliefs persist, so maybe a case can be made that if a speaker changes his
mind, he ought to (generally) indicate this, either because failing to do so would
jeopardize his communicative goals, or because there is a social or moral norm
requiring him to do so.
What I am certain about, however, is that this would not be enough. At best,
such a communication-based account would predict that a speaker has to retract
a previous assertion whose content was believed by his audience. But as we shall see,
a speaker has to retract even utterances he knows have not been believed. What I
have in mind are not utterances made in a situation in which the speaker was not
believed because he was taken to be unreliable. Here, I want to focus on cases of
perfectly cooperative communication.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
4.7.1
97
Loose talk—the basic facts
Cooperative conversation may well involve utterances of declaratives that the
speaker knows to be false, or at least, it would seem that way. Some examples of
the phenomenon I have in mind are in (4.45).15
(4.45)
Mary was here by three.
Fact: Mary arrived at 3:02.
(4.46)
I am going to a conference in Berlin next week.
Fact: The conference is Potsdam, which abuts Berlin, but is not part of it.
(4.47)
There were five hundred people at the rally.
Fact: There were exactly 489 people at the rally.
At least intuitively speaking, these sentences, in the indicated contexts, are false
(though they ‘come close to being true’). And yet, there are many contexts in
which a cooperative speaker who knows the indicated facts could use them. And
such a speaker would not be at fault. This is what Lasersohn (1999) calls Loose
Talk, or Pragmatic Slack.
The philosophical literature provides a similar example, (4.48):
(4.48)
France is hexagonal.
This example was introduced by Austin, who denied that, taken as a statement,
(4.48) is true or false—it is simply ‘rough’:
“Suppose that we confront ‘France is hexagonal’ with the facts, in this
case, I suppose, with France, is it true or false? Well, if you like, up
to a point; of course I can see what you mean by saying that it is true
15
(i)
(4.45) is a slight variant of an example that Lasersohn gives:
Marry arrived at three.
I am no longer entirely convinced that Lasersohn was right in supposing that (i) involves ‘loose talk’
rather than truth-conditional vagueness. But his basic point is not affected if he has misclassified
this example. (4.45) more clearly behaves in the way he has described.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
98
for certain intents and purposes. It is good enough for a top-ranking
general, perhaps, but not for a geographer. ‘Naturally it is pretty rough’,
we should say, ‘and pretty good as a pretty rough statement’. But then
someone says: ‘But is it true or is it false? I don’t mind whether it is
rough or not; of course it’s rough, but it has to be true or false—it’s a
statement, isn’t it?’ How can one answer this question, whether it is true
or false that France is hexagonal? It is just rough, and that is the right
and final answer to the question of the relation of ‘France is hexagonal’
to France. It is a rough description; it is not a true or a false one.”
(Austin 1962, p. 142)
Subsequent authors have taken a somewhat different view, saying that the
sentence in (4.48) is not true or false simpliciter, but that it depends on the context
whether it is—that is, they have argued that the sentence (and hence, presumably,
the predicate hexagonal) is vague: Whether or not we take it as true depends on a
contextually-given standard of precision. This is the view taken by Lewis (1979).
Lasersohn (1999), however, convincingly argues that this cannot quite be correct,
or at least, it cannot be correct in all cases in which we feel that the speaker only
‘comes close enough to the truth’ with what he says. Because in many such cases,
such as those cited above, there is good evidence that what the speaker said is
literally false, while proper truth-conditional vagueness does not result in falsity.
Thus we can conjoin a truly vague predicate like bald with a claim that there are
exceptions.
(4.49)
a.
Homer is bald: he has, like, three hairs left.
b.
Homer is bald, but he has a few hairs left.
This strongly indicates that Homer is bald can be literally true if Homer has some
hairs left. By contrast, statements such as Mary was here by three do not allow for
such conjunctions.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
(4.50)
99
#Mary was here by three, but she did not arrive until a few minutes after
three.
If Mary was here by three were truth-conditionally vague in the sense that Homer
is bald is, the former sentence could be true even if she arrived shortly after three
and (4.50) should be felicitous in such a context, but it is not. Lasersohn’s conclusion is that loose talk is different from truth-conditional vagueness: It involves
sentences that have context-independent truth-values, which yet can sometimes be
‘blamelessly asserted’ if they come close enough to the truth for practical purposes.
4.7.2
Loose talk in multi-sentence discourses
Turning our attention to multi-sentence (and multi-turn) discourses, we find that
an assertion that is only loosely true cannot be felicitously followed up by a more
precise one without further ado:16
(4.51)
(Yes,) Mary was here by three. # When she wasn’t here five minutes after
three, I got worried, but then she arrived in time.
(4.52)
A: I am going to this conference in Berlin.
B: Oh, where in Berlin is it held?
A: #In Potsdam(, which is just outside of Berlin).
(4.53)
A: Do we have enough coffee for the council meeting? How many people
will be there?
B: Thirty.
A: Oh, great, then we’ll have a quorum.
B: #No, because only 27 people will be there.
These discourses improve dramatically if we insert an expression that acknowledges the falsity of the previous assertion, such as actually:
16
This short section is adapted from the one with the same title in Lauer (2012).
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
(4.54)
100
(Yes,) Mary was here by three. Well, actually, she wasnt here until five
minutes after three, and I got worried, but then she arrived in time.
(4.55)
A: I am going to this conference in Berlin.
B: Oh, where in Berlin is it held?
A: Actually, it is in Potsdam(, which is just outside of Berlin).
(4.56)
A: Do we have enough coffee for the council meeting? How many people
will be there?
B: Thirty.
A: Oh, great, then we’ll have a quorum.
B: No, because (I was speaking loosely and) actually, only 27 people will
be there.
At first glance, this may seem unsurprising: After all, we have just seen that there is
good reason to believe that the first assertion in all these examples is literally false
if the second one is true. So, of course, a speaker who wants to make the second
assertion has to acknowledge this contradiction. Indeed, if the first assertion is
appropriately hedged, no such acknowledgement is necessary:
(4.57)
(Yes,) Mary was here around three. When she wasn’t here five minutes
after three, I got worried, but then she arrived in time.
(4.58)
A: I am going to this conference in the Berlin area.
B: Oh, where is it held?
A: In Potsdam(, which is just outside of Berlin).
(4.59)
A: Do we have enough coffee for the council meeting? How many people
will be there?
B: About thirty.
A: Oh, great, then we’ll (likely) have a quorum.
B: No, because only 27 people will be there.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
4.7.3
101
Loose talk in the commitment-to-belief account
Under the commitment-to-belief account of the sentential force of declaratives,
we can construe cases of loose talk as instances where a speaker commits to a
proposition that he knows to be false, or that he at least does not know to be true.
Thus, in uttering (4.60) while knowing or suspecting that Mary arrived shortly
after three, a speaker still commits himself to act as though she was here by 3:00:
(4.60)
Mary was here by three.
Why would a speaker ever do this? He might have reason to do so, if he believes
that the exact time of Mary’s arrival is irrelevant to his future (linguistic and nonlinguistic) action choices. That is, if he does not expect to be in a future decision
situation in which there are two actions a and a1 such that if Mary arrived at 3:00, a
is optimal, but if she arrived slightly after 3:00, a1 would be optimal. If there is no
such future decision situation, then acting as though one believes that Mary was
here by three is true is just the same as acting as though one believes that Mary was
here shortly after three is true.17
Now suppose, contrary to the speaker’s expectation, the conversational purposes shift so that more precision is required: Suddenly, it does make a difference
whether Mary arrived at three or a couple of minutes after three. Given that the
speaker has committed to a belief that she arrived at 3:00 sharp, we can explain
why a speaker needs to acknowledge the falsity of his previous utterance.
(4.61)
Actually, she wasn’t here until a few minutes after three.
What actually does in this case, on the commitment analysis, is rescind the previous commitment taken on with (4.60), at the same time, (4.61) induces a new
commitment to the effect that the speaker believes Mary arrived a few minutes
17
An additional consideration such a speaker needs to take into account is whether his audience
knows he is speaking loosely. If they take him to speak strictly—i.e., if they infer from (4.60) that
Mary was here by 3:00, then a cooperative speaker may want to prevent this misunderstanding,
and opt for a locution such as around three. He may, or he may not: If he does not expect that a
couple of minutes of difference makes a difference to his audiences’ future action, he may still opt
to speak loosely.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
102
after three.
By contrast, an utterance like (4.62), without actually or something that would
do its job, would only create a new commitment, but leave the prior commitment
intact. This is why (4.62) feels infelicitous after (4.60), even if the speaker was
speaking loosely: He takes on incompatible commitments with his two utterances.
(4.62)
She wasn’t here until a few minutes after three.
It should be stressed that the commitments we are talking about here are, in most
normal contexts, very ‘cheap’: The speaker can easily rescind them.18 But rescind
them, he must, if he wants to make another assertion that is incompatible with
them.
4.7.4
Communicative reasons for retraction
The prediction of the commitment-to-belief account is that the need for retraction is
independent of the communicative goals of the speaker: It does not matter whether
retraction is necessary to avoid a misunderstanding. Retraction is independently
necessitated by the fact that declarative utterances give rise to commitments.
Let us reconsider the Potsdam example in a context in which a certain amount
of slack is expected.19
(4.63)
A: I am going to this conference in Berlin.
B: Where is it held?
A: # In Potsdam, which is just outside of Berlin.
It is noteworthy that B’s question does not indicate that he took A to speak strictly,
and also that the continuation explicating that Potsdam is outside of Berlin does
18
How ‘cheap’ rescission is will depend a lot on the context of utterance—if the context was such
that precision was known to be necessary from the outset, rescission will be much harder than if
initial demands were loose.
19
It is helpful to assume that the conversation takes place somewhere far away from Berlin—in
the US, say. If the interlocutors are close to, or in Berlin or Potsdam, it is implausible to assume that
the speaker is speaking loosely, for reasons I cannot go into here.
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
103
not remove the sense of incoherence (if anything, it strengthens it). So, in this
example, there is no communicative reason to retract: Even if the speaker does not
retract, there is no danger that the addressee will misunderstand the speaker, and
assume (for example) that Potsdam is part of Berlin.
To put this differently: Given that the speaker can be assumed to speak loosely,
what A intends to convey with his initial utterance, and what B comes to believe
if he takes A to be otherwise sincere and well-informed is something like the
proposition expressed by (4.64).
(4.64)
A is going to a conference in the Berlin area.
If all goes smoothly, (4.64) will be common belief after A uttered I am going to this
conference in Berlin. A used this sentence to inform B that (4.64) is true, and he
succeeded (let us assume). All parties to the conversation are in agreement that
this is what happened. Why, then, can the speaker not act as though he had uttered
(4.64), in which case no retraction is necessary?
(4.65)
A: I am going to this conference in the Berlin area.
B: Where is it held?
A: In Potsdam, which is just outside of Berlin.
The commitment-based account explains why the speaker cannot just behave as if
he made the ‘loose’ assertion in (4.65): Because he did not, and regardless of the fact
that he communicated something weaker than the truth conditions of the sentence,
he committed himself to act as though he believes the strict truth conditions.
Any account that instead tries to explain retraction facts solely in terms of the
speakers’ communicative goals will not be able to explain the contrast between
(4.63) and (4.65) in a context in which slack is expected, because the two dialogues
are the same in terms of what is communicated. It is only on the level of what is
asserted that they differ.
To illustrate this another way: On the level of communicated content, a similar
sequence involving a vague predicate like bald is completely parallel to (4.63):
CHAPTER 4. THE SENTENTIAL FORCE OF DECLARATIVES
(4.66)
104
A: Homer is bald.
B: Has he any hairs left?
A: Yeah, like three.
Again, A’s first utterance communicates that Homer is close to being completely
bald, but not that he is completely hairless. A and B are aware of this. There is
no communicative reason, then, to retract his first utterance when asserting that
Homer has some hairs left. And in this case, because Homer is bald has vague
truth-conditions, the commitment is to the weak truth conditions, too, and there is
no need for retraction.
Loose talk provides us with an empirical argument to favor the commitmentbased approach over alternatives that do not predict that utterances of declaratives
induce commitment to the (strict) truth of their contents. Such accounts could conceivably explain the requirement for retraction where what is communicated is the
strict truth conditions, but they cannot explain why retraction/acknowledgement
is necessary when what is communicated is weaker than the truth-conditions of
the asserted sentence.
Chapter 5
Action Choice and Commitment
At the end of Chapter 4, I proposed that declaratives give rise to commitments to
act as though one believes the content of the declarative to be true and argued for a
normative convention of the form in (5.1).
(5.1)
Declarative Convention (normative)
=(4.43)
A speaker who utters a declarative with content p thereby becomes committed to choose his actions as though he believes p to be true.
Among other things, this convention is supposed to help us explain why declaratives (in contrast to interrogatives and imperatives) are generally useful for conveying information, i.e., why it is that, when a speaker utters a declarative $ ϕ, this
can lead to the addressee coming to believe that ϕ is true.
In Chapter 2, I developed a model in which we could show that an addressee
will come to believe the content of an utterance ϕ if his belief state satisfies the
conditions in (5.2).
(5.2)
Contextual assumptions: Trusting addressee
a.
‘Honest speaker’
BAd,t uttert pSp, ϕq ñ Sp,t pϕq
b.
‘Informed speaker’
BAd,t Sp,t pϕq ñ ϕ
105
=(2.34)
CHAPTER 5. ACTION CHOICE AND COMMITMENT
106
The problem raised by the existence of multiple clause types is that (5.2a) should not
hold for interrogatives and imperatives, but it should (usually) hold for declarative
sentences. That is, we now want to use the convention in (5.1), together with
plausible contextual assumptions, to derive (5.2a). The addressee-reasoning that
we want to capture is given informally in (5.3).
(5.3)
a.
[Convention]
When a speaker utters $ ϕ, he becomes committed to believe ϕ.
b.
[Contextual assumption]
The speaker does not want to be committed to believe ϕ unless he
actually believes it.
c.
[From (a) and (b)]
The speaker will decide to utter $ ϕ only if he believes that ϕ.
d.
[Belief about actions]
Since utterances are intentional actions, the speaker will utter $ ϕ only
if he decides to do so.
e.
[From (c) and (d)]
The speaker will utter $ ϕ only if he believes that ϕ.
This reasoning is rather straightforward on the surface, but there is much hidden
complexity. Besides reasoning about beliefs and commitments, it involves a premise
about desires or preferences, viz., (5.3b), and a conclusion about what the speaker
will decide to do, given his preferences, viz., (5.3c), as well as a seemingly innocent
assumption about intentional actions, viz., (5.3d).
If we want to capture such reasoning—if we want to explain how a convention
like (5.1) enables speakers to communicate that they believe in a proposition, our
system of dynamic pragmatics needs to be able to capture reasoning about how
speakers decide on (utterance) actions, based on their preferences and beliefs.
An additional benefit of representing action choice is that this will enable us to be
more explicit about the nature of normative conventions like (5.1). The convention
says that the speaker becomes committed to choose his actions in a certain way. In
order to give this notion content, we need to be able to talk about action choice.
CHAPTER 5. ACTION CHOICE AND COMMITMENT
5.1
107
Modeling action choice
At its most abstract, deciding what to do can be construed as selecting an action from a set of alternatives (minimally, doing a certain act or not doing it).
Which alternative an agent selects depends—besides the set of available action
alternatives—on what outcomes he wants to achieve, and what he believes about
the consequences of each of the actions. That is, we can abstractly characterize a
decision procedure as in (5.4).
(5.4)
A decision procedure is a function Opt that takes three arguments B, P and A;
where B and P are suitable representations of beliefs and preferences, and
A is a set of possible actions, such that OptpB, P, Aq Ď A.
This way of construing action choice is, of course, not a novel idea at all. It is
a familiar paradigm in many disciplines that are modeling how people choose
their actions. Most notably, this conception of action choice is the one of decision
and game theory. The most dominant approach there is to assume that beliefs
are specified as probability distributions, preferences are specified by means of
utility assignments, and Opt is defined in terms of the Expected Utility of the
actions. In other words, Opt returns the subset of A which contains only those
actions that maximize expected utility.
The central concern of these approaches is choice under uncertainty about the
consequences of the actions: Agents are taken to have preferences over fullyspecified outcomes, but are uncertain as to which action results in which outcome.
Expected utility theory and its alternatives are theories about how an agent makes
such a choice, given his probabilistic beliefs about the world.
However, the usefulness of the conception of action choice in (5.4) is not exhausted by these concerns. Indeed, in this dissertation, the conception of a decision
procedure that selects among actions, given certain input beliefs and preferences
will be very central indeed, while uncertainty about the outcomes of actions will
not play a central role at all.
CHAPTER 5. ACTION CHOICE AND COMMITMENT
108
Given a known Opt-function, we can use information about its output to draw
inferences about its inputs: Suppose that B1 and P1 are unknown, and for some set of
actions A, we are told that one of its members a is such that a P OptpB1 , P1 , Aq. Then,
using our knowledge of Opt, we can infer constraints on what B1 and P1 must be like.
Of course, we will never be able to fully know either, but especially if we had some
partial antecedent information about both, then learning that a P OptpB1 , P1 , Aq can
allow us to infer rather concrete things about B1 and P1 .
The reasoning in (5.3) is of just this kind: By observing an utterance, the addressee can conclude that this utterance satisfied the speaker’s preferences better
than the alternatives, given his beliefs; and so the addressee can ‘reason backwards’
to a conclusion about the speaker’s beliefs. We can bring this out by reformulating
(5.3). Let BSp and PSp be the speaker’s beliefs and preferences, and let ASp be any
set that contains utterpSp, ϕq.
(5.5)
a.
[Knowledge of convention]
According to BSp : When Sp utters $ ϕ, he becomes committed to believe ϕ.
b.
[Contextual assumption]
PSp specifies that the speaker will not commit himself to ϕ unless BSp
supports ϕ.
c.
[From (a) and (b) + knowledge of Opt]
utterpSp, ϕq P OptpBSp , PSp , ASp q only if BSp supports ϕ.
d.
[Belief about actions]
For any a P A: a happens only if a P OptpBSp , PSp , ASp q.
e.
[Observation]
utterpSp, ϕq happened.
f.
[From (d) and (e)]
utterpSp, ϕq P OptpBSp , PSp , ASp q
g.
[From (c) and (f)]
BSp supports ϕ.
CHAPTER 5. ACTION CHOICE AND COMMITMENT
109
The system PSen already contains a representation of beliefs, allowing us to
model premise (5.5a), as well as the conclusion (5.5g), and it allows us to talk about
events occurring, as in premise (5.5e). In order to capture the reasoning in (5.5)
fully, we minimally need to add a representation of preferences (so that we can
represent premises like (5.5b)) and we need to specify how Opt relates actions to
beliefs and preferences (so as to represent a premise like (5.5c)).
Section 5.2 introduces preferences, Section 5.2.2 a working definition of Opt.
Before, I will show how to implement (5.5d), the assumption that speakers only
perform actions that are selected by a given Opt. This involves some complications, which will lead us to revise one of the constraints we have put on beliefs in
Chapter 2.
5.1.1
Actions and agents
As a first step, we need to amend our model to specify which events are actions,
and who the agent of a given event is (i.e., who decides whether an action occurs).
Let Ag be the set of agents, then:
(5.6)
The models for PSen contain a partial function Agt : W ˆ T ÞÑ Ag.
Agt specifies the unique agent (if any) who acts at t in w. Given that we have
restricted our models such that only one event can happen at any given time,
only one agent can act at any given time, hence it makes sense to also require
that there is a unique agent who decides on actions. We also define the function
Act : W ˆ T ÞÑ E, which specifies action alternatives for the agent.
(5.7)
For all w, t : Actpw, tq “ tev | Dv : w «t v & Hapv pt, t ` 1q “ evu
if Agtpw, tq is defined, undefined otherwise.
Thus if Agtpw, tq “ i, then agent i decides which of the possible futures that diverge
at time t becomes actual. The alternative actions available to him are all events
that can happen just after t. We don’t require that Agtpw, tq is defined for all w, t to
allow for events that are not actions.
CHAPTER 5. ACTION CHOICE AND COMMITMENT
5.1.2
110
Beliefs about action choices
Given a fixed Opt, we now simply impose a generalized version of premise (5.5d)
as a constraint on our models. The constraint requires that all agents believe each
other to only perform actions that cohere with Opt. This is an idealization, even
with utterance actions. Sometimes a speaker may speak involuntarily, in a way
that does not cohere with his beliefs and preferences, and his audience may well
be aware of this fact. But pragmatic reasoning, or at least pragmatic reasoning of
the Gricean kind, is suspended in such a situation and hence, in situations that we
are interested in as pragmaticists, this assumption is warranted.
(5.8)
Belief in optimal actions constraint
For all w, v, t, t1 , i , i1 such that wRi,t v and Agtpv, t1 q “ i1 :
Hapv pt1 , t1 ` 1q P OptpBi1 ,t1 ,v , Pi1 ,t1 ,v , Actpv, t1 qq
where Pi,tvw is a representation of the preferences of i1 at v, t.
This constraint straightforwardly captures the assumption that agents believe each
other to choose their actions via Opt (based on their beliefs and preferences).
Given that we want to implement this constraint, however, we need to slightly
relax one of the other constraints we imposed on our models in Chapter 2: the
constraint No fore-knowledge, repeated below, which requires that the beliefrelations at any given time t do not distinguish between worlds that have not
diverged at t: If an agent takes a world to be possible, he must also take all its
possible futures to be possible.
(5.9)
No fore-belief constraint
=(2.16)
If v1 «t v2 , then wRi,t v1 iff wRi,t v2
It may not be immediately obvious that (5.8) and (5.9) are in conflict. On the face of
things, combining the two simply seems to have the result that an agent can take a
world v to be possible at t only if all possible futures of v are such that only optimal
CHAPTER 5. ACTION CHOICE AND COMMITMENT
111
actions are chosen at times after t.
That, precisely, is the problem. A world v which only has futures in which
optimal actions are chosen at all times is a world in which no non-trivial action
choices occur.1 Here is why: The actions available to an agent at t1 are determined
by the possible futures at t1 : For each possible action, there is at least one such
possible future in which the action occurs. This includes the actions that are nonoptimal, which will exist in all non-trivial choice situations. But then, every world
in which a non-trivial action choice is made will have possible futures at which
non-optimal actions are performed.
This is not an accident, but a necessity: By hypothesis, an agent chooses between
his actions by comparing what would be the case if he performed them. But that
means that even when he acts optimally, there must be ‘counterfactual’ possibilities
in which he does not, because otherwise we cannot define what it means to act
optimally). The relevance of such counterfactual possibilities in modeling action
choice has been recognized in the epistemic approach to game theory, in particular
in connection with the justification of backward-induction in games of perfect
information (Aumann 1995, Aumann 1996, Stalnaker 1996).
What this means that if an agent i believes about another agent that he will act
optimally in the future, then i has a fore-belief: On the one hand, he takes the world
to be such that the agent could perform both optimal and non-optimal actions (for
else, we cannot say that the agent will choose an action). On the other hand, i
believes that some of these actions (the non-optimal ones) will not occur. Seen this
way, beliefs about future action choices intrinsically are fore-beliefs.
However, we do not want to give up on the constraint in (5.9) altogether. We
still require it to hold with respect to non-action events. The following sequence of
definitions defines the weaker version of No fore-belief that ensures this.
(5.10)
For two worlds w1 , w2 such that for some t : w1 «t w2 , the point of divergence
divpw1 , w2 q is the unique time t1 such that w1 «t1 w2 and w1 0t1 `1 w2 .
1
Formally, such a world v must be such that at all future times t1 and all worlds v1 «t v : If
Actpv1 , t1 q is defined, then OptpBAgtpt1 q,t1 ,v1 , PAgtpt1 q,t1 ,v1 , Actpv1 , t1 qq “ Actpv1 , t1 q, i.e., it must be that, at
all future decision points, all alternative actions are equally optimal.
CHAPTER 5. ACTION CHOICE AND COMMITMENT
(5.11)
112
Two worlds w1 and w2 are external historical alternatives at t, written
w1 «et w2 iff w1 «t w2 and Agtpw1 , divpw1 , w2 qq is undefined.
(5.12)
No fore-knowledge about non-actions constraint
If v1 «et v2 , then wRi,t v1 iff wRi,t v2 , for all i.
5.2
Modeling Preferences
We now have extended our models so that they contain the assumption that agents
take each other to choose their actions, based on their beliefs and preferences. In
terms of the reasoning that we want to capture, we are now able to represent the
steps marked with a 3 below.
(5.13)
a.
[Knowledge of convention]
According to BSp : When Sp utters $ ϕ, he becomes committed to
believe ϕ.
b.
[Contextual assumption]
PSp specifies that the speaker will not commit himself to ϕ unless BSp
supports ϕ.
c.
[From (a) and (b) + knowledge of Opt]
utterpSp, ϕq P OptpBSp , PSp , ASp q only if BSp supports ϕ.
d.
[Belief about actions]
3 For any a P A: a happens only if a P OptpBSp , PSp , ASp q.
e.
[Observation]
3 utterpSp, ϕq happened.
f.
[From (d) and (e)]
3 utterpSp, ϕq P OptpBSp , PSp , ASp q
g.
[From (c) and (f)]
3 BSp supports ϕ.
All that is missing (besides a way to talk about commitments, which is the topic of
the second part of this chapter) is a suitable representation of preferences (which
CHAPTER 5. ACTION CHOICE AND COMMITMENT
113
we have presupposed anyway), so as to represent (5.13b), as well as as a characterization of Opt that allows us to conclude (5.13c). This is what we will do in the
present section. We begin with the representation of preferences in Section 5.2.1.
Thereafter, in Section 5.2.2, I discuss possibilities for Opt that support the inference
in (5.13c).
5.2.1
Preference structures
I will represent the preferences of an agent as a set of propositions, which are ordered
in terms of their importance. We can think of these propositions as capturing outcome
preferences, i.e., as properties of outcomes that the agent prefers.2 In other words,
if the proposition p is a preference of an agent, this means that he wants to see p
realized, all else being equal.
Now, of course, an agent typically has many different kinds of preferential
attitudes: desires, inclinations, appetites, personal moral codes, and so on. I
assume, however, that all these diverse preferential attitudes can be integrated
into one over-arching attitude that an agent uses to guide his action choices. This
assumption of an ‘overall’ preferential attitude is again familiar from decision
and game theory: There, too, an agent’s action-relevant preferences are typically
represented by a single utility assignment, which is viewed as the composite of
all preferential factors guiding the agent’s action choice. I call these ‘overall’
preferences Effective Preferences (a term introduced in Condoravdi and Lauer
2011), to highlight the fact that these are the preferences that are relevant for the
agent’s action choices.
(5.14)
A preference structure is a pair hP, ďi, where P is a set of propositions and
ď is a binary relation on P that is reflexive, transitive and total.3
2
In Chapter 9, when I apply the model of preference and action choice to conversational implicatures, it will be necessary to represent a second kind of preference, action preferences, for which a
propositional format is not well-suited.
3
A binary relation R on a set P is reflexive if pRp for all p P P, it is transitive if for all p1 , p2 , p3 P P,
p1 Rp2 and p2 Rp3 imply p1 Rp3 and it is total if for all p1 , p2 P P, either p1 Rp2 or p2 Rp1 (and possibly
both).
CHAPTER 5. ACTION CHOICE AND COMMITMENT
114
I take it that reflexivity and transitivity are intuitively necessary properties for any
relation that captures an ‘importance ranking’ which allows for ties. Totality might
not be as obviously adequate as a requirement—an agent might be genuinely undecided which of two preferences is more important, without necessarily considering
the two preferences equally important.4 Requiring totality has the convenient consequence that a preference structure decomposes into equivalence classes that are
linearly ordered by ď. As it will occasionally be useful to refer to these equivalence
classes, I introduce some notational conventions to talk about them:
(5.15)
Notational conventions: Given a preference structure hP, ďi,
a.
p ă q iff p ď q and q p
b.
p “ď q iff p ď q and q ď p
c.
rpsď “ tq | p “ď qu
d.
”ď is the set trps | p P P u
Preference structures in themselves can be used to represent any kind of preferential
attitude. Effective preferences have to satisfy a number of additional rationality
constraints, in particular they must be both realistic and consistent, relative to the
agents’ beliefs.
(5.16)
Realistic preference structures
An preference structure hP, ďi is realistic, relative to an information state
B iff for all p P P : p X B , H.
(5.17)
Consistent preference structures (Condoravdi and Lauer (2012))
A preference structure hP, ďi is consistent with respect to an information
Ş
state B iff for any X Ď P, if B X X “ H, there are p, q P X such that p ă q.
The definition of consistency is a generalization of the definition in Condoravdi
and Lauer (2011), which perhaps is easier to grasp intuitively:
4
This seems to be possible, in particular, for non-effective preferences, such as mere desires or
inclinations. If so, it would be more adequate to impose totality as a rationality constraint on effective
preferences, on a par with the realism and consistency constraints introduced below.
CHAPTER 5. ACTION CHOICE AND COMMITMENT
(5.18)
115
Consistent preference structures (simplified version, Condoravdi and
Lauer (2011))
A preference structure hP, ďi is consistent with respect to an information
state B iff for any p, q P P, if B X p X q “ H, then p ă q.
(5.18) says that a preference structure is consistent (relative to an information state B)
if every two preferences that are known (in B) to be incompatible are strictly ranked.
The underlying intuition is that if an agent believes that two of his preferences are
incompatible, in order to act, he has to decide which of the two is more important
to him. (5.17) generalizes this condition to ensure strict ranking of preferences also
in case more than two propositions are known to be jointly incompatible, even if
they are pairwise compatible. Consistency in the sense of (5.17) entails consistency
in the sense of (5.18).
With these notions in place, we amend the models for PSen again:
(5.19)
Models for PSen contain a function EP such for all i P Ag, w P W, t P T,
EPi pw, tq is a preference structure that is realistic and consistent with respect to Bi,t,w .
Given that effective preferences are those preferences that guide action choice, both
constraints on them are intuitively appropriate. An agent should not strive to
make true a proposition he knows is unattainable; and if he has two preferences he
knows to be inconsistent (an appetite for ice cream, say, as well as a desire to lose
weight), he’d better decide which of the two preferences is more important to him
before deciding to act.
Ranked preferences as ceteris paribus preferences
Employing ranked preferences (instead of just a flat set of propositions) has various
advantages which will become apparent in what is to follow. The most important
advantage is that it lets us conceive of the propositions in EP as ceteris paribus
preferences: If p P EP, that does not necessarily mean that the agent prefers p
over
p tout court, but rather only if ‘everything else is equal’. Hansson (1996)
CHAPTER 5. ACTION CHOICE AND COMMITMENT
116
articulates well how central such preferences are to our everyday reasoning:
“When discussing with my my wife what table to buy for our living
room, I said: ‘A round table is better than a square one.’ By this I did
not mean that, irrespective of their other properties, any round table
is better than any square table. Rather, I meant that any round table
is better (for our living room) than any sqaure-shaped table that does
not differ significantly in its other characteristics, such as height, wood,
finishing, price, etc. This is preference ceteris paribus or ‘everything
else being equal’. Most of the preferences that we express or act upon
seem to be of this type.”
(Hansson 1996, p. 307)
Such ceteris paribus preferences play a crucial role in pragmatic reasoning, in
particular in the derivation of implicatures, which will be discussed in Chapter 9.
To anticipate, it is often necessary to assume that speakers prefer uttering certain
forms over uttering others. An often used example is a preference for forms that
are short and simple over those that are longer and more complex (cf. Grice’s (1975)
Maxim of Brevity).
While it is intuitive to assume such a preference in general, in usual cases such
an assumption makes sense only if the preference is ceteris paribus: It should be
subordinate to the speaker’s communicative goals, for example. If a speaker desires
to convey a certain amount of information (say, p ^ q) that cannot be conveyed by
using a shorter form (say, ‘p’ or ‘q’ instead of ‘p ^ q’), then we usually would
not expect the speaker to give up on his communicative goal in order to satisfy
his preference for a shorter form—that is, his preference for shorter forms is only
ceteris paribus: In choosing between two forms, the speaker will opt for the shorter
form only if both forms satisfy his other goals equally. The order ď is intended to
capture this—intuitively, a preference p P P is only to be taken into account if the
preferences that strictly outrank p do not decide between two actions.
CHAPTER 5. ACTION CHOICE AND COMMITMENT
117
Interaction between beliefs and preferences
Effective preferences are those that an agent uses to decide for actions. It hence
makes sense to assume that agents are aware of their effective preferences. We
ensure this by requiring:
(5.20)
Preference introspection constraint
p P EPa pw, tq iff for all v P Bi,t,w : p P EPi pv, tq
(5.20) embodies an idealization, in particular together with the assumption that
agents take each other to always act optimally. In actual fact, an agent’s decisions
can be influenced by preferential attitudes he is unaware of. And yet, in such
a case, an agent will generally assume that he is aware of the factors driving his
action choices. In order to implement this, we could weaken (5.20) to only apply
to worlds contained in the agent’s belief state, but not necessarily the actual world.
However, since preferences that the agent is unaware of will not play a role in the
cases I discuss in this dissertation, I will assume the simpler (5.20) throughout.
We will also introduce an operator ep that allows reasoning about effective
preferences in the pragmatic language, but before we can do so, we will have to
specify how preferences (and beliefs) relate to action choice.
5.2.2
A working definition for Opt
The framework now can represent beliefs and preferences. All that is needed to
reason about action choice is a specification of Opt. This operator is, of course,
central to the whole enterprise. Yet, I will not spend much time in defending
the choice of Opt made here. The reason for this is that the precise nature of
the decision procedure is a very complex topic, which depends, in no small part,
on considerations beyond the scope of pragmatic theory. Ultimately, we want a
definition of Opt that is useful to model action choice in general, with utterance
choice just being a special case. But, of course, investigations of language use
will not tell us anything about how non-linguistic actions are chosen. A full, and
generally defensible, specification of Opt can only emerge from the joint work of
CHAPTER 5. ACTION CHOICE AND COMMITMENT
118
pragmaticists and researchers in other disciplines concerned with action choice,
such as psychology, behavioral economics, or the philosophy of action.
Luckily, however, we do not need to have a generally-applicable definition of
Opt in order to investigate pragmatic reasoning. We can make do with imposing
some constraints on how Opt functions.
In order to state these constraints, it is useful to conceive of OptpB, P, Aq as being
defined in terms of a ‘betterness’ relation ăB,P over the actions in A, such that a ă a1
iff a1 is strictly better to fulfill the preferences in P, given the beliefs in B. Given that
it is supposed to be a strict betterness relation, we require that ăB,P is irreflexive
and transitive (hence anti-symmetric). Then Opt is simply defined as:5
(5.21)
OptpB, P, Aq :“ ta P A | for no a1 P A : a ăB,P a1 u
if B , H and P is realistic and consistent given B. Else undefined.
The only substantive constraint on admissible ăB,P orderings I will impose here is
that it should respect the ‘importance’ ranking ď on P: As indicated above, the idea
is that a lower-ranked preference is taken into account only if the higher-ranked
preferences don’t make a decision. (5.22) captures this idea.
(5.22)
Lexicographicness
If a ăB,P a1 and P Ď Q such that for all p P P, q P QzP : q ă p,
then a ăB,Q a1 .
Beyond this, there are only very weak assumptions that should be uncontroversial,
given the interpretation of the propositions in P as properties of outcomes that the
agent prefers:
(5.23)
5
a.
For no a, a1 , B : a ăB,H a1 .
b.
If for all p P P : Bras Ď p, then for no a1 : a ăB,P a1 .
c.
If for all p P P : Bra1 s Ď
p, then for no a : a ăB,P a1 .
Again, defining Opt via a relation on actions is very much in keeping with the conception of
action choice in decision and game theory: There, the ‘betterness’ ranking is defined in terms of
expected utility, and the prediction of the theory is that agents will (or should) choose one of the
actions that yields the highest expected utility value.
CHAPTER 5. ACTION CHOICE AND COMMITMENT
119
(5.23a) says that an agent that does not have any preferences should be indifferent
between any two actions. (5.23b) requires that if an action is certain to satisfy all
preferences of the agent, it should be optimal and (5.23c), conversely, says that an
action that is certain to not satisfy any of the preferences of an agent should not be
strictly preferred over any other action.
Besides these requirements, the shape of ă (and hence, the nature of Opt) largely
depends on two questions: (i) How is uncertainty about the consequences of actions
treated? and (ii) How are preferences that are unranked with respect to each other
treated?
For the sake of concreteness, I will define a particular instantiation of ăB,P that
embodies very simple answers to both questions. This is defensible largely because
the cases we will consider in the remainder of the thesis will be independent from
the correct answer to (i) and (ii). The definition of ăB,P offered below hence should
be seen as nothing more of an example implementation that is useful for working
through example cases: All that I require in general are the constraints given above.
(5.24)
Given a belief state B and a preference structure hP, ďi; and E P ”ď let
a ĺE a1 iff tp P E | Bras Ď pu Ď tp P E | Bra1 s Ď pu
with «E and ăE defined in the obvious way in terms of ĺE .
(5.25)
Betterness of actions (exemplary version used in the rest of this thesis)
a ăB,P a1 iff for some E P ”ď : a ăE a1 and for all E1 ą E : a «E1 a1
With respect to question (i) above, this definition simply assumes that all that
matters is whether or not an agent is certain that an action realizes a given preference
(i.e., whether or not Bras Ď p), and hence is a lexicographic variant of a ‘maximin’
rule (i.e., it optimizes the ‘worst case’). This is sufficient for the cases discussed in the
remainder of this dissertation, as substantial uncertainty about the consequences
of actions will not play a role.
With respect to question (ii), the definition makes a very weak claim: An action
CHAPTER 5. ACTION CHOICE AND COMMITMENT
120
a is strictly better than another action a1 iff a satisfies all the preferences that a1
satisfies, and also satisfies at least one preference that a1 does not satisfy. At least as
long as we restrict attention to cases in which the agent is certain which preference
any given action satisfies, this seems like a minimal requirement on a ‘betterness’
relation for agents.6
5.2.3
Reasoning about preferences
As a final step, we now introduce an ep operator into the pragmatic language PSen ,
so that we are able to reason about effective preferences.
The intended interpretation for epi,t pφq is that at time t, agent i’s effective preference structure is such that he acts as if φ were a maximal preference of his. To
implement this, we first define an operator ` which adds a new preference to an
existing preference structure as a maximal element:
) !D
)
!D
E
E
(
(5.26)
hP, ďi ` φ “ P Y φ , ď Y φ, p | p P P Y p, φ | p P maxpP, ďq
Secondly, we define an equivalence relation over preference structures, as follows:7
(5.27)
EP „i,t,w EP1 iff
@v P Bi,t,w , t1 ě t : if Agtpw, tq “ a, then
OptpBi,t,w , EP, Actpw, t1 qq “ OptpBi,t,w , EP1 , Actpw, t1 qq
That is, two preference structures are „i,t,w -equivalent iff at all future decision
situations which the agent expects to face, EP and EP1 determine the same set of
optimal actions.8 The idea is that two preference structures are equivalent if, given
6
This definition is essentially the same as the definition of rationality used by Aumann (1995,
1996) and related work in game theory when there is no uncertainty about outcomes (in particular,
in games of ‘perfect information’): “a player i is rational if and only if it is not the case that he knows
that he would be able to do better” (Aumann 1996, p. 138). Of course, for Aumann, ‘do better’ is
defined as maximizing a utility value, in which case this definition comes out as a consequence of
expected utility maximization: “This is weaker than (i.e., implied by) expected utility maximization;
if an agent is maximizing expected utility, surely he cannot have an option that he knows, with
certainty, will yield a preferred outcome.” (Aumann 1995, p. 9, n. 4)
7
Recall that because the Ra relations are transitive and Euclidean, we have @v P Bi,t,w : Bi,t,v “ Bi,t,w .
8
In full generality, we probably want to restrict the set of future decision situations that are
CHAPTER 5. ACTION CHOICE AND COMMITMENT
121
what the agent believes, it does not make a difference which one he uses to choose
his actions.
With these notions in place, we can define ‘preference-support’:
(5.28)
EP $i,t,w φ iff EP ` φ „i,t,w EP and EP ` Wzφ /i,t,w EP
EP supports φ if adding φ as a maximal element to EP would not change the agent’s
decision in any situation he expects to face. The second conjunct in the definition is
necessary to exclude ‘spurious’ support—that is, if an agent expects that neither a φ
preference nor a not-φ preference would make a difference for his future decisions,
we do not want to say that his preferences support φ—e.g., we do not want to say
that an agent prefers it to rain tomorrow in virtue of the fact that he does not have
control over the weather.
With this, we can state the interpretation clause for ep:
(5.29)
w epi,t pφq iff EPi pw, tq $i,t,w φ
(final)
This ensures that if p is a maximal element of EPi pw, tq, then w epi,t ppq, which
is how Condoravdi and Lauer (2011) defined the ep operator. This was sufficient
there, as we were mostly interested in talking about the public version of effective
preferences, which only have maximal elements (cf. Section 5.3 below).
The definition in (5.29) is more general, as it also allows epppq to be true if p is a
non-maximal element, but the agent does not expect to face a decision situation in
which p is defeated by a higher-ranked preference.
5.2.4
Summary
With the representation of preferences, and a definition of Opt, we are now able to
formalize most parts of the reasoning schema we started out with—except for the
parts pertaining to commitment:
taken into account somewhat. It may be that an agent only expects to face a decision situation
after learning some new pertinent fact. In that case, this decision situation should not be taken into
account.
CHAPTER 5. ACTION CHOICE AND COMMITMENT
(5.30)
a.
122
[Knowledge of convention]
BSp rutterpSp, ϕqs Sp is committed to ϕ
b.
[Contextual assumption]
epSp pSp is committed to ϕ ñ BSp ϕ)
c.
[From (a) and (b) + knowledge of Opt]
utterpSp, ϕq P OptpBSp , PSp , ASp q ñ BSp ϕ
d.
[Belief about actions]
For any a P A: a happens only if a P OptpBSp , PSp , ASp q.
e.
[Observation]
utterpSp, ϕq
f.
[From (c) and (d) and (e)]
BSp ϕ
More generally, this reasoning schema captures the following inference: If Sp
believes that an action a has consequence c, and he prefers to avoid c unless that c1
obtains, then from observing a, we can conclude that he believes c1 obtains.
The relevant consequence c here is ‘Sp becomes committed to believing ϕ’ and
the condition is ‘The speaker believes ϕ’. The rest of this chapter lays out how to
capture the notion of commitment in the system of dynamic pragmatics.
5.3
Modeling commitment
The notion of commitment I am aiming at here may not completely match the
everyday use of the word commitment, and it certainly is at variance with ways
in which the term has been used in linguistics and philosophy. Before I go ahead
with saying how this notion should be understood formally, I want to briefly lay
out some basic properties I take commitments, in the intended sense, to have.
Commitments are public I take commitments to be always inter-personal. That
is, a commitment is always made towards an agent, or a group or community of
agents. I take this to be a fundamental part of our notion of ‘commitment’: One
CHAPTER 5. ACTION CHOICE AND COMMITMENT
123
cannot take on a commitment unless there is at least a witness to whom one makes
it. Of course, we sometimes talk of private resolutions as if they were commitments:
If an agent privately resolves to exercise at least three times a week, we may say that
he ‘makes a commitment’ to exercising that often. I take such uses to be derived
from the inter-personal sense.
Commitments are commitments to act
I also take commitments to be necessarily
connected to action: A commitment is kept by making the right action choices; it
is violated by making the wrong ones. That means I take locutions such as ‘a is
committed to believe / the belief that p’ to be short for ‘a is committed to act as
though he believes p’. Consequently, if an agent is ‘committed to the belief that p’,
he does not violate this commitment merely by failing to believe that p. Instead,
he can be in perfect compliance with his commitment, as long as his action choices
are consistent with (his other commitments and) the belief that p.9
Conceiving of commitments in this way quite severely limits the things one can
be committed to: An agent can only be committed to things that influence his action
choices, i.e., an agent can only be committed to mental attitudes (strictly speaking,
to act as though he had a mental attitude), in particular beliefs and preferences.
Saying that an agent ‘is committed to p’, where p is a proposition about an agentexternal fact makes sense only as an abbreviation for ‘is committed to (act though
he has) the belief that p’ or, possibly, ‘is committed to (act as though he has) a
preference for p’.10
Commitments survive the present discourse
The notion of commitment em-
ployed here is more general than the ‘mere discourse commitments’ in theories
like that of Hamblin (1971) and the work of Gunlogson (2003, 2008). These authors
9
This will be instrumental in explaining why a speaker would be willing to ever engage in ‘loose
talk’, e.g., saying that France is hexagonal even though he clearly believes it is not. In the theory
developed here, such an utterance will commit the speaker to the belief that France is hexagonal—
and, by and large, we will predict that a speaker will only utter the sentence if he believes the
non-hexagonality of France will not make a difference to his future action choices.
10
Both kinds of abbreviatory uses are common, e.g., John is committed to Toulouse being the
capital of France and John is committed to getting the president re-elected.
CHAPTER 5. ACTION CHOICE AND COMMITMENT
124
characterize commitments in terms of the future discourse actions the commitment
excludes. For example, if an agent commits to a belief that p, this precludes future
discourse moves that express a belief that is incompatible with p.
This was adequate for the purposes of Hamblin and Gunlogson, but it is too
limited in the present context for two reasons. Firstly, the kinds of commitments
created by linguistic utterances often concern actions that take place after the current discourse has ended. Secondly, many of the commitments we are interested in
here concern non-linguistic actions. The commitments resulting from promises and
orders are particularly obvious examples: These can of course concern linguistic
actions (Don’t say another word!, Say it ain’t so!), but mostly, they do not (Be at
the airport at noon!).
The discourse commitments of Hamblin and Gunlogson come out as a special
case on the perspective taken here: Commitments constrain actions in general,
including linguistic actions. To the extent that it is necessary to also represent
commitments that are in effect only for the present discourse, we can allow commitments that are only temporary, i.e., that constrain actions only up to a given
point in time, such as the end of the current discourse, or the end of the current
discourse segment.11
5.3.1
Commitments exclude possible future states of the world
Commitments will be construed as excluding possible future states of the world in
our branching-time model. This is somewhat parallel to the way discourse commitments are modeled in Hamblin (1971) and Gunlogson (2008), where commitments
constrain the set of ‘legal’ (Hamblin) or ‘expected’ (Gunlogson) future states of the
discourse. That is, discourse commitments are modeled by excluding all future
discourse states at which the commitment is not honored.
This will not do if we model commitments as excluding possible future states
11
The latter is what happens when a speaker agrees to a hypothesis ‘for the sake of the argument’:
For the current discourse segment, he commits to act as though he believed the hypothesis, but this
commitment is not as enduring and open-ended as the commitment that he would have taken on
otherwise.
CHAPTER 5. ACTION CHOICE AND COMMITMENT
125
of the world: Of course, taking on a commitment does not make it impossible to
violate the commitment. But there are future states that are indeed made completely impossible by taking on a commitment: Those where the commitment is
not honored, and yet, the agent does not count has having violated a commitment.
One-off commitment to action: An illustration
To bring out the intuition, let us consider a somewhat artificial case of a commitment
to a one-off action. At time t, John takes on the commitment to raise his hand at a
future time time t1 . After t, there are two kinds of possible future worlds:
1. future worlds in which John raises his hand at t1 .
2. future worlds in which John does not raise his hand at t1 , and counts as having
violated his commitment.
Before t, there was a third kind of possible future, namely those in which John does
not raise his hand, yet is not at fault for having violated the commitment to do so.
A necessary complication: Commitments can be rescinded, or canceled.12 So
there are really three kinds of possible futures after t: The two characterized above,
plus:
future worlds in which the commitment gets rescinded before t1 , and John
3.
does not raise his hand at t1 .
The taking on of the commitment thus can be construed as making impossible all
future worlds of the kind characterized below:
future worlds in which the commitment does not get rescinded between t
4.
and t1 in which John does not raise his hand at t1 and does not count as
having violated the commitment to do so.
12
In the real world, commitments vary with respect to which agent ‘controls’ a commitment, i.e.,
which agent has the right to rescind the commitment: Promises can primarily be rescinded by the
promisee, while declarative commitments to belief can often be freely rescinded by the agent who
made the commitment in the first place (I take that back). I ignore this complication here.
CHAPTER 5. ACTION CHOICE AND COMMITMENT
126
Commitments to courses of actions
Agents do not always commit to single actions, as above, but frequently commit
to complex future courses of action. Suppose, for example, that John commits to
never drink alcohol again. There is no single future time at which this commitment
will be honored or violated. Instead, this commitment will be honored or violated
every time John has the opportunity to take a drink.
This complicates matters, as ‘being in accordance with the commitment’ is no
longer a stable property. In the case of a one-off action, to be performed at time t1 ,
there was a single decisive point (viz., t1 ) at which the commitment was honored
or violated. If John performed the necessary action at t1 , at all future times, he
was in accordance with his commitment. For open-ended commitments, this is no
longer true: If John has not taken a drink up to time t2 , he is in accordance with
his commitment, but there are still many possible futures after t2 at which he no
longer is, namely, all those at which he takes a drink after t2 .
The notion of ‘having violated the commitment’, however, is still a temporally
stable one: If, at any given time, John takes a drink, he will henceforth count as
having violated his commitment. For this reason, it is expedient to characterize
commitments in general using the notion of failing to act in accordance with a commitment, rather than the—prima facie—simpler notion of acting in accordance with
a commitment.
So, informally, for a given property P of actions (such as ‘the action is not a
drinking of alcohol’), we explicate ‘commitment to P’ as follows:
(5.31)
When an agent i takes on a commitment to P at time t, he excludes all
possible future worlds w1 from becoming actual which are such that, for
some t1 after t:
a.
in w1 , i does not act in accordance to P at t1 AND
b.
in w1 , P has not been rescinded prior to t1 AND
c.
in w1 , i is not at fault at t1 .
This characterization employs the primitive predicate ‘being at fault’, which I will
CHAPTER 5. ACTION CHOICE AND COMMITMENT
127
leave unanalyzed here. Seeing as the commitments in questions are taken to arise
on the basis of social conventions, it might be more appropriate to circumscribe it
as ‘is considered at fault by the community’, but I shall use the simpler formulation
here.13
5.3.2
Commitments to beliefs and preferences
The preliminary characterization of the notion of commitment in the previous section was done in terms of predicates of courses of action. In the present section, I
will implement this indirectly in the system of dynamic pragmatics, modeling commitment to belief (doxastic commitment) and commitment to preferences (preferential
commitment). This indirection is required because we want commitments to interact: Whether or not any given action choice of an agent violates his commitments
will, in general, not depend on single commitments of the agent, but on their sum.
Suppose, for example, that an agent is committed to a preference for being at the
airport at noon, and also to a belief that, in order to be at the airport at noon, he has
to leave home before 10:30. Suppose further that the agent is still at home at 10:45.
We want to be able to say that he is in violation of his commitments, but his being
at home is compatible with both his preferential commitment and his doxastic
commitment, taken individually. It is only together that the two commitments are
incompatible with his being at home.
I hence introduce predicates pb (for ‘public belief’) and pep (for ‘public effective
preference’), which are indexed to individuals and times, and take propositional
arguments. Model-theoretically, these are interpreted using two set-valued functions that specify the beliefs and preferences an agent is committed to, PB and PEP.
I require these to have the following (joint) closure properties.
(5.32)
For any i, t, w
a.
13
PBi pt, wq is closed under logical inference with an SD45-logic for pba,t .
Ultimately, we probably want to index ‘being at fault’ with the commitment(s) the agent has
violated (or rather—to avoid circularity—the commitment-creating events with respect to which he
is ‘at fault’). This, again, is a complication I avoid here.
CHAPTER 5. ACTION CHOICE AND COMMITMENT
b.
If p P PEPi pt, wq then pepi,t ppq P PBi pt, wq.
c.
If pepi,t ppq P PBi pt, wq then p P PEPi pt, wq.
d.
If p P PEPi pt, wq, then
128
p < PBi pt, wq
We use these two sets to interpret the operators, in analogy to the interpretation of
B and ep:
(5.33)
14
a.
w pbi,t pφq iff φ P PBi pt, wq.
b.
w pepi,t pφq iff PEPi pt, wq $i,t,w φ
That is, we treat pep as an effective preference structure that only contains maximal
elements.
(5.33a) simply ensures that pb behaves like a doxastic operator—in particular,
it requires positive and negative introspection. This ensures that:
(5.34)
a.
pbi,t ppbi,t ppqq ñ pbi,t ppq
b.
pbi,t ppq ñ pbi,t ppbi,t ppqq
(5.33b) says that if an agent is committed to believe that he is committed to prefer p,
he is also committed to prefer p—Condoravdi and Lauer (2011)’s Doxastic reduction for preferential commitment. (5.34c) is the inverse. While it will not play an
essential role in what is to follow, it seems so plausible that I impose it nonetheless.
I model the commitments here using sets of formulas because it allows an
easy way to state the closure properties syntactically. The desired logical properties for pb alone could be implemented easily in the standard way, by adding
a time-indexed accessibility relation Ca ptq for each agent a, which is required to
be transitive, Euclidean and serial, but ensuring introspection pep and pb would
require some additional structure I wish to avoid here.
In the present context, we ignore the possibility of rescission of commitments
and require that doxastic and preferential commitments are increasing:
14
I take it to be obvious how the definition of „i,t,w has to be adjusted for pep—by replacing
reference to the agent’s actual belief state with reference to his public one.
CHAPTER 5. ACTION CHOICE AND COMMITMENT
(5.35)
If t ă t1 and PBi pt, wq Ď PBi pt1 , wq.
(5.36)
If t ă t1 then PEPi pt, wq Ď PEPi pt1 , wq.
129
Now we connect these two constructs to at-fault-ness. To this end, we introduce
another time-indexed predicate AtFault into the pragmatic language. AtFault
applies to an individual if he counts as having violated a commitment. At-faultness is connected to PB and PEP by the following constraint on the interpretation
of the AtFault-predicate:15
(5.37)
A model for PSen is admissible only if, for all w, t, i: hii P IpwqptqpAtFaultq
iff there is t1 ď t:
(i)
Agtpw, tq “ i
(ii) Hapw pt, t ` 1q “ a
Ş
(ii) a < Optp PBi pt1 , wq, PEPi pt1 , wq, Actpw, tqq
This just says that an agent is AtFault at t if, at some previous time t1 , he performed
an action that was not consistent with his public beliefs and public preferences at
t1 .
The Declarative Convention can now be implemented in the obvious way:
Declaratives add their contents to the PB-set of the speaker. However, I do not take
PB to only contain elements that have been added by explicit utterance events—
that would leave the system far too weak. Minimally, it has to contain certain
‘obvious facts’, for example about events that have happened in the course of
the conversation, as well as certain basic background assumptions that could not
possibly be doubted by a rational agent. In this way, PB is similar to Stalnakerian
speaker presuppositions—things that are taken ‘for granted’ automatically are
PBs. In particular, I take utterance events to be recorded in the PBs of all utterance
participants, in the way we have ensured that speakers have perfect knowledge
about utterance events in Chapter 2 by means of the PAL-constraint.
15
For this version of the constraint to deal with failures to act, we have to construe inaction as the
performance of a ‘null action’.
CHAPTER 5. ACTION CHOICE AND COMMITMENT
5.4
130
Modeling conventions of use
We are now almost in the position to implement the normative conventions of use
governing the effects of clause types. The only thing that is left to introduce is the
modal operator Result indexed to events, with the intended interpretation ‘φ is
true as a result of e’.
(5.38)
If e is an event-constant or variable and φ a proposition, then Resulte pφq
is a formula.16
Result is interpreted via the function Res : W ˆ T ˆ T ÞÑ ℘pWq. This function should
obey a number of constraints which I will not go into here, save for requiring that
it is closed under entailment, and that it is factive:17
(5.39)
If X P Respw, t, t1 q, then w P X.
Now we can capture normative conventions of use for sentence types by imposing additional constraints on admissible models, such as the following:18
(5.40)
Declarative Convention constraint
A model for PSen is admissible only if for all ϕ, e, i, i1 : If ϕ P L$ and
Hapw pIpeqq “ utterpi, i1 , ϕq then
w Resulte ppbi,t pϕqq
Anticipating the treatment of imperatives in Chapter 6, we can also state the
16
Recall that event constants and variables refer to event tokens, which in the current system are
simply identified with pairs of times.
17
Another plausible requirement is that Result is causally closed in the following sense: If e
is causally sufficient for e1 and Resulte1 pφq, then Resulte pφq. I don’t introduce this as an official
requirement because modeling causal relationships would require a further complication of the
current setup that do not concern us here.
18
Stating these conventions as global constraints on models has the effect that all agents always
believe the convention to be in place (and are, in fact, also committed to believe they are). This
makes sense, given that we assume knowledge of these conventions to be part of the basic linguistic
competence of speakers of a language.
CHAPTER 5. ACTION CHOICE AND COMMITMENT
(5.41)
131
Imperative Convention constraint
A model for PSen is admissible only if for all ϕ, e, i, i1 : If ϕ P L! and
Hapw pIpeqq “ utterpi, i1 , ϕq then
w Resulte ppepi,t pϕqq
These constraints encode the conventions in a rather straightforward way: If a
speaker utters a declarative or an imperative, the content of his utterance gets
added to his doxastic commitment or his preferential commitments, respectively.
We immediately obtain two results: Firstly, if a speaker utters incompatible
declaratives, then any future action will make him AtFault, as Opt is undefined if
its belief state argument is empty. Secondly, after a speaker utters two imperatives
that are incompatible, given his doxastic commitments, any future action will make
him AtFault, given the consistency requirement Opt imposes on its preference state
argument.
This is of course far too strong in general, but the result relies crucially on the
assumptions stated in (5.35) and (5.36), i.e., on the assumption that commitments
are non-decreasing. Thus we already capture the fact that if an agent commits to
incompatible beliefs or preferences, this requires rescission of a prior commitment.
5.5
Communicating contents, again
We can now show how it comes about that an addressee comes to believe the
content of declaratives that a speaker utters. Recall that, in Chapter 2, we already
derived this, relying on the following two contextual beliefs of the addressee:
(5.42)
Contextual assumptions: Trusting addressee
a.
‘Honest speaker’
BAd,t uttert pSp, $ ϕq ñ Sp,t ϕ
b.
‘Informed speaker’
BAd,t Sp,t ϕ ñ ϕ
=(2.34)
CHAPTER 5. ACTION CHOICE AND COMMITMENT
132
We now replace (5.42a) with the following condition (as with (5.42), this can be
seen as a specific instance of a more general belief about the speaker’s preferences):
(5.43)
‘Cautious speaker’
BAd,t epSp,t p Sp,t`1 ϕ ñ
pbSp,t`1 pϕqq
For simplicity, assume that the speaker chooses only between two actions, and
that the addressee knows this:
(5.44)
@w P BAd,t : Agtpw, tq “ Sp and Actpw, tq “ tutterpSp, Ad, ϕq, Ku
K stands for the ‘null action’ of doing nothing. That is, we assume that the speaker
only faces the choice whether to assert ϕ or not.19 Given the declarative convention,
we also have, for all worlds v (and hence in particular for all worlds v P BAd,t ):
(5.45)
BSp,t,v rutterpSp, Ad, ϕqs pbSp,t`1 ϕ
Similarly (assuming Sp is not already committed to ϕ):
(5.46)
BSp,t,v rKs 2 pbSp,t`1
Assuming that the addressee is uncertain whether the speaker believes ϕ to be true,
there are two kinds of worlds in BAd,t :
(5.47)
a.
Worlds vϕ such that vϕ Sp,t ϕ.
b.
Worlds v
ϕ
such that v
ϕ
Sp,t ϕ
Given (5.45), for vϕ -worlds, we have (5.48a) and (5.48b), which (in the absence of
additional relevant preferences) ensures (5.48c), where Optpwq is an abbreviation
for the outcome of Opt, given Sp’s beliefs and preferences at w.20
19
Of course, in realistic scenarios, there will often be many alternative utterances the speaker
could make, and even if we assume that ϕ is very topical, before the utterance, the addressee likely
takes it to be possible that the speaker utters ϕ instead. But for the sake of illustration, it is useful
to use a somewhat artificial example.
20
That is, Optpwq “ OptpBSp,t,w , EPSp pw, tq, Actpw, tqq.
CHAPTER 5. ACTION CHOICE AND COMMITMENT
(5.48)
For v
(5.49)
a.
BSp,t,vϕ rutterpSp, Ad, ϕqs Sp,t ϕ ^ pbSp,t`1 ϕ
b.
BSp,t,vϕ rKs Sp,t ϕ ^
c.
utterpSp, Ad, ϕq P Optpvϕ q
ϕ -worlds,
133
pbSp,t`1 ϕ
we have instead:
a.
BSp,t,v ϕ rutterpSp, Ad, ϕqs b.
BSp,t,v ϕ rKs c.
Optpv
ϕq
Sp,t ϕ ^
Sp,t ϕ ^ pbSp,t`1 ϕ
pbSp,t`1 ϕ
“ tKu
But, given the Belief in optimal action choice constraint, this ensures that only
vϕ -style worlds are such that Happt, t ` 1q “ utterpSp, Ad, ϕq, and so:
(5.50)
BAd,t rutterpSp, Ad, ϕqs Sp,t ϕ
And, assuming ‘Informed speaker’:
(5.51)
5.6
BAd,t rutterpSp, Ad, ϕqs ϕ
Conclusion
This chapter has introduced a good deal of machinery, which was illustrated with a
rather simple example of a pragmatic inference: The inference to the truth of what
the speaker said. The usefulness of the machinery will reveal itself in the chapters
to follow.
It is worth reflecting, however, on what we have done. We started in Chapter 2
with a system that let us show that an addressee who makes the ‘Honest speaker’
assumption will learn that the speaker believes in the truth of what he says. In
Chapter 3, I demanded justification for this contextual belief, on the basis that it is
warranted only for declaratives, but not for other sentence types.
Now, at the end of the present chapter, we have again derived the same result,
replacing the contextual assumption ‘Honest speaker’ with the contextual assumption ‘Cautious speaker’. A critical reader may wonder—what have we gained? We
CHAPTER 5. ACTION CHOICE AND COMMITMENT
134
have replaced one assumption with another. Where is the justification for ‘Cautious
speaker’?
Let us compare the two contextual assumptions, repeated below.
(5.52)
‘Honest speaker’
BAd,t uttert pSp, $ ϕq ñ Sp,t ϕ
‘A speaker will utter a declarative with content ϕ only if he believes it to
be true.’
(5.53)
‘Cautious speaker’
BAd,t epSp,t p Sp,t`1 ϕ ñ
pbSp,t`1 pϕqq
‘A speaker prefers not to commit himself to the proposition ϕ unless he
believes it to be true.’
‘Honest speaker’ makes essential reference to the clause type of the uttered sentence:
It is an assumption about when sentences in L$ will be uttered. ‘Cautious speaker’
makes no reference to sentences of the object language Sen at all, but only talks
about propositions (which it happens to identify by sentences of Prop).
Justification of ‘Honest speaker’ hence requires appeal to narrowly linguistic
concepts, while justification of ‘Cautious speaker’ does not. It is because of interlocutors’ linguistic knowledge that ‘Honest speaker’ is justified.
At the end of Chapter 4, I have proposed that the linguistic knowledge that
justifies ‘Honest speaker’ (in appropriate contexts) is related to the dynamic effect
that is conventionally associated with declarative sentences, and that this effect
consists in the creation of commitments to choose one’s actions in a particular way. To
make this hypothesis precise, the current chapter has introduced a model of action
choice (making as few assumptions as necessary about how it works, by specifying
general constraints on Opt relative to the chosen representation of preferences and
beliefs), and then gone one to specify a model of the notion of commitment suitable
for the statement of clause-typing conventions.
The usefulness of the apparatus introduced in this chapter is, of course, not
exhausted by explaining the simple pragmatic inference derived in the last section.
CHAPTER 5. ACTION CHOICE AND COMMITMENT
135
This is just the beginning. We now have a fairly general framework at our disposal
that can be applied to a variety of pragmatic phenomena. In Chapter 6, I apply
the framework to the other two main clause types, imperatives and interrogatives.
Chapter 7 shows how the current set-up can account for the fact that explicitly performative utterances have the result of making the content of the uttered sentence
true. Chapter 8 discusses exclamatives, arguing that they should not be viewed
as giving rise to commitments. And finally, Chapter 9 applies the framework to
model conversational implicatures.
Chapter 6
Commitments to preferences
The previous chapters were largely concerned with declarative sentences and their
effects. Other sentence types have been mentioned, but mainly in order to provide
a contrast to declaratives, which were argued to be associated with a convention
that ensures that a speaker who utters a declarative comes to be committed to
believe that that content of the declarative is true.
The present chapter is concerned with sentences that commit their speakers to a
preference. The main example are imperatives. According to the analysis adopted in
this chapter, an utterance of an imperative conventionally commits the speaker to a
preference for the imperative to become fulfilled (Section 6.1). Section 6.2 discusses
a class of declarative sentences that have imperative-like uses. While utterances
of these sentences, like all declarative utterances, conventionally give rise to a
commitment to a belief, I will argue that this doxastic commitment can be viewed
as giving rise to a preferential commitment indirectly. Section 6.3 sketches how the
emerging system of clause-typing could be extended to the third ‘major’ sentence
type: interrogatives. Interrogatives are analyzed on the model of imperatives, as
being conventionally associated with a preferential commitment of a very particular
kind. The chapter concludes with Section 6.4, which discusses the question what
the denotation of sentences of various types should be and revisits the question
whether information relating to (sentential) force should be represented in the
compositionally-determined denotation of sentences.
136
CHAPTER 6. COMMITMENTS TO PREFERENCES
6.1
137
Imperatives
A central problem for an analysis of imperatives is the problem of functional
heterogeneity, which was first appreciated in its entirety by Schmerling (1982):
Imperatives come with a wide variety of uses, and it is difficult, at least at first
glance, to find a unifying feature of all of them. Consider the examples in (6.1),
repeated from page 46.
(6.1)
a.
Stand at attention!
(Command)
b.
Don’t touch the hot plate!
(Warning)
c.
Hand me the salt, please.
(Request)
d.
Do the right thing!
e.
Take these pills for a week.
(Advice)
f.
Please, lend me the money!
(Plea)
g.
Get well soon!
h.
Drop dead!
i.
Okay, go out and play.
(Permission)
j.
Okay then, sue me, since it’s come to that.
(Concession)
k.
Have a cookie(, if you like).
(Exhortation)
(Well-wish)
(Curse)
(Offer)
Commands such as (6.1a) are the stereotypical (though not necessarily typical) use
of imperatives. They have two features, which they share with other ‘directive’
uses, such as requests (6.1c), warnings (6.1b) and exhortations (6.1d): (i) They act
as an inducement for the addressee to fulfill the imperatives and (ii) they express
that the speaker desires, in some sense, the addressee to fulfill the imperative.
However, neither of these properties is present on all uses: A well-wish like
(6.1g) and an ill-wish like (6.1h) do not act as inducements, they merely express a
desire. And permission and invitation uses like (6.1j-k) do not appear to express a
speaker desire at all, though they may act as inducements in some sense. This is
even more obvious in cases of disinterested advice, as in (6.2).
CHAPTER 6. COMMITMENTS TO PREFERENCES
(6.2)
138
[Strangers in the street of Palo Alto]
A: Excuse me, how do I get to San Francisco?
B: Take the north-bound Caltrain.
B’s utterance does not, in any obvious sense, express a desire on B’s part, even
though, given A’s mutually-known, salient goal, it will likely act as an inducement
for A to get on the Caltrain.
The problem of functional heterogeneity makes it particularly difficult to assign
imperatives a uniform sentential force—i.e., a uniform, conventionally-determined
essential effect that is present for all utterances of imperatives, as it is not obvious
what this uniform effect should be.
Condoravdi and Lauer (2012) argue, however, that we can specify such a uniform effect, and derive the varied uses of imperatives in context. The following
sections summarize the proposal, and show how the system of dynamic pragmatics can be used to demonstrate, on a very fine-grained level, how specific uses
of imperatives arise through the interplay of the conventionally specified effect of
imperatives and features of the context in pragmatic, interactional reasoning.
I will not follow the strategy of Chapter 4 and try to discuss various alternative
ways the force of imperatives could be specified. The main reason is that an
excellent exhaustive review of the literature exists, viz., that in Kaufmann (2012).
I would have little to add to Kaufmann’s discussion of previous proposals. As
Kaufmann’s discussion shows, few of the proposals predating her book are even
initially plausible once one appreciates the functional heterogeneity of imperatives.
That leaves the two main current alternative approaches, Kaufmann’s own and that
of Portner (2005, 2007). I refer the reader to Condoravdi and Lauer (2012) for a
discussion of these, as well as an in-depth comparison with the analysis proposed
here. A noteworthy feature of both approaches is that they take imperatives to be
semantically underspecified much in the way that modals like must are in analyses
following Kratzer (1981), and they explain functional heterogeneity as arising from
this underspecification. The analysis proposed here and in Condoravdi and Lauer
(2012), by contrast, assumes no semantic underspecification.
CHAPTER 6. COMMITMENTS TO PREFERENCES
6.1.1
139
Imperatives as creating commitments to preferences
Condoravdi and Lauer (2012) propose that the essential effect of imperatives is the
creation of a commitment to a preference:
(6.3)
Imperative Convention (NL statement)
When a speaker utters an imperative that has the content ϕ, he thereby
commits himself to prefer ϕ to be actualized.
This assumes that the content of imperatives is a proposition—the proposition that
is true if the imperative is fulfilled, i.e., that (6.4a) roughly has the same denotation
as (6.4b):
(6.4)
a.
Leave!
b.
You will leave.
This assumption is not at all essential. The analysis of imperatives proposed here
is compatible with a wide range of possible denotations for imperative, as I show
in Section 6.4. In the meantime, I will assume such propositional denotations for
the sake of concreteness. That is, I assume that
(6.5)
0 8Sen 0 8Prop
!φ
“ φ
With this, we can state (6.3) as a constraint on pragmatic models (this definition
was anticipated already in Section 5.3.2 above):
(6.6)
Imperative Convention constraint
(5.41)
A model for PSen is admissible only if for all ϕ, e, i, i : If ϕ P L! and
1
Hapw pIpeqq “ utterpi, i1 , ϕq then
w Resulte ppepi,t pϕqq
In view of the stereotypical uses of imperatives, i.e., commands, this proposal may
seem surprising: Imperatives commit the speaker to a preference? Isn’t the main
CHAPTER 6. COMMITMENTS TO PREFERENCES
140
point of a command-imperative like (6.7) to get the addressee committed to perform
the commanded action?
(6.7)
Be at the airport at noon!
Indeed, on command uses, this is generally the point of uttering an imperative, but
we deny that it is the conventionally-determined effect of imperative utterances.
One reason for this is conceptual. It would be quite strange if an utterance,
in virtue of linguistic convention, could commit the addressee to something. At
best, it could attempt to commit the addressee to something. The other reason is
empirical and lies in the problem of functional heterogeneity: There are many uses,
such as advice, permission, and wish uses, for which it is implausible to say that
they involve an addressee commitment, even an attempted one.
Assuming that imperatives give rise to speaker-commitments to preferences
brings into sharp focus, however, the flip-side of the problem of functional heterogeneity. While imperatives can have a wide range of uses, there are others that are
categorically excluded. Most notably, imperatives cannot be used to promise or
threaten:
(6.8)
a.
Get a promotion!
(cannot be a promise that the speaker will ensure that the addressee gets a
promotion)
b.
Lose your job!
(cannot be a threat that the speaker will ensure that the addressee loses his
job)
An analysis that hypothesizes that imperatives commit the speaker to a preference
for the imperative to be fulfilled highlights this problem, as promises arguably are
just utterances that commit the speaker to a preference for what is promised (cf.
Chapter 7). As Condoravdi and Lauer (2012) argue, however, the main alternative
accounts of imperatives, those of Kaufmann (2012) and Portner (2005, 2007), face
the same problem, at least to a certain degree. The commitment-based analysis
CHAPTER 6. COMMITMENTS TO PREFERENCES
141
merely makes the issue particularly obvious.
I shall not dwell on this problem here, but refer the reader to Condoravdi and
Lauer (2012), where we propose to solve it by assuming that, in addition to the
creation of a preferential commitment, imperatives come with a secondary effect
which limits the involvement of the speaker in bringing about the truth of the
proposition denoted by the imperative, ruling out promise and threat uses.
In the following, I briefly summarize some desirable consequences of the proposed analysis. In Condoravdi and Lauer (2012), we argue that all these consequences are necessary for a satisfactory analysis of imperatives, and also compare
the commitment-based analysis with that of Kaufmann and Portner, showing that
neither meets all the desiderata that we argue for.
Contextual inconsistency
As pointed out in Section 5.3.2, a consequence of the
current analysis is that subsequent imperatives by the same speaker are required,
just like declaratives, to be consistent. This is a welcome prediction.
Here is why: Imperatives require retraction or revision even when they are not
used with the same ‘force’. This is observed by Portner (2007, p. 367): “it’s odd
to give conflicting imperatives even when they are of different subtypes (unless
you have changed your mind, of course), as shown in example [(6.9)]. This pair of
sentences cannot be coherently uttered by a single speaker.”
(6.9)
Stay inside all day!
(order)
#Since you enjoy the nice weather, go out and play a little bit. (suggestion)
This does not follow from the fact that the two imperatives, intuitively, convey that
the speaker wants them to be fulfilled. As I noted before, it is quite consistent to
have inconsistent desires—unless one decides to act on them:
(6.10)
I want to stay inside to avoid getting a sunburn, but I also want to go out
and play!
Why, then, can (6.9) not be taken to be an indication of such inconsistent desires?
CHAPTER 6. COMMITMENTS TO PREFERENCES
142
On the present account that is because (6.10) involves commitment to effective
preferences, which need to be consistent.1
Speaker endorsement
A speaker who utters an imperative with content p cannot,
without being subject to blame, act so as to prevent the realization of p. This
is something that is universally true for all uses of imperatives, even ones that
intuitively do not involve a desire of the hearer or deontic authority, such as the
advice-use in (6.2). As a consequence, certain continuations are infelicitous after
an imperative, regardless of which ‘force’ the imperative has:
(6.11)
a.
Go, # but I won’t allow you to.
b.
Go, # but I don’t want you to.
Declarations that the speaker would prefer it if his imperative were not fulfilled are
possible, but only if they explicitly concern non-effective preferences, as is arguably
the case for wish:
(6.12)
Okay, go to the party, but I wish you would not.
Functional heterogeneity
The main idea underlying a uniform analysis of im-
peratives (instead of one relying on underspecification, such as those of Kaufmann
(2012) and Portner (2007)) is that the distinctive features of the varied uses of
imperatives can be explained by interaction of the uniform dynamic effect of imperatives together with contextual conditions and pragmatic reasoning. The rest of
this section illustrates how we can model this interaction in the system of dynamic
pragmatics.
The presentation will focus on two exemplary classes of uses. On the one
hand, Section 6.1.2 discusses uses that are directive in a narrow sense, such as
1
Kaufmann (2012)’s account arguably is unable to derive the consistency requirement. Portner
(2007) appears to intend to simply stipulate it as a requirement on ToDo-lists—-but in Portner
(2012), the author suggests that allowing for inconsistency is the right way to deal with Permission
uses. If so, then Portner’s account cannot jointly account for Permission uses and the consistency
constraint. See Condoravdi and Lauer (2012) for discussion.
CHAPTER 6. COMMITMENTS TO PREFERENCES
143
orders, requests, and pleas. These are the stereotypical uses of imperatives, and
it is not immediately obvious that our analysis predicts them. On the other hand,
Section 6.1.3 focuses on advice uses. These are particularly interesting from the
perspective of dynamic pragmatics, as they involve quite intricate inferencing
of the Gricean sort, whereas directive uses do not. Relatedly, advice uses are
somewhat unique among the uses of imperatives in that their main point often
seems to be to convey information (about how a certain goal can be reached), which
is the stereotypical purpose of declaratives.
6.1.2
Deriving directive uses
(6.13) illustrates some directive uses of imperatives. Each example is paired with
a reportatively used declarative that can be used to describe the corresponding
imperative after the fact.
(6.13)
a.
b.
(i)
Stand at attention!
(ii)
He ordered me to stand at attention.
(i)
Hand me the salt, please.
(ii)
He requested to be passed the salt.
(Command)
(Request)
We want to answer two questions about directive uses: (i) Under which circumstances can an utterance of an imperative be described with a directive verb of
saying such as order and request? (ii) Why do utterances of imperatives, in these
contexts, act as an inducement for the addressee to fulfill them? (Why can they
create an effective preference of the addressee for fulfilling them?)
Imperatives and directive verbs of saying
I take order and request2 to have the same truth-conditional content, summarized
in (6.14), and assume they differ only in their presuppositions. Roughly, order
2
In Condoravdi and Lauer (2012), we also took warn to be equivalent to these directive predicates
in terms of its truth-conditional content, but I am no longer sure that this is adequate—in particular,
I think that the preference expressed by a warning does not necessarily require that the addressee
form an effective preference. Warnings seem more akin to cases of advice in this way.
CHAPTER 6. COMMITMENTS TO PREFERENCES
144
presupposes that the speaker believes himself to have deontic authority over the
addressee, whereas request presupposes that the speaker takes it to be likely that
p does not interfere with the current effective preferences of the addressee.
(6.14)
utteru paq and Resultu ppepa pepb ppqqq
The assumed dynamic effect of imperatives means (6.15) will be true after the
utterances of imperative with content p:
(6.15)
utteru paq and Resultu ppepa ppqq
Our task is now to specify properties of the context under which (6.15) entails (6.14).
The condition I want to propose here is the following: Directive uses arise only if
the addressee is taken to have control about p, that is, if it is taken for granted that
p will be realized if and only if b’s preferences are such that he would choose it.
(6.16)
PBa pw, tq p ô epb ppq
(6.16) is indeed sufficient to ensure that (6.15) entails (6.14): If it is true, then p and
epb ppq are indistinguishable in the speaker’s public belief state, and hence for any
preference structure PEP in which p is a maximal element: PEP „a,t,w PEP ` epb ppq.
So if (6.16) holds, then an utterance of an imperative automatically ensures
that the truth conditions of directive predicates are satisfied. But it is a rather
strong condition. Is it justified to assume it holds whenever an imperative is used
with directive force? I will try to motivate the two directions of the implication
separately.
Motivating p ñ epb ppq.
This direction can be motivated rather easily. It says that
p will be actualized only if b’s preferences require it, i.e., that the actualization of p
depends on b’s choice. This is a natural assumption for many agentive predicates,
e.g., the ones exemplified in (6.17).
CHAPTER 6. COMMITMENTS TO PREFERENCES
(6.17)
a.
Close the window!
b.
Leave!
c.
Pay your taxes!
d.
Call me!
145
Not all directive uses of imperatives involve predicates that are necessarily agentive. In particular, such directive uses can happen with stative predicates and those
that involve refraining from actions (which arguably are stative, too), as in (6.18).
(6.18)
a.
Sit still!
b.
Don’t say another word!
It seems plausible, however, to say that a speaker will utter such imperatives (as
directives) only if he takes it for granted that they will not become realized unless
the addressee forms the correct effective preference.3
Motivating p ð epb ppq.
It is perhaps not quite as obvious that agentivity, insofar
as it is required for directive imperatives, also ensures the opposite implication,
i.e., that p will be actualized if b effectively prefers it. This is certainly plausible
for imperatives such as Leave or Shut up, but does it hold in general? There is
some reason to think that when an imperative is used directively, it does. Suppose
the speaker has authority over the addressee, and wants him to be at the airport at
noon, but he knows that whether the addressee can be there in time depends on
factors like the traffic conditions, which are not under the addressee’s control. In
that case, it seems a speaker would use (6.19a) rather than (6.19b).
(6.19)
a.
Try to be at the airport at noon!
b.
Be at the airport at noon!
It does seem intuitively correct to say that if a speaker uses (6.19b) (as a command
or request), he indeed must presuppose that the addressee can ensure that he is
3
Perhaps it would be adequate to add this speaker presumption as another presupposition for
order, request, etc.
CHAPTER 6. COMMITMENTS TO PREFERENCES
146
at the airport at noon. This strongly suggests that, for directive uses, it is indeed
necessary that p is considered, in the given context, to be under the control of the
addressee in the strong sense demanded by (6.16). I have to concede, however,
that I cannot exhaustively demonstrate at present that imperatives can only have
directive force if (6.16) holds. Observations like (6.19) suggest that this is so, but it
might turn out that while (6.16) is a sufficient condition for directive uses, it is not
a necessary one.
I want to stress that in the present account (6.16) is just a contextual condition which
may or may not be in place in a given context. It is not a feature conventionally
associated with imperatives. In the past, (6.16) has sometimes been taken to be a
general property of all utterances of imperatives, as in Belnap and Perloff (1990)’s
imperative content thesis (for these authors, the term ‘agentive’ implies (6.16)):
(6.20)
Imperative Content Thesis (Belnap and Perloff 1990, p. 173)
Regardless of its force, the content of every imperative is agentive.
But this is too strong as a requirement for all imperative uses: Recall that imperatives (at least in English) can be used to express mere desires, as in well-wishes,
‘absent wishes’:4
(6.21)
a.
Get well soon!
b.
[On the way to a blind date, to oneself]
Be blond/rich/nice!
This indicates that it is wrong to assume that the Imperative Content Thesis
captures a conventional fact about imperatives. Rather, it articulates a contextual
4
The requirement is also too strong in order account for ‘advertising imperatives’, like (i) and
(ii), as Paul Kiparsky (p.c.) points out:
(i)
Win a trip to Finland!
(ii)
Be the first on your block to own a Tesla!
CHAPTER 6. COMMITMENTS TO PREFERENCES
147
condition that may or may not be in place, and that is necessary for directive uses
only. The set-up of the dynamic pragmatics lets us see how contextually-assumed
agentivity results in directive uses: Directive predicates like order and request
apply to imperative utterances only if the speaker of the imperative takes it for
granted that (6.16) holds.
Directive imperatives act as inducements to action
Why do directive imperatives (i.e., imperatives with contents that require an addressee choice for their realization) typically act as incentives to action? This will
depend on socio-cultural circumstance: An imperative functions as an order only
if the speaker has deontic authority over the addressee, which we can now cash
out as follows:
(6.22)
An agent a has deontic authority over another agent b, with respect to
p if pepa,t ppq implies that, at t, b is (socially/legally/institutionally/. . . )
obligated to act though he prefers p.5
If the speaker does not have deontic authority in this sense, the extent to which
the addressee will be induced to fulfill the imperative will depend in complex ways
on his other preferences. In many cases of requests, they will be mediated by a
desire to help the speaker to fulfill (publicized) goals, which may be motivated in
various ways—by a desire to stay in the good graces of the speaker, for example,
or by genuine concern for a friend’s well-being.
In the limiting case, though, why would a request made by one stranger to another ever induce the addressee to act? I propose, again following Condoravdi and
Lauer (2012), that agents are generally cooperative-by-default, in the sense defined in
(6.23).
5
I do not treat permission and concession uses here, but this dependence of the obligation
on a public preference will play an instrumental role in accounting for those in the current setup—permission for p can be viewed as the retraction or modification of the permitter’s existing
preferential commitment for p. Given the characterization of deontic authority employed here,
such a rescission or modification will remove the prohibition against p.
CHAPTER 6. COMMITMENTS TO PREFERENCES
(6.23)
148
Cooperativity by default
An agent a is cooperative-by-default iff he adds any topical goal g of
another agent to his effective preference structure.
(6.23) is very weak: It only requires that topical goals of other agents be added to
the effective preference structure, but not that they be ranked in a particular way.
The effect this has is that an agent will not thwart another agent’s goals without
reason, i.e., randomly. The reason may be entirely selfish—laziness, say—but (6.23)
requires that, if an agent does not act to help others fulfill their publicized goals,
he must have some selfish preference motivating this. To realize how weak this
principle is, observe that the reason may even be pure antagonism: An agent
who fails to aid another simply because he takes pleasure in seeing others’ goals
thwarted can still be cooperative-by-default, if he ranks his desire to see others
suffer over the goals of others.
(6.23) ensures that if the action requested by a speaker does not ‘cost’ the addressee anything, i.e., if it does not interfere with his private preferences, he will act
so as to fulfill the request. If we assume (6.23) is generally true of (social) agents,
we can hence explain why a speaker might have reason to assume he can get a
complete stranger to move by uttering (6.24).
(6.24)
6.1.3
Step out of the way, please.
Deriving advice uses
Advice uses come in two kinds: Those in which it is taken for granted that the
speaker and addressee share a salient goal, and those where this assumption is
not warranted. (6.25) is an example of the first kind, assuming that a conversation
between doctor and patient presupposes that both want the addressee to be healthy.
(6.25)
[Doctor to patient]
Take these pills for a week.
CHAPTER 6. COMMITMENTS TO PREFERENCES
149
(6.26) is an example of the second kind: Between strangers, it is not plausible to
assume that they share a topical goal.
(6.26)
[Strangers in the streets of Palo Alto]
A : Excuse me, how do I get to San Francisco?
B : Take the north-bound Caltrain.
Here, I want to focus on the first kind of case, supposing that the second kind can
be assimilated to the first if we assume agents to be cooperative-by-default.
Advice cases are quite interesting from a pragmatic perspective, because they
involve an imperative that does not act as an inducement to action (the way an
order, request or plea does), but also gives information about the satisfaction of a
goal: (6.26a) makes it so that the addressee can conclude something along the
lines of (6.27), and that may well be the main purpose of the speaker in uttering the
imperative.
(6.27)
Taking the pills is the best action to further the goal of curing the addressee’s current illness.
We want to understand how this happens. Let pills be a proposition letter standing
for the proposition that the addressee takes the pills and cure be a proposition
letter standing for the proposition that the addressee gets cured. Then we can
symbolically represent (6.27) as:
(6.28)
Bestpcure, pillsq
(6.28) stands for a possibly quite complex set of facts involving medical knowledge
and standards of certainty, etc., so it would be futile to define it as a proposition in
our current system. I use it as an abbreviation for the natural language statement
in (6.27) here.
The addressee will come to believe (6.27) through an utterance of !pills if the
following assumptions hold throughout his belief state.
CHAPTER 6. COMMITMENTS TO PREFERENCES
(6.29)
a.
150
Sincerity
The doctor S will commit to a preference for ϕ only if he has it.
@ϕ : pepS pϕq ñ epS pϕq
b.
Goal
The doctor S shares the (topical, salient) addressee’s goal of curing
his current illness.
epS pcureq
c.
Non-interference
The doctor has no other (topical, salient) preferences relevant to the
addressee’s taking the pills.
d.
Expertise
If the doctor thinks that ϕ is the best way to realize cure, it is.
@ϕ : BS pBestpcure, ϕqq ñ Bestpcure, ϕq
Assuming that the addressee’s pre-utterance information state BA supports all these
conditions, we can derive (6.30d) (omitting all temporal parameters):
(6.30)
a.
BA rutterpS, p!pillsqqs “ Resultu ppepS ppillsqq
b.
BA rutterpS, p!pillsqqs “ pepS ppillsq
c.
BA rutterpS, p!pillsqqs “ epS ppillsq
d.
@w P BA rutterpS, p!pillsqqs :
EPS pwq „S,t,w EPS pwq ` pills
(Decl. Convention)
(Result is factive)
(Sincerity)
(def. ep)
Let us make, for a second, the (unrealistic) simplifying assumption that BA also
supports the assumption that the speaker has no effective preference other than
cure. Then (6.30d) comes down to (6.31):
(6.31)
@w P BA rutterpS, p!pillsqqs :
tcureu , H „S,t,w tcure, pillsu , H
(6.31) is true if, according to the speaker’s information state BS pwq, there is no
decision situation in which jointly satisfying cure and pills requires a different
action than satisfying cure alone. Assuming that the speaker believes that he has
CHAPTER 6. COMMITMENTS TO PREFERENCES
151
the ability to ensure pills at all6 ), this means that he must believe that his preference
for cure is best satisfied by ensuring pills, that is, the speaker must believe that
Bestpcure, pillsq.
Things would not change if the addressee takes it to be possible that the speaker
has other effective preferences, as long as these preferences are orthogonal to taking
the pills—i.e., that none of his other preferences plays a role in action choices having
to do with ensuring pills. Indeed, A may realistically assume that there are many
such preferences—e.g., the doctor might prefer to have a beer at the end of the
day, that his daughter gets into a good college, that England wins the world soccer
cup, and so forth. None of these preferences (presumably) are relevant in choosing
whether to ensure pills becomes true.
So, by Non-interference, we can conclude that
(6.32)
a.
BA rutterpS, p!pillsqqs BS pBestpcure, pillsqq
b.
BA rutterpS, p!pillsqqs Bestpcure, pillsq
(by Expertise)
But (6.32b) is just what we wanted to derive: If an addressee has an epistemic
state BA which supports our contextual assumptions, then observing an utterance
of !pills will make him believe that the best way to cure his current illness is to take
the pills. The contextual assumptions are rather specific, of course, but we can now
investigate what happens if we selectively give them up.
For example, let us assume that the addressee does not trust the medical knowledge of the doctor, i.e., we have all assumptions in place except for Expertise. We
no longer predict that (6.32b) holds, but we still predict that (6.32a) holds: The
addressee learns that the doctor believes that taking the pills is the best way to cure
him.
Perhaps more interestingly, let us hypothesize that the addressee fails to believe
in Non-Interference. Concretely, he knows (or takes it to be possible) that the
6
That is, that he believes he can convince the addressee to take the pills. Note that we need not
assume here that (the addressee believes that) the speaker believes that his current utterance will
suffice to convince the addressee. The speaker could believe that, but he could also simply take it to
be possible to convince the addressee with a later utterance or other action—e.g., by showing him
a medical study that establishes the efficacy of the pills.
CHAPTER 6. COMMITMENTS TO PREFERENCES
152
doctor gets a kickback from the manufacturer of the pills whenever a patient
fulfills a prescription for them. This alone would not change things, if the patient
still believes that the doctor’s desire for money is kept strictly dominated (in virtue
of his professional ethics) by his desire to cure his patient. But let us assume—
violating belief in Non-interference—the patient takes it to be possible that the
doctor’s desire for getting the kickback is active, i.e., that:
(6.33)
Dv1 P BA pwq : v epS pkickbackq
Given that making the patient take the pills will make kickback true, an (undominated) effective preference for kickback is interfering. In v-type worlds, we cannot
conclude from (6.34) that the speaker believes that taking the pills is the best way
to cure the patient.
(6.34)
EPS pvq „BS pvq EPS pvq ` pills
The doctor’s decision for pills in v might also be motivated by his desire for kickback.
So if the patient is suspicious in this way, we derive the weaker conclusion in (6.35):
(6.35)
BA rutterpS, p!pillsqqs Bestpcure, pillsq _ epS pkickbackq
This accords with intuition: If the patient thinks that the doctor has incentive to
prescribe him the wrong pills, he will come to believe that either taking the pills
will cure him, or the doctor wants to profit from his (presumed) gullibility.
We need not make such sinister assumptions. We derive a similar effect if
we assume that the doctor and patient share another benign preference that is
‘interfering’. Suppose, for example, that the patient does not have health-insurance
(or large amounts of spare money), and hence can only afford generics. This is a
benign interfering preference, but it is an interfering preference: The patient will no
longer conclude (necessarily) that taking the pills is the best way to get him cured,
but rather that taking the pills is the best way to get him cured with generics—there
might be a medication that is better for him, but does not have a generic equivalent.
So we get the strong implication that the patient comes to believe the pills to be
CHAPTER 6. COMMITMENTS TO PREFERENCES
153
the best way to cure him only if we assume him to believe there are no interfering
preferences—otherwise, he will believe that taking the pills is the best way to cure
him, taking into account his other preferences. This again accords with intuition.
We thus see how the hearer’s inference depends in complex ways on his assumptions about speaker beliefs and preferences (and whether these preferences
are well-aligned with his own). We can now generalize the conditions under which
we expect an imperative to give advice relative to a salient shared goal:
(6.36)
An addressee of an imperative p! will conclude that p is the best way to
satisfy g if he believes the following:
a.
The speaker and the addressee share the goal g.
b.
The speaker does not have any other effective preference that is plausibly relevant to decisions about p.
c.
The speaker is an epistemic authority on ways to achieve g.
d.
The speaker will only commit to preferences he actually has.
e.
The speaker takes it to be possible that there is a way to convince the
addressee to perform p.
Salience, topic-hood and the QUD
The foregoing assumed that there is only one shared goal between the interlocutors,
and motivated this by assuming it is the only salient or relevant one. It would be
desirable if our model included a mechanism that explains (or at least constrains)
how this happens. We only have to enrich the utterance situation slightly to make
this assumption worthy of exploration. Doctors and patients may have a number
of (shared) preferences in addition to curing the patient’s current illness (if it is even
established that there is a single illness to be cured). For example, it is reasonable
to assume that they also desire to prevent future illnesses that may be unrelated
to the current one. Even if such a background goal exists, a patient might well
(reasonably! ) infer that the doctor’s utterance of Take these pills for a week!
pertains to the current illness—but why?
CHAPTER 6. COMMITMENTS TO PREFERENCES
154
I believe the correct answer has indeed to do with salience, and perhaps topichood of the conversation. The doctor will not utter his imperative out of the blue;
it likely is made as part of a longer conversation about how to cure what ails the
patient right now. Developing a mechanism to model topic-hood is well beyond
the scope of this thesis.
I want to briefly mention a dominant approach to capture something like topichood or salience, and explain why I don’t think it helps with the problem we
face here. In question under discussion models (Roberts 1996, Ginzburg 1996), every
utterance is taken to address a salient question that is represented as the top-level
element of a QUD stack (or a more general structure—Ginzburg allows the order
of the possible QUDs to be partial). Perhaps we should integrate such widelyemployed representational parameters into our model?
This seems tempting, for isn’t the tendency to infer that taking the pills will cure
the current illness particularly strong if the doctor utters it in response to an overt
question like (6.37)?
(6.37)
Doctor, what can I do to fix this?
Indeed, it is. And given that the QUD model allows the question that an utterance
addresses to be implicit, we could hypothesize that even if the patient does not
utter an interrogative like (6.37), but only describes his problems, a QUD like (6.37)
can be inferred.
But it is not clear what we gain from this assumption. Why does (6.37)—
whether it is implicit or overt—exclude other possible preferences from consideration? What’s worse, we do not even always want to exclude other possible
preferences—recall the preference for only taking generic medications above. So if
the QUD model is to have any bite (in the sense that if the QUD is (6.37), the only
preference taken into account is the preference for fixing the current problem), we
would need to assume that the QUD in the generic scenario is instead (6.38).
(6.38)
Doctor, what can I do to fix this without paying for non-generics?
CHAPTER 6. COMMITMENTS TO PREFERENCES
155
This may not be implausible as a QUD, but the question remains: How does the
QUD model help us to understand which considerations should enter into the
interpretation of the imperative utterance and which should not? I submit that it
does not, unless it is paired with a substantive theory about how QUDs evolve in
dialogue, and how they relate to conversational-external goals of the interlocutors.
This is not to say that the preference-based model laid out in this thesis is any
better than the QUD model in telling us how salience and topic-hood influence
which preferences and beliefs are taken into account at any given point in the
conversation—it plainly is not. But it is also not clear how a QUD-less preferencebased model is any worse. While QUDs intuitively relate to topic-hood (and may be
helpful in explaining other information-structural properties), assuming that the
QUD is (6.38) is not any more illuminating than hypothesizing that the only two
preferences taken into account in the interpretation of the imperative are those for
fixing the current problem and only buying generic medications.
So I shall not extend the system of dynamic pragmatics with a QUD representation, and leave the proper way to model salience and topic-hood in the system
for future research.
6.1.4
Conclusion
In this section, I have illustrated how the system of dynamic pragmatics helps us
to spell out, in a quite detailed fashion, how some of the varied uses of imperatives
arise on the basis of a uniform dynamic effect.
A theory such as that of Condoravdi and Lauer (2012) arguably requires a rich
theory of pragmatic inference to account for the functional heterogeneity of imperatives. A theory that instead relies on semantic underspecification, such as that
of Kaufmann (2012), may be seen as less dependent on pragmatic inference, but
this comes at a price. In order to correctly account for advice uses, Kaufmann
(2012) imposes the following disjunction as one of a number of felicity conditions
on utterances of imperatives (c is the context of utterance, g is the ‘ordering source’
parameter of a Kratzerian modal, ∆ is a function that specifies the decision problem
CHAPTER 6. COMMITMENTS TO PREFERENCES
156
(if any) that the interlocutors are trying to resolve in a context):
(6.39)
Ordering source restriction (Kaufmann 2012, p. 162)
either (i) in c there is a salient decision problem ∆pcq Ă ℘pWq such that in
c the imperative provides an answer to it, g is any prioritizing ordering
source, and speaker and addressee consider g the relevant criteria for
resolving ∆pcq;
or else, (ii) in c there is no salient decision problem ∆pcq such that the
imperative provides an answer to it in c, and g is speaker bouletic.
In order to account for advice uses, Kaufmann has to hard-code the fact that
the imperative provides a solution to a decision problem into the conventional
felicity conditions of imperatives. Because of this, she then is forced to provide the
second disjunct as an escape-hatch, in order to be able to account for other uses
of imperatives that are not connected to the resolution of decision problems at all
(such as well-wishes and absent wishes, for example).
The analysis adopted here leaves reasoning about decision problems entirely
in the pragmatics. As a result, the conventional meaning and dynamic effect of
imperatives can be greatly simplified.
6.2
Performative uses of desideratives
In Condoravdi and Lauer (2009), we drew attention to the fact that assertions about
speaker desires can be used to perform a range of speech acts that is strikingly
similar to those that can be performed by means of imperatives:
(6.40)
a.
[Mother to child]
I want you to clean your room before playing!
b.
[Mother to child]
You do NOT want to touch that cookie!
c.
(Command)
(Prohibition/Warning)
[Doctor to patient]
I want you to take these pills for a week.
(Advice)
CHAPTER 6. COMMITMENTS TO PREFERENCES
d.
157
[Recipe]
You want to stir the mixture well.
e.
(Advice)
[Affirming an offer]
No, really, I want you to take the last cigarette.
f.
(Invitation)
[Among collaborators]
I want you to write this up before our next meeting.
g.
If it is that important to you, I want you to go.
(Request)
(Concession)
We spelled out an analysis of theses uses in much the same way as I did above for
imperatives, but without appeal to the crucial notion of commitment: We specified
contextual conditions under which the use in question would arise. As an example,
a crucial ingredient for deriving order uses was the condition in (6.41):
(6.41)
Speaker Authority over the hearer: It is commonly assumed that if the
hearer believes that the speaker prefers unsettled, hearer-controllable φ
over not-φ, the hearer is socially or institutionally obligated to execute an
action that ensures that φ.
Besides omitting reference to the notion of commitment (or rather, approximating this notion by reference to what is ‘commonly assumed’, i.e., public knowledge), this formulation does not employ the distinction between effective and other
preferences—the obligation arising from an order-use of a desiderative assertion is
tied to the addressee knowing about any desire of the speaker.
As Sam Cummings (p.c.) correctly observed at the time, this predicts that if the
speaker is taken to have authority, then any assertion of a speaker desire should
count as a command. But this is not correct, the same mother that gives a command
with (6.40a) could employ (6.42) only to inform the child that she desires him to
clean his room, but leave up to the child if it wants to clean it (picture a mother
who firmly believes in anti-authoritarian parenting).
(6.42)
I want you to clean up your room, honey. Nothing would make me
happier.
CHAPTER 6. COMMITMENTS TO PREFERENCES
158
In Condoravdi and Lauer (2009), we took want to simply entail the existence of an
unspecified desire, which initially seems plausible. But in our work on anankastic
conditionals (Condoravdi and Lauer to appear), Cleo Condoravdi and I have since
argued that there is reason to think that the meaning of want is underspecified,
much in the way in which the meaning of modal auxiliaries is usually taken to be
underspecified. A similar intuition was expressed by Hare (1968) when considering
the contrast between (6.43) and (6.44).7
(6.43)
If you want to have sugar in your soup, you should ask the waiter.
(6.44)
If you want to have sugar in your soup, you should get tested for diabetes.
“Let us consider the meaning of ‘If you want’ in the two cases. In the
‘diabetes’ case, a first approximation would be to say that it means the
same as ‘If you, as a matter of psychological fact, have a desire’. I
am very much inclined to deny that it means anything like this in the
‘waiter’ case.”
(Hare 1968)
Given the way I have modeled preferences and desires so far, we can assume that
want depends on a contextual parameter, viz., a preference structure, which may
be the agent’s effective preference structure, or it may be one of the underlying
preference structures which model psychological desires, inclinations, appetites,
etc. We could then say that the ‘performative’ readings arise only when want
targets the agent’s effective preferences.
To make this work, at least without modifying the set-up of the system so far,
however, we need another ‘introspection’ principle for commitment to beliefs and
preferences, namely the one in (6.45).
(6.45)
If epa,t ppq P PBa pt, wq, then p P PEPa pt, wq.
Compare (6.45) with the principle we already have introduced in Section 5.3.2:
7
See Levinson (2003), and references therein, for similar observations.
CHAPTER 6. COMMITMENTS TO PREFERENCES
(6.46)
159
If pepa,t ppq P PBa pt, wq then p P PEPa pt, wq.
(6.46) ensures that if an agent is doxastically committed to being committed to a
preference, he is also (plainly) committed to that preference. (6.45) instead says
that if an agent is committed to believe he has a preference, he is also committed
to that preference. These principles sound so similar that it may be tempting to
assume that if one of them is appropriate, so is the other. This is not quite the case,
as we shall see: I think that (6.46) is most obviously necessary, but things are far
less clear with (6.45).
But let us assume (6.45) is in place. Importing the ep-predicate into our object
language, we can represent (6.47a) as (6.47b).
(6.47)
a.
Mother: I want you to clean your room NOW!
b.
epm pcleanq
And then proceed in the now familiar manner:
(6.48)
a.
BA rutteru pm, pepm pcleanqqqs Resultu ppbm pepm pcleanqqq
(Decl. Convention)
BA rutteru pm, pepm pcleanqqqs pbm pepm pcleanqq
BA rutterpu, m, pepm pcleanqqqs pepm pcleanq
(Result factive)
(6.45)
(6.48c) is of course just what we derive for imperatives, as well, so from this point
on we can derive the performative effect of the want sentence in just the way
we would do this for the imperative Clean your room now. And given that the
derivation in (6.48) does not make any use of contextual assumptions, we could
do the same for other uses, as well. To the extent that we find the explanation of
the functional heterogeneity of imperatives convincing, we have an account of the
varied uses of want as well.8 This is quite unified and quite pleasing, assuming is
a plausible principle.
8
Assuming we can also explain why some of these want uses require rather particular forms—
e.g., while we can reinforce an offer with want, it is not clear we can make an initial one. Similarly,
order-uses do generally require a certain intonation.
CHAPTER 6. COMMITMENTS TO PREFERENCES
160
Doing away with PEP?
Given the success with explaining why desiderative assertions (if want targets effective preferences) have performative uses like imperatives, it may seem tempting
to turn around and simplify the analysis of imperatives, as well, and indeed of the
whole system, in the following way.
We get rid of the independent specification of PEP in the models, and instead
define the pep-operator as an abbreviation:
(6.49)
pepa pφq :“ pba pepa pφqq
The desired introspection properties follow immediately, both the pbppepq and
the pbpepq version. And we can simply represent pb as an accessibility relation,
without worrying about how to keep the PEPs in sync. Imperatives simply become
assertions about effective preference. And our models are simpler, because they
contain one fewer operator. What is not to love about this solution?
This solution would, indeed be pleasing in its representational parsimony. But
it does lack one essential feature of the analysis that independently represents
PEPs. Commitments to effective preferences, like all commitments, are stable: They
persist until and unless they are rescinded. The same is not true for plain effective
preferences (or, realistically speaking, beliefs). Effective preferences change over
time, and speakers are aware of this. They might form new underlying preferences, or their evaluation of their relative importance may change. And even if this
does not happen, agents will frequently have to revise their effective preferences
to maintain consistency (e.g., when they learn that two previously unranked preferences are in fact incompatible). But if effective preferences are not temporally
stable, then commitment to preferences, as defined in (6.49), will not be temporally
stable, either. This becomes apparent if we put in the omitted temporal parameters:
(6.50)
pepa,t pφq :“ pba,t pepa,t pφqq
Now consider what happens at a later time t1 . Given that doxastic commitment is
stable, we still have:
CHAPTER 6. COMMITMENTS TO PREFERENCES
(6.51)
161
pba,t1 pepa,t pφqq
But that is not the same, nor does it entail (given that eps can change) that
(6.52)
pepa,t1 pφq :“ pba,t1 pepa,t1 pφqq
And we certainly do not want to revise ep to be timeless—speakers can and do
speak about the fact that their preferences have changed in the past or might
change in the future, and there is no reason to believe that such talk is always about
non-effective preferences.
So it seems we cannot do without an independent representation of the notion of
‘commitment to preferences’, next to ‘commitment to belief’, because without one,
only the latter notion would be stable—and we need commitment to preferences
to be stable if we want to use the notion to explain the workings of predicates
like promise: A speaker who promises to get Sally from the airport does not just
commit to act as though he wants to get her unless he changes his mind. If he fails
to show up at the airport, he cannot defend himself by arguing ‘No, no, I did not
violate my commitment. I acted according to the preference for being there for a
while, but two hours before I had to leave, I changed my mind, and so I did not
go.’ Such an argument would be ridiculous—and on the PEP-analysis, we can say
why: Because commitment to preferences, like all commitment, is stable. On the
pba,t pepa,t q-analysis of preferential commitment, this argument would be perfectly
plausible. The agent had the requisite commitment at the time, and he has not
since acted in a way that is incompatible with him believing that he did—until he
changed his mind. But that is not how commitment works.
Unfortunately, these considerations also cast some doubt on the plausibility of
the pbpepq-introspection property proposed above. This property does not cause
instability of peps, of course, but on reflection, it is not clear how plausible it is.
Why should doxastically committing oneself to having a preference at time t create
an enduring commitment to a preference for p? Wouldn’t it be just as natural, if not
more natural, to say that such a doxastic commitment is just that: What endures
is your commitment to believe that you had the preference in question at t, but
CHAPTER 6. COMMITMENTS TO PREFERENCES
162
that does not affect your persisting preferences to commitment. Committing to a
preference (indefinitely), and committing to believe to have a preference at t are
just two different kinds of things—why should the latter give rise to the former?
6.3
Interrogatives
In the present section, I want to sketch, in a very brief fashion, how interrogative
sentences can be integrated into the theory of the form–force mapping proposed
here.9 Of course, interrogatives are a complex topic, so this should be seen as a
pointer to future work rather than a fully spelled out proposal.
For concreteness, I assume that the interrogatives of Sen (which, due to the
propositional nature of the language, are all polar interrogatives), denote a partition
of the set of possible worlds, as in Groenendijk and Stokhof (1984):
!
)
Sen
Prop
Prop
v?ϕw “ vϕw
(6.53)
, v ϕw
Interrogatives, just like declaratives and imperatives, are associated with a convention of use, the Interrogative Convention:
(6.54)
Interrogative Convention (NL statement)
When a speaker utters an interrogative with content ?ϕ, he thereby incurs
the following commitment:
pepSp pDp P?ϕ : pbAddr ppqq
That is, with an interrogative ?ϕ, a speaker commits himself to a preference for
the addressee to be committed to one of the possible answers to φ. More briefly,
we can paraphrase this as: The speaker requests that the addressee assert one of
the possible answers to his question. Viewed this way, the proposal bears some
similarity to the ‘imperative-assertoric’ approach pioneered by Lewis and Lewis
9
This section summarizes the preliminary proposal for an analysis of interrogatives in Lauer and
Condoravdi (2012).
CHAPTER 6. COMMITMENTS TO PREFERENCES
163
(1975), which is arguably the analysis with the broadest coverage of uses among
the ‘imperative paraphrase’ accounts which were en vogue in philosophical logic
for in the 1960s and 1970s.10 According to Lewis and Lewis’ proposal (elaborated
somewhat in Åqvist (1983)), an interrogative ?ϕ is equivalent to the imperative ‘Tell
me truly whether v?ϕw’—and is thus a request for the true answer. The account
in the present framework does not involve the notion of truth, and it does not
specify that the speaker requests a speech act. Instead, he only is committed to a
preference for a doxastic commitment. The latter fact makes the present account
more promising in view of rhetorical questions, such as (6.55).11
(6.55)
[B does not stop complaining about how bad the movie was]
A: Well, who insisted that we see it?
On the present account, we can view rhetorical questions in the way Rohde (2006)
analyzed them, as ‘redundant interrogatives’—they are redundant because the
commitment that the speaker commits to preferring already exists.12
How does this proposal fare in accounting for the other uses of interrogatives?
A final answer to this question of course requires careful spelling out of the various
conditions that give rise to the various uses, but, at first glance, it does pretty well.
Consider the sample of uses in (6.56).
(6.56)
10
a.
Is it raining?
b.
What is the formula for sulphuric acid?
c.
[A is desperately looking for her keys]
(requesting information)
(testing knowledge)
The most notable exponent of this kind of theory was Åqvist (1965, et seq.).
It has sometimes been claimed (most recently and most explicitly by Han (2002)) that rhetorical
questions always are associated with (or even assert) their negative answers—i.e., no for a polar
question, no-one for a who-question, nothing for a what-question. Examples like (6.55) show that
this is not quite right. Rohde (2006) provides more data challenging this generalization.
12
Rohde also cashes out ‘redundancy’ in terms of prior commitment. She says that there is
a ‘felicity condition’ on the use of rhetorical questions ensuring prior commitment. This could
mean that she takes rhetorical questions to be associated with this condition by convention—thus
assuming that rhetorical questions constitute their own sentence type. In the present proposal,
we could simply define a rhetorical question as an interrogative that is uttered in the presence of
(obvious) pre-existing addressee-commitment (which might also be what Rohde had in mind).
11
CHAPTER 6. COMMITMENTS TO PREFERENCES
B: Could they be in the car?
d.
164
(pointing out a possibility)
Senator, should taxes be raised to balance the budget?
(combative
question)
e.
And doesn’t this line bisect each of these spaces? (Socratic question)
Combative questions like (6.56d) come out as somewhat of a base case—they can be
taken to be motivated solely by a desire to get the addressee to commit. Information
questions like (6.56a) arise straightforwardly in situations in which the speaker
lacks information, assumes that the addressee has the information, and further
assumes that the addressee would only commit himself in something he believes
to be true. Exam questions like (6.56b) similarly do not pose a problem. In this
case, the speaker does not ask to acquire information, but rather simply to get the
addressee to commit to an answer, so that he can evaluate his knowledge. Socratic
questions like (6.56f) can be viewed as cases where the speaker wants the addressee
to consider the correct answer to the question, and forces him to do so by requesting
a commitment. Similar things can be said about cases like (6.56d).
Again, this is meant only as a short illustration of how an account of interrogatives would fit into the current, commitment-based view of clause-typing. Many of
the details need to be spelled out. The suggested analysis holds promise, however,
for a successful treatment of interrogatives that covers their varied uses.
6.4
Denotation types and the form–force mapping
In the statement of the conventions of use, and the associated discussion, I have
assumed that declaratives and imperatives both denote propositions, while interrogatives denote sets of those. In terms of the semantics of the language Sen:
(6.57)
a.
v$ ϕwSen “ vϕwProp
b.
Prop
v!ϕwSen “ vϕw
!
)
Sen
Prop
Prop
v?ϕw “ vϕw
, v ϕw
c.
CHAPTER 6. COMMITMENTS TO PREFERENCES
165
In the present section, I want to show that these assumptions are non-essential:
While the conception of the form–force mapping proposed in this dissertation
(and Condoravdi and Lauer 2011 and Condoravdi and Lauer 2012) puts some
constraints on denotational types, it by no means requires the types in (6.57). In
this section, I will illustrate how the theory is compatible with a wide variety of
assumptions about denotational types. In each case, I will not argue in favor of any
particular proposal (such arguments would have to rely on considerations external
to the theory proposed here). The intention is simply to demonstrate how various
proposals for denotation types could be adopted in the current theory without
changing its essential nature.
The present section concerns possibilities that maintain the assumption that
sentences have the same kind of denotation in their embedded and matrix uses.
Section 6.5 will discuss the possibility of lifting this assumption in order to represent
force directly in the semantic denotations of sentences, as proposed by Krifka (to
appear).
6.4.1
Denotations for imperatives
We start with imperatives. There is great variation in the semantic literature with
respect to the question of what imperatives denote, and the choice made in the
preceding discussion—a proposition that characterizes the fulfillment conditions
of the imperative—is not a popular one.13 What I want to show in the sequel is
that, in the present set-up, we could adopt any sensible proposal for what imperatives denote. All that adopting a non-propositional denotation requires is a slight
modification of the Imperative Convention, repeated here for the propositional
case:14
13
The reason for this is not entirely clear to me—one consideration seems to be the pre-theoretical
intuition that imperatives, unlike declaratives, do not relate to truth, and cannot be used to make
claims. But that does not mean their compositionally-determined denotations cannot be propositions. As pointed out in Section 3.2, a proposition (in the linguist’s or logician’s sense—i.e., a set of
possible worlds), by itself, is no more closely related to truth and claiming as it is to fulfillment or
desiring.
14
Throughout this section, I use a natural language formulation of the conventions, as those
are much easier to read then the corresponding constraints on pragmatic models. I spell out the
CHAPTER 6. COMMITMENTS TO PREFERENCES
(6.58)
166
Imperative Convention (propositional denotation)
A speaker Sp who utters an imperative with content p thereby incurs the
following commitment:
pepSp ppq
Properties Hausser (1978) proposes that imperatives denote properties of individuals: “roughly that property which the speaker wants the hearer to acquire”
(p. 84). We could easily adopt this analysis, replacing (6.58) with (6.59).
(6.59)
Imperative Convention (property denotation)
A speaker Sp who utters an imperative towards addressee Ad with content
P thereby incurs the following commitment:
pepSp pPpAdqq
Predicates restricted to addressees
Portner (2005) adopts a variant of Hausser
(1978)’s analysis. Motivated by the fact that imperatives can have overt subjects
(such as Everyone go home!), Portner proposes that imperatives always have subjects, assuming a null subject referring to the addressee(s) when no overt subject is
present.15 He then takes imperatives to denote predicates that are presuppositionally restricted to only apply to the subject. i.e., Leave! has the denotation in (6.60)
(where c is the context of use):
(6.60)
λx : x “ addresseepcq . x leaves
normative consequences in terms of the operators pb and pep (omitting all temporal parameters),
though.
Another possible motivation to give imperatives non-propositional denotations may be the fact
imperatives rarely embed, and generally cannot occur as complements to predicates that take
declarative complements. If imperatives have a non-propositional denotation, this can be seen as
due to a type clash.
15
This assumption is commonly made, see Kaufmann (2012, Chapter3, 105–122) for extensive
discussion.
CHAPTER 6. COMMITMENTS TO PREFERENCES
167
Again, we could adopt this analysis easily. All we need to do is to define a function
def that takes predicates into the individual(s) for which they are defined.16 Then
our convention would go as follows:
(6.61)
Imperative Convention (restricted predicate denotation)
A speaker Sp who utters an imperative towards addressee Ad with content
P thereby incurs the following commitment:
pepSp pPpdefpPqqq
Event descriptions
A final example I want to mention is the idea that imperatives
denote something like actions or predicates of actions, or something of the kind
(Mastop (2005) is a recent exponent of this idea). This idea is quite intuitive as long
as one focuses on the stereotypical uses of imperatives, i.e., directives. In some
sense, these are always about actions (cf. Section 6.1.2), even when they involve
otherwise non-agentive predicates. It may hence seem plausible to say that such
non-agentive predicates somehow get ‘coerced into’ or ‘conceptualized as’ actions
in such uses. Functional heterogeneity, however, casts doubt on this idea. The
well-wish Get well soon! arguably does not involve any kind of action at all. A
more defensible variant of this view is to construe the denotation of imperatives as
related to events (or more generally eventualities in view of stative imperatives such
as Please, be at home!). For concreteness, let us assume that imperatives denote
functions of type hev, ti, where ev is the type of events.
Again, it would be easy to adjust the Imperative Convention accordingly. Let
Happen be a predicate of events true if the event is instantiated. Then:
(6.62)
Imperative Convention (properties of events)
A speaker Sp who utters an imperative towards addressee Ad with content
16
This could be a single (possibly plural) individual, in case we allow for plural individuals, or a
set of individuals. The denotation below is formulated in terms of the single-individual case, but it
should be obvious how it could be translated to the set-of-individuals case.
CHAPTER 6. COMMITMENTS TO PREFERENCES
168
P thereby incurs the following commitment:
pepSp pDe : Happenpeq ^ Ppeqq
Upshot
The discussion of these examples illustrates the following fact: All that
the theory proposed here requires is that the denotation of imperatives—whatever
it is—can be used to determine a proposition that characterizes the fulfillment conditions of the imperative. As long as this is possible, the Imperative Convention
needs to be adjusted only very slightly to be compatible with any kind of denotation.
In Section 6.4.4 below, I will argue that there is at least one further option
for the denotation of imperatives that leaves the predictions of the current theory
unaffected, according to which the denotation of imperatives does not even directly
determine fulfillment conditions. Before turning to that question, I briefly want
to show that the considerations of the present section apply, mutatis mutandis, to
interrogative sentences, as well.
6.4.2
Denotations for interrogatives
Just as the hypothesized Imperative Convention does not, in any essential way,
depend on a particular hypothesis concerning the denotational type of imperatives,
neither does the Interrogative Convention proposed in Section 6.3. All that the
convention requires is that there is a way to derive, from the denotation, a set of
possible answers to the interrogative. The generalized version of the Interrogative
Convention is stated in (6.63):
(6.63)
a.
Let Ans be a function that takes the denotation of an interrogative
to the corresponding set of propositions characterizing possible answers.17
b.
Interrogative Convention (generalized version))
When a speaker utters an interrogative with content ?ϕ, he thereby
CHAPTER 6. COMMITMENTS TO PREFERENCES
169
incurs the following commitment:
pepSp pDp P Ansp?ϕq : pbAddr ppqq
Of course, the influential theory of interrogatives of Groenendijk and Stokhof (1984)
(building on Hamblin (1958)) assumes that the denotation of interrogatives just is
the set of possible answers. In that case Ans can just be the identity function, and
(6.63b) amounts to the statement in Section 6.3.
But (6.63) is compatible with a large range of alternative hypotheses about
interrogative meaning. As an example, take the Structured Meaning approach
to question meaning, which has a long tradition (Hiz (1978) traces it back to Ajdukiewicz (1938)). Krifka (2001a) summarizes the basic idea as follows:
(6.64)
Question meanings are functions that, when applied to the meaning of
the answer, yield a proposition.
On this approach, answers are not necessarily taken to be propositions, but we still
can take Ans to return a set of propositions. To take Krifka’s example: If Who read
Die Kinder der Finsternis? denotes the function λx.READpKFqpxq, we can simply
apply this function to all the individuals in the domain to derive the set of possible
answers referred to in (6.63).
As in the preceding section, the lesson is that considerations about possible
uses of matrix sentences do not, generally, determine a denotation type, they only
constrain the set of possible types. If we can derive additional constraints on
possible types from investigating the semantics of embedded occurrences (as has
been done in the case of interrogatives in the large body of literature subsequent
to Karttunen (1977)), then these constraints, together with the constraints derived
from pragmatic considerations may—but do not need to—determine a unique
appropriate denotation type.
17
I leave it as an open question for now how ‘possible answer’ should be spelled out, so as to
allow for both mention-all and mention-some readings of questions.
CHAPTER 6. COMMITMENTS TO PREFERENCES
6.4.3
170
The (possible) indeterminacy of denotation types
The present theory is compatible with a large range of hypotheses concerning the
denotational type of imperatives. At the same time, imperatives can, at least in
English, be embedded only to a very limited degree. In particular, there do not
seem to be any verbal or adjectival predicates that embed imperatives.
One may wonder, then: How can we figure out what the denotational type of
imperatives is? The short answer is that we cannot. If a sentence type truly never
occurs embedded in other sentences, then we only have considerations about
possible uses to guide our hypotheses about the denotational type. And such
considerations will never single out one single type. But this is no reason to
despair. It simply means that the question ‘What do imperatives really, truly,
denote?’ is an uninteresting question. To the extent that imperatives do not embed,
the interesting question to ask is how they are used, and what kind of content must
their denotation determine? If we adopt the generalized version of the Imperative
Convention from Section 6.4.1, the answer is: The denotation must determine the
fulfillment conditions of the imperatives. There simply is nothing more to be said
about the denotation of imperatives.
It is even possible, if imperatives do not embed, that there is no fact of the
matter about what the denotation of imperatives is. What this means depends,
of course, on what we take to be the determinants of semantic facts, a difficult
philosophical question I wish not to go into here. But let us assume, for the sake
of the argument, that what determines semantic facts is the linguistic knowledge
of competent speakers of the language. Now, consider the following thought
experiment.18 A speech community C speaks a language L, which has an imperative
sentence type such that imperative sentences never occur as constituents in other
sentences. All speakers in C are perfectly identical in their linguistic knowledge
about L,19 with one exception: There are two subgroups of speakers in C. On the
18
On some level, this thought experiment is just a Quinean indeterminacy argument.
This assumption is of course one of the crucial problems for the thesis that semantic facts
are determined by the linguistic knowledge of the speakers of the language, but grant me the
assumption for the sake of the thought experiment.
19
CHAPTER 6. COMMITMENTS TO PREFERENCES
171
one hand there is the group Cp , which takes imperatives to denote propositions,20
and believes them to be associated with the appropriate convention of use (i.e., the
propositional version we have been using in previous sections). On the other hand,
there is the group Ce which takes imperatives to denote properties of eventualities,
and also believes them to be associated with the corresponding convention of
use. The speakers in Cp and Ce could interact with each other smoothly, using
imperatives to communicate successfully all day long, without ever realizing that
they have incompatible beliefs about the denotational type of imperatives. Nor
could an external observer (a linguist studying L, say) ever determine which of the
individuals belongs to Cp and which belongs to Ce . In such a community, there
simply would be no fact of the matter as to what the denotation type of imperatives
is. There would only be a fact of the matter about how imperatives are used, and
a fact of the matter about what imperative denotations need to be able to do (i.e.,
determine fulfillment conditions). To the extent that imperatives in English truly
do not embed, I see no reason not to believe that the same is true for speakers of
English.
It is not strictly speaking true, or at least not obvious, however, that imperatives
never occur embedded in other sentences: English allows for conditional imperatives, as in (6.65a), which at least could be construed as conditional declaratives with
an imperative consequent. Similarly, Imperatives can occur as the first element in
conjunctive sentences (6.65b) and in disjunctive sentences (6.65c).21
(6.65)
a.
If you get lost, call me!
b.
Mow the lawn and I will give you $5.
c.
Freeze or I’ll shoot.
It is conceivable that investigation of constructions such as (6.65) will yield further
20
When I say that speakers in Cp ‘take imperative to denote propositions’, I refer, of course, to
their implicit grammatical knowledge, not to a semantic theory of imperatives they believe in.
21
All these three constructions have seen considerable attention in the semantics literature, but
a consensus on how they should be analyzed has not yet emerged. On the topic of conditional
imperatives, see Kaufmann and Schwager (2009), and references therein. On conjunctions and
disjunctions of imperatives and declaratives (so-called ‘pseudo-imperatives’), see Kaufmann (2012,
Chapter 6, p. 221–254), and references therein.
CHAPTER 6. COMMITMENTS TO PREFERENCES
172
constraints on what the denotational type of imperatives should be.
6.4.4
One convention for declaratives and imperatives?
As indicated at the end of Section 6.4.1, there is a another possibility for the denotation of imperatives that leaves the predictions of the proposal of Condoravdi and
Lauer (2012) unaffected. This version of the proposal is analogous to the proposal
of Kaufmann (2012). Kaufmann takes imperatives to denote propositions, but not
propositions that characterize the fulfillment conditions of the imperative. Instead,
on her proposal, imperatives are headed by a modal operator Imp much like English
must or should, which is analyzed as a graded modal in Kratzer (1981)’s theory of
modality.22 Schematically, we can characterize her proposal as in (6.66).
(6.66)
vLeave!w “ Imppyou leaveq
Following Kaufmann’s lead, we could assume that imperatives contain an abstract
imperative operator, whose truth conditions are given in terms of pep:
(6.67)
vLeave!w “ pepSp pyou leaveq
And assume that Imperative Convention is (6.68).
(6.68)
Imperative Convention (Kaufmannian version)
When a speaker Sp utters an imperative S with content φ, he thereby
incurs the following commitment:
pbSp pφq
Given the semantics in (6.67), that would mean that a speaker who utters an imperative incurs the commitment pbSp ppepSp ppqq. Given the introspection properties
22
This is motivated, in part, by the fact that modals can be ‘used performatively’ to create
obligations, etc., instead of reporting them (cf. Kamp (1978)). Kaufmann imposes a number of
conventional felicity conditions on imperatives that are intended to ensure that imperatives can
only be ‘used performatively’.
CHAPTER 6. COMMITMENTS TO PREFERENCES
173
in Section 5.3.2, this entails pepSp ppq—that is, (6.68) comes out as equivalent to the
original version of the Imperative Convention.23
Of course, (6.68) is identical to our original Declarative Convention, so if we
adopt (6.67), we can replace the two conventions with a single one. Given that,
under present assumptions, imperatives and declaratives are the only sentence
types that have a propositional denotation, we can do so expediently by having
the denotation reference the denotational type:
(6.69)
Convention about sentences with propositional denotations
When a speaker Sp utters an sentence S with content φ of type hs, ti, he
thereby incurs the following commitment:
pbSp pφq
Deciding between this version of the theory and one on which imperatives semantically characterize fulfillment conditions involves a trade-off in complexity. The
version here is simpler in that it requires only a single convention where the previous ones needed two, but this simplicity is traded against increased complexity
of the denotations of imperatives, as we need to assume an abstract imperative
operator.
6.4.5
The form of the clause-typing conventions
In the statement of the various conventions, I so far have glossed over an issue that
constitutes another way in which the present theory is very flexible: The issue of
clause type identification. The natural language descriptions of the conventions
had the form ‘When a speaker utters a declarative/imperatives/interrogative . . . ’,
but I have not said what I take this to mean.
23
In particular, the analysis would still predict that imperative utterances are ‘automatically
sincere’: Even though imperatives, on this version of the analysis, create doxastic commitments, the
proposition the speaker gets committed to is true as soon the sentence is uttered (cf. the discussion
of explicit performatives in Chapter 7). The same is not true for Kaufmann’s analysis, as we point
out in Condoravdi and Lauer (2012).
CHAPTER 6. COMMITMENTS TO PREFERENCES
174
One way to construe this is as an abbreviation of a morpho-syntactic description.
This corresponds most directly to how sentence types are treated in Sen—each
sentence contains an operator that marks it as being of a given syntactic type,
which (at least in the case of imperatives and declaratives) is uninterpreted. We
can make this construal more explicit:
(6.70)
a.
Declarative Convention (syntactic version)
When a speaker Sp utters a sentence that is morpho-syntactically
declarative and has the content φ, he thereby incurs the following
commitment:
pbSp pφq
b.
Imperative Convention (syntactic version)
When a speaker Sp utters a sentence that is morpho-syntactically
imperative and has the content φ, he thereby incurs the following
commitment:
pepSp pφq
c.
Interrogative Convention (syntactic version)
When a speaker Sp utters a sentence that is morpho-syntactically
interrogative and has the content φ, he thereby incurs the following
commitment:
pepSp pDp P φ : pbAd ppqq
But this is not the only option. If we assign a unique denotation type to each
sentence type, the conventions can just as well be formulated in terms of these
types. Suppose, for concreteness that declaratives are of type hs, ti, imperatives are
of type hev, ti and interrogatives are of type hhs, ti , ti. Then the conventions can be
specified as follows:
(6.71)
a.
Declarative Convention (denotational type version)
When a speaker Sp utters a sentence that has a content φ of type hs, ti,
CHAPTER 6. COMMITMENTS TO PREFERENCES
175
he thereby incurs the following commitment:
pbSp pφq
b.
Imperative Convention (denotational type version)
When a speaker Sp utters a sentence that has a content φ of type hev, ti,
he thereby incurs the following commitment:
pepSp pφq
c.
Interrogative Convention (denotational type version)
When a speaker Sp utters a sentence that has a content φ of type
hhs, ti , ti, he thereby incurs the following commitment:
pepSp pDp P φ : pbAd ppqq
There are other options, but the contrast between these two versions may suffice
to illustrate the point. Which of these is more attractive will depend largely on
two factors: On the one hand, using the denotational type versions requires that
each type of sentence that has its own use also has a unique denotational type. As
such, it is unsuitable if we have independent reason (e.g., from the investigation
of embedded uses) to assume that two sentence types with distinct uses have the
same denotational type. On the other hand, the syntactic version is less attractive
if we have a class of sentences that is morpho-syntactically heterogeneous, but
whose members have a uniform use. In that case, we would be forced to state
multiple conventions for this single class (or to use a cumbersome disjunctive
morpho-syntactic description in the statement of the convention).
Finally, if we have a class of sentences that is morpho-syntactically heterogeneous, but that shares its denotation type with other sentences that do not have the
same use, neither the syntactic, nor the denotation-type versions are attractive. In
such a situation, an attractive solution is to opt for semantic representationalism
CHAPTER 6. COMMITMENTS TO PREFERENCES
176
of the kind advocated in Potts (2005). In Potts’ system, natural language sentences are translated into a typed λ-calculus in an entirely compositional fashion.24
The semantic representation language then allows for expressions with the same
model-theoretic interpretation to have distinct semantic types (in Potts’ case, types
for ‘at-issue’ meanings an ‘conventional implicature’ meanings).
Which version of the theory of clause-typing presented her is most desirable
thus depends on general considerations of parsimony and theoretical elegance, as
many theoretical choices do. But it also depends on what independent constraints
we have on denotational types, and on the morpho-syntactic analysis of sentences
that share a single use. It hence is not a question that can be resolved just in terms
of considerations about use alone.
6.4.6
Conclusion
The purpose of this section was to illustrate that the theory of clause typing proposed here and in Condoravdi and Lauer 2011 and Condoravdi and Lauer 2012 is
compatible with a wide range of assumptions about the denotational semantics of
sentences of various types. Even though the rest of this dissertation makes particular assumptions about denotational types (declaratives and imperatives denote
sets of possible worlds, interrogatives denote sets of those), the overall theory does
not rely on these assumptions in any essential way.
I consider this a significant virtue of the theory. Considerations about possible
uses do not determine a unique denotatum for sentences, they only contribute some
general constraints on what a sentence can denote. Accordingly, the architecture
of a general theory of use, including a theory of conventional constraints on use,
should not require any particular kind of semantic denotatum, but should only
determine some constraints on them. The theory proposed here does just that.
24
That is, unlike in strongly representational theories such as DRT, the semantic representation
maintains the structure of the interpreted phrase structure.
CHAPTER 6. COMMITMENTS TO PREFERENCES
6.5
177
Representing force in the compositional system
In this section, I want to take up the question whether sentential force can and
should be represented within the system of semantic composition. Up to now we
have operated under the assumption that it is not. Matrix sentences have the same
denotation as embedded sentences have, and these denotations capture only the
content of the sentence, not its force.
This is not the only possibility, however. As mentioned in Section 3.3, an
alternative is to assume that matrix sentences have a denotation which encodes
their sentential force. The most straightforward way to implement this is to assume
that matrix sentences contain an unpronounced ‘force operator’ whose semantics
specifies the force of the sentence. On the conception of sentential force proposed
in the previous chapters, this force is consists in essential normative effect of the
sentence.
The most obvious way to implement this is to assume force operators denote
update functions that take the current context of utterance into a new context of
utterance. In order to implement this smoothly, it would be expedient to slightly
adjust our models in the following way: Instead of associating public beliefs and
preferences directly with world/time pairs, we would associate each world/time
pair with suitable representation c of a context, which in turn determines the sets
cpPBi q and cpPEPi q for each discourse participant i, as well as other features of the
content, such as cpSpq for the speaker of the current utterance. Then the semantics
of the force operators would be as follows, where crx{ys is the context that is like c,
except that the component x has the value y:
(6.72)
a.
Declarative operator
v$ w “ λϕ.λc.crPBcpSpq {cpPBcpSpq q Y tϕus
b.
Imperative operator
v!w “ λϕ.λc.crPEPcpSpq {cpPEPcpSpq q Y tϕus
c.
Interrogative operator
!
)
v?w “ λϕ.λc.crPEPcpSpq {cpPEPcpSpq q Y Dp P ϕ : pbcpAdq ppq s
CHAPTER 6. COMMITMENTS TO PREFERENCES
178
This treatment would essentially be that of Krifka (to appear). Krifka’s representation is not completely identical of course, but it is very similar in the essential
aspects: He also uses a forward-branching model of time, and he also assumes that
utterances reduce the set of future possibilities (he calls this an ‘option space’), i.e.,
for him, too, utterances exclude certain future states of affairs. The commitments
that he assumes are different from the ones argued for in the preceding chapters,25
but the basic idea is the same.
Even if we assume that sentences have such dynamic denotations, we still need
to assume a convention of use—again represented as a constraint on admissible
models—that ensures that if a sentence denotes a certain update function, then an
utterance of the sentence applies this function to the current context. Let conpw, tq
refer to the context at w, t and ˚ a metavariable ranging over t$ , ?, !u.26
(6.73)
Utterance effect constraint
A model for PSen is admissible only if for all ˚ϕ, t, a, b:
If Hapw pt, t ` 1q “ utterpa, b, ϕq then
conpw, t ` 1q “ v˚ϕw pconpw, tqq
As discussed in Section 3.3, even though (6.73) is uniform in that it represents a
single convention that governs clauses of all types, it is not at all obvious that
this setup is more parsimonious than one that employs multiple conventions of
use. After all, we now need a convention of use plus one operator per sentence
type. And we had to give up the attractive, simple picture according to which the
denotation of a sentence is the same in its matrix and embedded uses, assuming a
number of silent operators that are otherwise unmotivated.
Krifka discusses a number of ways in which we can exploit a compositional
representation of force in the treatment of various phenomena. Here I will focus
25
In particular, Krifka’s imperative operator directly imposes an obligation on the addressee,
which is quite problematic, given the functional heterogeneity of imperatives discussed above.
26
Krifka defines a second update operation, ` that essentially does the job of this convention of
use—so despite the fact that he has dynamic denotations, he too needs an additional operation that
applies the denotation to the current context when an utterance happens.
CHAPTER 6. COMMITMENTS TO PREFERENCES
179
on the phenomenon that I think has the potential to make the most compelling
argument for representing force in the compositional system: illocutionary modifiers. In Chapter 7, I will discuss another of Krifka’s cases, which involves logical
operators in explicit performatives.
Illocutionary modifiers are items like frankly which intuitively predicate something of the current speech act.
(6.74)
Your tie and shirt frankly don’t go together.
Krifka proposes that frankly combines with the force operator in order to modify it, but frankly can be treated similarly on the assumption that force is extracompositional. We can assume that frankly indexically refers to the current utterance, and predicates something of it (namely that it is done frankly). All that
remains to be explained, then, is why the contribution of frankly does not become
part of the truth conditional content of the sentence—but that could be accounted
for in various ways, e.g., in a multi-dimensional system à la Potts (2005).
The reason this works smoothly in the case of frankly is that its contribution
leaves the conventional dynamic effect of the sentence unaffected: A declarative
utterance with frankly commits the speaker to believe in the truth of the content
of the sentence, just as an utterance of the same sentence without frankly would.
Krifka does not discuss other such modifiers, but it is quite possible that there are
ones that work differently in that they change the basic dynamic effect of the uttered
sentence. If such operators exist, they would constitute a compelling argument for
the representation of force in the compositional system.
A potential example are the ‘illocutionary evidentials’ Faller (2002) describes in
Cuzco Quechua. According to Faller, declaratives containing the hearsay evidential
-si do not create speaker commitments. If this is so, we cannot treat -si in the
way I just suggested for frankly, as -si needs to cancel the conventional effect of
the declarative—putting -si into a sentence prevents the usual commitment from
coming into effect.27
27
Some particles in languages like German, in particular the ones that have been called the
CHAPTER 6. COMMITMENTS TO PREFERENCES
180
For items like evidentials, which typically form a small closed class of morphemes which are often in complementary distribution, another treatment naturally suggests itself. We could see them as marking distinct sentence types—i.e.,
we could assume that Cuzco Quechua does not have one declarative, but several,
each of which is associated with its own extra-compositional convention of use.
On such a construal, there would be no commitment for -si to cancel—instead, -si
would function simply as a marker for the ‘hearsay-declarative’, much like word
order and intonation function as markers for sentence types in English.
It is an open question whether there are expressions that (i) do not affect the
truth-conditional content of sentences, (ii) affect the commitment undertaken with
a sentence and (iii) for which it is implausible to assume that they serve to mark
their own clause type. If there are such expressions, their existence constitutes
a compelling argument for a compositional implementation of clause-typing à la
Krifka. But unless and until we can make such a compelling case, I prefer to keep
the specification of normative effects of utterances out of the denotational semantics
of natural language sentences.
Abtönungspartikel (‘downtoning particles’) might constitute a similar case—for example, Zimmermann (2004) proposes that wohl weakens that degree of commitment that the speaker undertakes
with his utterance.
Chapter 7
Explicit performatives
Explicitly performative utterances (explicit performatives, for short) are utterances of
sentences like those in (7.1a-c). They are of interest because they are a rare instance
where ‘saying makes it so’: If a speaker sincerely says that he promises, he has
thereby promised.
(7.1)
a.
I promise to come to the party.
b.
I order you to sign the contract.
c.
I claim that the king is illegitimate.
Explicit performatives are particularly interesting from the perspective of a theory
of clause-typing, as they apparently are an instance where illocutionary force of an
utterance is made explicit. This is how explicit performatives have been typically
conceptualized since the seminal work of Austin (1962).
One might hence suspect that the theory of clause-typing presented in the previous chapters faces difficulties in accounting for explicit performatives, given that
it intentionally avoids any reference to illocutionary concepts. In fact, the opposite
is the case: The analysis of explicit performatives is entirely straightforward.
181
CHAPTER 7. EXPLICIT PERFORMATIVES
7.1
182
Saying makes it so
To appreciate how surprising it is, initially, that speakers can do something (such as
promising) by saying that they do, it is useful to consider essentially non-linguistic
acts, such as frying an egg.
(7.2)
John: I (#hereby) fry an egg.
By uttering (7.2), John does not bring about the frying of an egg. We would not
report the utterance in (7.2) by saying John fried an egg. Instead, we would report
it, for example, by saying John said he fried an egg.
Things are different with explicit performatives. If John says I promise to come
to the party, we can report his utterance by saying John promised to come to the
party (as well as the more prolix John said he promises to come to the party).
This is a general property of explicit performatives: We can report utterances of
the (a)-sentences in (7.3)–(7.5) by using the corresponding (b)-sentence.
(7.3)
(7.4)
(7.5)
a.
I promise to come to the party.
b.
He promised to come to the party.
a.
I order you to sign the contract.
b.
He ordered me to sign the contract.
a.
I claim that the king is illegitimate.
b.
He claimed that the king is illegitimate.
There is, of course, an obvious difference between fry an egg and promise to p: The
latter refers to an action that can be performed by a linguistic utterance, whereas the
former cannot. But this cannot be the only difference between the two predicates
that is responsible for the the fact that, with performatives, saying makes it so
There are many predicates that refer to communicative actions which cannot be
‘used performatively’—i.e., which cannot be used to perform the action merely by
saying that one does.
CHAPTER 7. EXPLICIT PERFORMATIVES
(7.6)
a.
I (hereby) insult you.
b.
I (hereby) annoy you.
c.
I (hereby) intrigue you.
d.
I (hereby) frighten you.
183
Insulting, annoying, intriguing and frightening are certainly ‘things that can be
done with words’ (to use Austin’s phrase). And yet, these are not things a speaker
can do merely by saying that he does.1 Explicit performatives, then, raise two
questions:
1. How can, with certain verbs, saying that one does something constitute doing
that thing?
2. What is the difference between verbs that can be used in this way and verbs that
cannot?
The traditional way of approaching this question answers the second question
first, and then uses this answer to tackle the first: A class of illocutionary acts
is characterized independently, and distinguished from another class of acts that
can be performed by making an utterance but are not illocutionary acts (these are
called, following Austin, the perlocutionary acts). It is then hypothesized that
verbs like promise, order and claim that can be used in explicit performatives refer
to illocutionary acts, while verbs that cannot be so used, like insult, annoy and
frighten instead refer to perlocutionary acts. This hypothesis is then used in trying
to explain why illocutionary verbs can be used in explicit performatives.
The strategy here will be to go the other way round: I will specify suitable
meanings for some of the ‘illocutionary’ verbs, and show how it is that they can
be used in explicit performatives. Then I will step back and ask, given the nature
of this explanation, what a verb has to be like in order to fit this use, in effect
1
With the exception, perhaps, of certain, very special contexts—e.g., if it is clear that any further
word out of John’s mouth would annoy Mary, then he can indeed annoy her by saying I annoy
you—but he also could say anything else. In such a special context, there is nothing about the
content of the uttered sentence that makes it achieve what it does.
CHAPTER 7. EXPLICIT PERFORMATIVES
184
reconstructing the distinction between ‘illocutionary verbs’ and ‘perlocutionary
verbs’ in terms of the properties necessary for explicit performatives.
7.2
Explicit performatives as ‘self-verifying’
Approaches to explicit performatives come in two broad classes, and while I defer
comparison with selected existing approaches to Section 7.6, I want to briefly situate
the analysis that flows naturally from the system of clause-typing developed in the
preceding chapters with respect to these classes.
The distinction between the two classes is perhaps most easily appreciated by
considering the characterization of explicit performatives offered by Heal (1974):
“Roughly speaking, an explicit performative utterance occurs when
(i) a sentence is uttered and an action is thereby performed, and
(ii) the grammatical form of the sentence makes it look at first glance
as though the speaker states that he performs that action.”
(Heal 1974, p. 106)
The two classes of accounts that I want to distinguish differ on whether they take
the impression described in (ii) to be misleading or not.2
Austin (1962) firmly is in the former camp—he takes it to be obvious that,
promising by saying I (hereby) promise to p does not involve stating that one is
promising:3
“In these examples it seems clear that to utter the sentence (in, of course,
the appropriate circumstances) is not to describe my doing of what I
should be said in so uttering to be doing or to state that I am doing it:
2
Heal herself argues for the position adopted here—that the impression is not mistaken. She
does not spell out how (i) is to be explained, though.
3
Technically, these remarks by Austin pertain to examples like I name this ship the Queen
Elizabeth—but Austin held that these are equivalent, in the relevant respects, to acts of promsing,
etc.
CHAPTER 7. EXPLICIT PERFORMATIVES
185
it is to do it. None of the utterances cited is either true or false: I assert
this as obvious and do not argue it. It needs argument no more than
that ‘damn’ is not true or false[.]”
(Austin 1962, p. 6)
However, in the years since Austin said this, a large number of authors have
held the opposite—that, just as it appears, when a speaker utters an explicit performative, he states that he is doing something and thereby does it (Lemmon 1962,
Hedenius 1963, Heal 1974, Bach and Harnish 1979, Ginet 1979, Bierwisch 1980,
Leech 1983, a. o.).
Though these approaches differ in the details, the idea is to analyze explicit
performatives on the model of sentences like (7.7).
(7.7)
I am speaking now.
On standard assumptions, (7.7) expresses the proposition that the speaker is speaking at the time of utterance. As a consequence, whenever the sentence is uttered, the
proposition it expresses is thereby made true. The idea is to assume that explicitly
performative utterances are ‘self-verifying’ in just this way.4
Such approaches can be called ‘assertoric’, since they construe the utterance of
explicit performatives as being an assertion which makes its content true. If such
an approach can be made to work, it is intuitively appealing and parsimonious:
The special ‘performative force’ of explicit performatives comes about only in
virtue of the fact that utterances of the sentence happen to make the sentence
true. And given that explicit performatives look and behave syntactically like
other declarative sentences in every way, there is no grammatical basis for the
4
To say that explicitly performative utterances are self-verifying is, of course, to speak somewhat sloppily—or at least involves some unstated assumptions. And utterance can only be selfverifying—make itself true—if utterances can be true or false, which is a non-trivial assumption.
We do not need to assume that utterances (rather than uttered sentences) have truth-conditions in
order to make sense of such talk: We can also just take talk of the truth or falsity of a (declarative)
utterance to metonymically refer to the truth or falsity of the uttered sentence. In any case, when,
here and in the following, I speak of explicit performatives being self-verifying, I mean this in the
sense of Lemmon (1962)’s title—they involve ‘sentences that are verifiable by their use’.
CHAPTER 7. EXPLICIT PERFORMATIVES
186
once popular claim (initiated by Ross (1970)) that in explicit promises, I promise
spells out an illocutionary operator that is silent in sentences that are not explicit
performatives. Such an analysis, in any case, leaves unanswered the question of
how force is related to compositional meaning and, consequently, does not explain
how the first person and present tense are special, so that first-person present tense
forms can spell out performative prefixes, while others cannot. Minimal variations
in person or tense remove the ‘performative effect’:
(7.8)
I promised you to be there at five.
(is not a promise)
(7.9)
He promises to be there at five.
(is not a promise)
Explicit performatives look like other declarative sentences, so the most straightforward assumption is that they are like all other declaratives in that they have
denotations that can be true and false; and that they are used in just the way
declarative sentences are normally used, i.e., to make statements or claims. On the
analysis advocated in this dissertation, this means that explicit performatives, like
all other declaratives, commit their speaker to their compositionally determined
content.
We can give precise formal content to what we want to explain. The idea is that
an explicitly performative utterance is automatically true once it is uttered, that
is, we are looking for a semantic analysis of the verbs involved that ensures the
following:
(7.10)
(7.11)
(7.12)
If a speaker utters I promise φ in context c and world w, then w P
0
8c
I promise φ .
0
8c
If a speaker utters I order φ in context c and world w, then w P I order φ .
0
8c
If a speaker utters I claim φ in context c and world w, then w P I claim φ .
CHAPTER 7. EXPLICIT PERFORMATIVES
7.3
187
Truth conditions for performative verbs
If we assume that explicit performatives are just utterances of regular declaratives,
and that all their constituents have the same meaning they would have in other
sentences, a good strategy is to first investigate reportative uses of performative
verbs, in order to arrive at a hypothesis for their meaning. On the one hand, we
ultimately require an analysis of these verbs that is adequate for both performative
and reportative uses, as Szabolcsi (1982) stressed:
“Interestingly enough, researchers who find the felicity conditions of ‘I
promise that. . . ’ or ‘I congratulate you’ so complex and necessary to
study closer never devote much attention to the “special truth conditions” of ‘Peter is congratulating Mary’ or ‘Yesterday Peter promised
that. . . ’, although it is inevitable that in deciding whether Peter is in the
set of those who are congratulating Mary etc. one should consider the
very same questions.”
(Szabolcsi 1982, p. 530)
On the other hand, while native speakers have intuitions about the truth conditions
of reportative uses, it is unlikely that they have reliable intuitions about performative uses. The reason for this is that, by assumption, performative sentences are
automatically true once uttered, and so speakers are unlikely to have intuitions
about conditions under which they would be false.
Let us consider, then, the question what has to be the case for (7.13) to be true.
(7.13)
John promised to be at the airport at noon.
(7.13) is true if there was an utterance5 by John, and with that utterance, John
incurred a certain commitment. What kind of commitment? To comply with a
promise to p is to choose one’s actions so as to make p true—that is, to choose
5
Or some other communicative act. For simplicity, I will talk about ‘utterances’.
CHAPTER 7. EXPLICIT PERFORMATIVES
188
one’s actions as if one effectively prefers that p. So (7.13) says that John committed
himself to pep John pJohn be at the airport at noonq.6
What about (7.14)?
(7.14)
John ordered Mary to be at the airport at noon.
Again, to comply with an order (or a request, or a similar directive) is to act as
though one preferred what has been ordered. One need not actually prefer it, but
one must choose one’s actions as if one did.
So we can understand (7.15) as saying that there was an utterance by John
(to Mary), and that utterance publicized John’s (effective) preference for Mary to
act as though she effectively prefers to be at the airport at noon. In terms of our
formalization of these attitudes, the commitment resulting from John’s utterance is
pep John pepMary pMary be at the airport at noonqq.7
Finally, (7.15) is intuitively true if John made an utterance that committed him
to a belief that the king is illegitimate:
(7.15)
John claimed that the king is illegitimate.
I summarize the hypothesized semantics of our three performative predicates in
(7.16).
6
Promise has some additional presuppositions. Firstly, (7.13) presupposes that John took it for
granted, at the time of the promising, that his addressee had a stake in whether or not he will be
at the airport at noon. When we think of promises, we usually think of cases where the promiser
takes the addressee to desire what is promised, but a sentence such as I promise that p can also be
used to threaten—i.e., to commit to acting in a way that the promisee disprefers. Crucially, such
utterances can still be reported using the verb promise: He promised to make my life miserable.
Secondly, promise likely presupposes that the promiser believes that he is able to ensure the truth
of what is promised.
7
Order again comes with additional presuppositions. One of these is that the orderer presumes
to have authority over the addressee with respect to order, in the sense explicated in (6.22) in
Section 6.1.2. Nota bene: order presupposes only that the speaker presumed to have authority, not
that he actually did, as attested by the coherence of (i).
(i)
John ordered Mary to be at the airport at noon, but he did not have the authority to do so.
CHAPTER 7. EXPLICIT PERFORMATIVES
(7.16)
a.
189
John claimed that p is true iff John made an utterance resulting in
pb John ppq.
b.
John promised to p is true iff John made an utterance resulting in
pep John ppq.
c.
John ordered Mary to p is true iff John made an utterance resulting
in pep John pepMary ppqq
What all three of these have in common, of course, is that the meaning of the
performative predicates all are spelled out in terms of the normative consequences
of utterances. This is, of course no accident: As we will see in the following, it is
precisely this property that enables claim, promise and order to work in explicit
performatives.
7.4
Deriving self-verification
With the hypothesized meanings for performative predicates, we have in place everything that is necessary to derive the ‘performative effect’: First-person, present
tense sentences with these predicates will be self-verifying. Informally, the reasoning goes as follows. Suppose John utters (7.17), and call this utterance u˚ .
(7.17)
I promise to be at the party.
The denotation of (7.17), in this context, can be paraphrased as in (7.18):
(7.18)
There is an utterance event u by John that commits him to
pep John pJohn be at the partyq.
(7.17) is a declarative, so by the declarative convention, John’s utterance commits
him to the belief that the content obtains:
(7.19)
As a result of u˚ , John is committed to believe that there is an utterance u
that commits him to pep John pJohn be at the partyq.
CHAPTER 7. EXPLICIT PERFORMATIVES
190
But this, in turn, means that (if commitment to belief is closed under logical consequence):
(7.20)
As a result of u˚ , John is committed to believe that he is committed to
pep John pJohn be at the partyq.
But being committed to believing one is committed to a preference means that one
is also committed to the preference.8 And so:
(7.21)
As a result of u˚ , John is committed to pep John pJohn be at the partyq.
But now, reconsider the content of John’s utterance.
(7.22)
There is an utterance event u by John that commits him to
pep John pJohn be at the partyq.
=(7.18)
u˚ , the utterance of I promise to be at the party itself, serves as a witness for
the existential claim made by the sentence. But that means that the sentence is
self-verifying in the appropriate sense.
Given the formal set-up in Chapter 5, we need not leave things on this informal
level: If we extend the object language of our dynamic pragmatics with predicates
like promise, we can prove that these are self-verifying.
7.4.1
Introducing illocutionary verbs into Sen
We enrich Sen with a set of operators that model performative verbs:
(7.23)
For a given PSen -model, Perf is the smallest language that contains Sen
and all sentences of the form Ppa, ϕq, for Ipaq P Ag, ϕ P Sen and P P
tclaim, promise, orderu, as well as propositional combinations of these.
8
With Condoravdi and Lauer (2011), I take this to be an obviously plausible desideratum for any
account of commitments to beliefs and preferences. Formally, this is captured in the introspection
properties required in Section 5.3.2.
CHAPTER 7. EXPLICIT PERFORMATIVES
191
For the sake of simplicity, I introduce claim, promise and order as sentence-forming
operators. Ultimately, we would want to treat them as predicates in a first-order
language with quantification over individuals (and times, which we ignore here),
but for present purposes, treating them as sentence-forming operators is sufficient.
The denotation of Sen-sentences is as before. The new formulas of Perf are
interpreted as follows:9
(7.24)
a.
b.
c.
ˇ
(
vclaimpa, ϕqw “ w P W ˇ w Du : utteru pa, ϕq ^ Resultu ppba pϕqq
ˇ
(
vpromisepa, ϕqw “ w P W ˇ w Du : utteru pa, ϕq ^ Resultu ppep pϕqq
a
vorderpa, b, ϕqw “
ˇ
(
w P W ˇ w Du : utteru pa, b, ϕq ^ Resultu ppepa pepb pϕqqq
As in the informal paraphrases, the denotation of claim, promise and order predicate
the existence of an event token (i.e., that an utterance takes place), which has certain
normative consequences.
7.4.2
Performatives with promise and order are self-verifying
We can now represent a sentence like (7.25) as in (7.26) (assuming that a is the
speaker of (7.25)):
(7.25)
I promise ϕ.
(7.26)
promisepa, ϕq
And we can show that (7.26) is indeed self-verifying, i.e., that:10
Fact 1. If w uttert pu, a, ppromisepa, ϕqqq then w promisepa, ϕq.
9
Note that Perf is untensed, and that formulas like claimpa, pϕqq are hence timelessly true or false,
just like the propositional atoms of the underlying propositional language. To reduce notational
clutter, I omit the temporal parameters of pb and pep throughout. In definition in (7.24), the
appropriate temporal parameter would obviously be the first component of the interpretation of u.
10
Here and in the following, I occasionally enclose Perf -formulas in p¨q when they occur as
arguments to predicates, to make the formulas easier to parse.
CHAPTER 7. EXPLICIT PERFORMATIVES
(7.27)
a.
w utterpu, a, ppromisepa, ϕqqq
b.
w Resultu ppba ppromisepa, ϕqqq
c.
promisepa, ϕq Du1 : Resultu1 ppepa pϕqq
d.
Du : Resultu1 ppepa pϕqq pepa ϕ
e.
promisepa, ϕq pepa ϕ
f.
w Resultu ppba ppepa pϕqqq
g.
pba ppepa pϕqq pepa pϕq
h.
w Resultu ppepa pϕqq
i.
w utterpu, aq
j.
w utterpu, aq ^ Resultu ppepa pϕqq
k.
w promisepa, ϕq
192
(assumption)
(a) + Decl. Convention
(def promise)
1
(Result factive)
(b) and (c)
(Result closed)
(pbppepq introspection)
(Result closed)
(a)
(i) + (h)
(j)+ def promise
Step (7.27g) is the crucial step where we exploit the closure property of (5.32c) of
Section 5.3.2, corresponding to the principle Doxastic reduction for preference
commitment in Condoravdi and Lauer (2011, p. 156). We immediately obtain the
same result for order:
Fact 2. If w uttert pu, a, b, porderpa, b, ϕ1 qqq then w orderpa, b, ϕ1 q.
The proof proceeds exactly as in (7.27), with epb pϕ1 q in place of ϕ.
7.4.3
Performatives with claim are self-verifying
The proof is analogous for claim, but I spell it out in full, seeing as it relies on
a different introspection principle (viz., Condoravdi and Lauer (2011)’s Positive
introspection for doxastic commitment, captured here in (5.34a) in Section 5.3.2).
We again can represent (7.28) as (7.29) (assuming a is the speaker of (7.28)):
(7.28)
I claim that ϕ.
(7.29)
claimpa, ϕq
And then show:
Fact 3. If w utteru pa, pclaimpa, ϕqqq then w claimpa, ϕq.
CHAPTER 7. EXPLICIT PERFORMATIVES
(7.30)
7.4.4
a.
w utterpu, a, pclaimpa, ϕqqq
b.
w Resultu ppba pclaimpa, ϕqqq
c.
claimpa, ϕq Du1 : Resultu1 ppba pϕqq
d.
Du : Resultu1 ppba pϕqq pba ϕ
e.
claimpa, ϕq pba ϕ
f.
w Resultu ppba ppba pϕqqq
g.
pba ppba pϕqq pba pϕq
h.
w Resultu ppba pϕqq
i.
w utterpu, aq
j.
w utterpu, aq ^ Resultu ppba pϕqq
k.
w claimpa, ϕq
193
(assumption)
(a) + Decl. Convention
1
(def claim)
(Result factive)
(b) and (c)
(Result closed)
(pb introspection)
(Result closed)
(a)
(i) + (h)
(j)+ def claim
Summary
In foregoing sections, I showed that, given an appropriate specification of the
truth-conditional content of performative verbs, the self-verification of explicit
performatives immediately results from the system set up in Chapter 5. Even
though spelling out the proof is somewhat involved, the treatment of performatives
is ultimately extremely simple: Performative verbs have regular, truth-conditional
meanings, which happen to be of such a kind that when a performative sentence
is uttered, this utterance itself ensures the truth of the sentence.
It is important to stress that the steps laid out in (7.27) and (7.30) are not intended
as a representation of pragmatic reasoning the hearer goes through when observing
an utterance. They are simply proofs that show that self-verification is ensured.
Of course, this also means that when a hearer observes an utterance of an explicit
performative, he thereby also comes to believe that the uttered sentence is true, i.e.,
that the promise, order, claim, etc. indeed happened.
CHAPTER 7. EXPLICIT PERFORMATIVES
7.5
194
Further predictions
In this section, I draw out some further predictions of the account of explicit
performatives presented here. In Section 7.5.1, I briefly discuss how self-verification
fails when performative verbs are used reportatively. In Section 7.5.2, I spell out
some predictions of the account with respect to combination of verbs with logical
operators, showing that the observations that Krifka (to appear) explains in terms of
illocutionary operators follow straightforwardly without assuming those. Finally,
in Section 7.5.3, I return to the question why verbs like insult and annoy cannot be
used in explicit performatives.
7.5.1
Reportative uses of performative predicates
I have shown that, with first person subjects, sentences involving performative
verbs like promise are self-verifying. Of course, we want to also assure ourselves
that with non-first person subjects, self-verification does not obtain. That is, a
sentence like (7.31) should not be self-verifying:
(7.31)
Mary: John promises to be at the party.
This is indeed the case. It is possible that w utteru pa, ppromisepb, ϕqqq, but w 2
promisepb, ϕq, because a’s utterance event u cannot serve as a witness for the truth
of the sentence she asserts, for two reasons. Firstly, promisepb, ϕq asserts that there
is an utterance event by b, while u is an utterance by a. Secondly, the proof of
self-verification fails at step (7.27g):
(7.32)
pba ppepb pϕqq 2 pepb pϕq
There is no introspection property that ensures that pba ppepb pϕqq entails pepb pϕq,
and there should not be one: Just because I am committed to believe you are
committed to prefer something does not mean that you are so committed.
An analogous argument can be made regarding tense, once we extend our object
language to represent it. Tense would put restrictions on the temporal location of
CHAPTER 7. EXPLICIT PERFORMATIVES
195
the utterance event, and non-present tense would again ensure that the utterance
cannot serve as a witness for the truth of the asserted sentence, explaining why
(7.33) cannot be used performatively.
(7.33)
I promised that I will be at the party.
In general, for self-verification to obtain, the utterance of a sentence must be suitable
for serving as a witness for the existential claim the sentence makes. If it does not,
the sentence can only be used reportatively.
7.5.2
Performatives and logical operators
Krifka (to appear) makes a number of observations concerning the interaction of
logical operators with explicit performatives, which he proposes to explain by hypothesizing non-Boolean meanings for these operators which apply to illocutionary
operators that are part of the compositionally-determined denotation of sentences
(cf. Section 6.5). In this section, I show that the account of explicit performatives
proposed above and in Condoravdi and Lauer (2011) directly predicts these observation on the assumption that the logical operators have their traditional Boolean
denotations.
Explicit performatives and negation
The first operator Krifka discusses is negation:
(7.34)
I do not promise to come.
Krifka characterizes such sentences as follows:
“Denegation can be understood as explicitly refraining from performing
a speech act, like an act of promise (cf. Hare 1970). As such, they are
not regular speech acts. This is reflected by the fact that we cannot use
hereby, different from usual explicit performatives.”
(Krifka to appear)
CHAPTER 7. EXPLICIT PERFORMATIVES
196
Krifka then proceeds to propose that not in (7.34) does not denote regular Boolean
negation, but instead is an operator that applies to an illocutionary operator to yield
a new one. But is unclear why that would be necessary: It seems we can explain
everything that there is to explain about (7.34), maintaining the assumption that
not is regular, Boolean negation. On the analysis given here, (7.34) commits the
speaker to the belief that he does not promise. This seems sufficient to capture the
impression that its speaker is ‘explicitly refraining from promising’: Just like saying
nothing or making an alternative, unrelated utterance, (7.34) does not commit the
speaker to act as though he prefers to come, but it does so in a particular way: By
saying what he does, the speaker is addressing the issue of whether he is promising,
thus either making this issue salient or responding to the pre-existing salience of
this issue. Is there anything about (7.34) that cannot be explained by the fact that
its speaker avoids committing himself to come, and makes the fact that he does so
salient and explicit?11
Krifka appears to think there is something more to be explained. His analysis
predicts that an utterance of (7.34) precludes any future promise to come. But this
predictions strikes me as far too strong. With (7.34), the speaker does not take on a
commitment that he will not make such a promise in the future, he simply explicitly
withholds this promise in the present. This is the difference between (7.34) and
(7.35).
(7.35)
I will not promise to come.
Krifka does not seem to share this intuition of a difference between the two sentences (“We can execute the same change of the option space by expressing a future
assertion, I will not promise to come.”), but it seems plain to me that with (7.35),
the speaker commits himself to not promising in the future while with (7.34), he
11
Hare (1970), which Krifka cites, does not provide any arguments for why (7.34) should not be
viewed as simply the assertion of a negated proposition, though he likely would endorse such a
view. All that Hare (1970) argues is that even if we regard I promise as a ‘performative prefix’ (a
term that he does not use, of course), we can deal with such negated examples by assuming that
there is such a thing as illocutionary negation. But that is not an argument for the existence of such
a negation. It appears that illocutionary negation solves a problem that does not exist in the first
place unless we assume illocutionary operators in the semantic representation.
CHAPTER 7. EXPLICIT PERFORMATIVES
197
does not. I leave this issue for future research.
Explicit performatives with conjunctions and disjunctions
Krifka points out the following contrast:
(7.36)
a.
I hereby promise you to sing, and I hereby promise you to dance.
equivalent to: I hereby promise you to sing and to dance.
b. #I hereby promise you to sing, or I hereby promise you to dance.
The conjunction of two explicit performatives is again a well-formed explicit performatives (and is self-verifying), while the disjunction of two explicit performatives
is odd.
Krifka’s explanation for this is complex: He maintains that and in (7.36a) is not
Boolean conjunction, but a different kind of operator, ‘dynamic conjunction’, which
conjoins speech acts. The infelicity of (7.36b) is explained by the fact that or only
has a Boolean reading, and that Boolean or cannot apply to speech acts.
But it seems we can account for the contrast in (7.36) much more simply. Indeed,
the contrast follows directly from the analysis of explicit performatives proposed
here.
Conjoined explicit performatives are self-verifying.
(7.36a) commits the speaker
to the existence of an utterance that commits him to pepSp psingq and, at the same
time, commits him to the existence of an utterance that commits him to pepSp pdanceq.
But being committed to believe in two such commitments again means having both
commitments. But that means that the very utterance of (7.36) can serve as witness
to both existential claims.
Disjoined explicit performatives are not self-verifying
(7.36b), on the other
hand, does not commit the speaker to a preference for either singing or dancing. It
does commit the speaker to the belief that he is either committed to sing or he is
committed to dance. But this doxastic commitment does not entail a commitment
CHAPTER 7. EXPLICIT PERFORMATIVES
198
for singing, and it does not entail a commitment for dancing. But then, the utterance
in (7.36) cannot serve as a witness for either of the disjoined claims.
Krifka’s analysis of explicit performatives
Krifka’s observations directly follow from the analysis of explicit performatives
proposed here. Interestingly, Krifka’s analysis of explicit performatives is quite
similar.12
Recall from Section 6.5) that Krifka assumes that all sentences are headed by
illocutionary operators. However, he does not take explicit performatives like I
promise to come to be headed by a promise-operator. Instead, the promise operator
occurs within the argument to the usual declarative operator Assert. That is, for
Krifka, the structure of an explicit promise is not (7.37), but (7.38).
(7.37)
Promise(I come).
(7.38)
Assert(Promise(I come))
The inner promise operator in (7.38) is used descriptively, i.e., (7.38) is an assertion
that a promise happens. This already sounds very familiar, especially in light
of the fact that Krifka, too, takes assertion to be defined as the taking on of a
commitment. He does not spell out formally how the performative effect arises,
but he summarizes his conception as follows:
“Now, this proposition states that at this index i [. . . ] the speaker has
the promissive obligations to come, and that the speaker has just at
this point obtained them. The only way how the speaker can heed the
assertive obligation is that the minimal change that created the assertive
obligation also created the [promissive] obligation. This is exactly how
explicit performatives work: The assertion coincides with the promise.”
12
Condoravdi and Lauer (2011), which first presented the analysis presented here, was written
before Krifka (to appear) was available, and it likewise seems that the latter was written before the
former was available. So the two analyses were developed independently, but converged on a very
similar conception.
CHAPTER 7. EXPLICIT PERFORMATIVES
199
(Krifka to appear)
This is not quite the conception proposed here—in particular, it is not clear how
the step “The only way how the speaker can heed the assertive obligation is that
the minimal change that created the assertive obligation also created the [promissive] obligation.” compares to the self-verification as it is derived in the present
approach.13 But it is similar enough to suspect that the account of the interaction
of logical operators with explicit performatives made above apply also to Krifka’s
conception. If so, then even assuming Krifka’s illocutionary operators does not
require assuming illocutionary negation or a speech act conjoining version of and
to account for explicit performatives.
7.5.3
‘Illocutionary’ vs. ‘perlocutionary’ verbs
The preceding sections show how, given appropriate assumptions about the lexical
meanings involved, sentences with promise and order can be self-verifying, and
how this self-verification fails if the predicates are embedded under negation or
disjunction (but not under conjunction), or when they are ‘used reportatively’. But
I have not yet addressed the second question that I initially posed: What is the
difference between verbs that can be used in this way and verbs that cannot?
We can now answer this question, given the way the self-verification of explicit
performatives is explained in the current account. Exactly those predicates can be
used in explicit performatives which . . .
(i) predicate the existence of a communicative event, and
(ii) the only things they require of this event can be characterized in terms of
a. the resulting commitments of the speaker, and
b. possibly additional (presuppositional) constraints on the speaker’s attitudes.
13
This is unclear largely because Krifka leaves it underspecified what it is to have ‘assertive
commitments’—he talks about ‘being liable for the truth of p’ and ‘the commitment to guarantee
that the content of the assertion is true’, but he does not spell out what exactly this means
CHAPTER 7. EXPLICIT PERFORMATIVES
200
It is hence clear why, on the present account, verbs like annoy and insult
cannot be used in explicit performatives. Their meanings do not only involve
commitments made by the speaker, but reactions of the audience. For (7.39) to be
true, John has to have done something (possibly an utterance), and further, John’s
action must have provoked a feeling of annoyance in Mary.
(7.39)
John annoyed Mary.
But then, (7.40) cannot be self-verifying (at least in the way explicit performatives
are self-verifying), for the mere fact that (7.40) was uttered does not guarantee that
the addressee is annoyed.
(7.40)
I (hereby) annoy you.
Similar things can be said about insult, which arguably entails that the addressee
of the insult takes offense.14 .
If we want to, we can now define two classes, the ‘illocutionary verbs’ and ‘perlocutionary verbs’. Illocutionary verbs are verbs of saying whose truth-conditional
content only involves speaker-commitments; perlocutionary ones are those whose
14
Actually, that is not quite true. There are uses of insult which deny an effect on the hearer (I
am grateful to Paul Kiparsky for supplying this example):
(i)
He insulted the audience in Urdu, but they had no idea what he was saying.
Somewhat mysteriously, such uses appear to require that there is a third party which regards the
communicative act as offensive for the the addressee. In any case, these uses are generally not
licensed in virtue of attitudes of the speaker alone, which goes some way to explain the nonperformativity of insult. If not knowing general relativity is not taken to be a sign of a negative
quality by anyone but John, one cannot say (ii), but has to say (iii):
(ii)
John insulted them by explaining general relativity(, which only specialists understand
anyway).
(iii)
John tried to insult them by explaining general relativity(, which only specialists understand
anyway).
This contrast suggests that there is something besides the speaker’s intention to offend that is
necessary for an action to count as an insult.
CHAPTER 7. EXPLICIT PERFORMATIVES
201
truth-conditional content involves further effects on the audience.15 And then, of
course, we can go one step further and call acts named by illocutionary verbs the
‘illocutionary acts’ and verbs named by perlocutionary verbs the ‘perlocutionary
acts’. The distinction between these two classes of acts, however, still would not
figure in any substantive way in our theory of language use, over and above the fact
that illocutionary acts are those that can be performed by explicit performatives,
while perlocutionary acts cannot.16
7.6
Comparison to existing approaches
Searle (1989) opens with the observation that “[t]he notion of a[n explicit] performative is one that philosophers and linguists are so comfortable with that one gets
the impression that som[e]body must have a satisfactory theory” (p. 535). As
might be expected, he then goes on to explain why he thinks existing theories are
not satisfactory. In doing so, he mounts a challenge for theories of explicit performatives, esp. those of the assertoric type. To my knowledge, this challenge has
not been answered by anyone subscribing to a Searlean conception of speech acts
(Although Bach and Harnish (1992) reply to the challenge, their conception of both
speech acts and explicit performatives is quite different from Searle’s, as discussed
below). I hence confine myself here to briefly summarizing Searle’s argument and
his way of dealing with the problem it raises, and to a comparison with Bach &
Harnish’s approach, mentioning a few variant approaches along the way.
15
Actually, not all verbs that predicate ‘perlocutionary effects’ are verbs of saying, or even verbs
of communication—annoy is not, for example, and insult may well not be.
16
The class of illocutionary acts would not correspond to the class of illocutionary acts according
to Searle or Bach and Harnish. These authors make clear that the existence of an illocutionary verb
is not necessary for something to count as an illocutionary act—but perhaps we could approximate
their conception by calling illocutionary acts all those acts that could be named by a hypothetical
illocutionary verb in the sense defined above. But that would not make the class of illocutionary
acts any more interesting from the present perspective.
CHAPTER 7. EXPLICIT PERFORMATIVES
7.6.1
202
Searle (1989): Explicit performatives as declarations
Searle outlines a set of desiderata for a theory of explicit performatives. Central for
our present concerns are the following:
(a) performative utterances are performances of the act named by the performative
verb;
(b) performative utterances are self-guaranteeing;
(c) performative utterances achieve (a) and (b) in virtue of their literal meaning,
which, in turn, ought to be based on a uniform lexical meaning of the verb
across performative and reportative uses.
It should be obvious that the account proposed in this chapter meets all three of
these requirements. Yet Searle claims that an assertoric account cannot meet all
three desiderata. Why?
According Searle, making a promise requires that the promiser intend to do so,
and similarly for other performative verbs (the sincerity condition). It follows that
no assertoric account can meet (a-c): An assertion cannot ensure that the speaker
truly has the necessary intention. As Searle puts it:
“Such an assertion does indeed commit the speaker to the existence
of the intention, but the commitment to having the intention doesn’t
guarantee the actual presence of the intention.”
(Searle 1989, p. 546)
This is obviously true and I think Searle’s argument is a death-blow for assertoric
accounts of explicit performatives that take a Searlean view of speech acts. But I think
we should not take such a view. What Searle’s argument shows, to my mind, is
that the conditions for promising, etc. should not depend on things like private
intentions of the speaker. Instead, what counts for such acts are public facts, such
as whether the speaker is, or is not, publicly committed to a belief or preference.
CHAPTER 7. EXPLICIT PERFORMATIVES
203
The account of explicit performatives presented here takes promising to be
constituted by the creation of a public commitment, nothing more and nothing
less. And this is why self-verification obtains: If a speaker says, without any
indication of insincerity, I promise p, he has thereby promised, and his private
intentions do not matter one bit. He cannot later claim that when he made his
utterance, he did not promise because he failed to have the right intentions. His
utterance committed him. Whether he desired or intended to be so committed by
his utterance is immaterial.
Searle, however, thinks that intentions are necessary. Instead, he advocates
giving up the idea the explicit performatives are assertions, and proposes to view
them as declarations, similar to utterances such as The meetings is adjourned. Such
declarations work because there is an extra-linguistic institution that ensures that
if the right speaker says that the meeting is adjourned, it thereby is.
Promises, orders, and the like are different from these cases only in that the
fact that is created is a linguistic fact (namely that an order, promise, etc. happened). Hence, they do not depend on an extra-linguistic institution, but only on
the institution of language itself.
So far so good, but what kinds of linguistic conventions make it the case that
I can declare with I promise p, but not with I insult you? And further, what
makes it the case that I promise p can only be a declaration (for recall, Searle wants
to explain, as I do, why explicit performatives are self-verifying—they cannot be
mistaken, they cannot be lies)? This is where things get a bit murky. Searle says
we need three ingredients:
“First, we need to recognize that there is a class of actions where the
manifestation of the intention to perform the action, in an appropriate
context, is sufficient for the performance of the action.
Second, we need to recognize the existence of a class of verbs which
contain the notion of intention as part of their meaning. To say that a
person performed the act named by the verb implies that he or she did
it intentionally, [. . . ]
CHAPTER 7. EXPLICIT PERFORMATIVES
204
Third, we need to recognize the existence of a class of literal utterances
which are self referential in a special way, they are not only about themselves, but they also operate on themselves. They are both self-referential
and executive.”
(Searle 1989, p. 551)
The first two of these are reasonably clear. It is not exactly obvious what it is to
‘manifest’ an intention, but I think we can understand this notion as something that
(i) one can only do if one has the intention in question and (ii) displays this intention,
i.e., makes it clear that one has it.17 The second element seems unproblematic in
general. The third, I must confess, I do not understand. While it is quite clear
what Searle means by ‘self-referential’, I am at a loss in understanding what kind
of property ‘executive’ is supposed to be, and why explicit performatives have this
property, but other sentences do not.
Granting Searle’s conception of speech acts requiring an intention, we can also
grant that an explicit performative makes itself true if it ‘manifests the intention’
to perform the act in question. But then we must show that explicit performatives
always and automatically manifest the intention to perform the act that they describe. It must be executiveness that ensures this, but how it does so, and where it
comes from, is patently unclear. Searle again:
“[A] way to manifest the intention to perform an illocutionary act is
to utter a performative sentence. Such sentences are self-referential
and their meaning encodes the intention to perform the act named in
the sentence by the utterance of that very sentence. Such a sentence
is ‘I hereby order you to leave.’ And an utterance of such a sentence
functions as a performative, and hence as a declaration because (a) the
verb ‘order’ is an intentional verb, (b) ordering is something you can
17
Perhaps the manifestation has also to be caused by the intention in question. In this case, it
would be like the conception of ‘expressing’ employed in the accounts of assertion due to Williams
(2002) and Owens (2006) briefly discussed in Chapter 4.
CHAPTER 7. EXPLICIT PERFORMATIVES
205
do by manifesting the intention to do it, and (c) the utterance is both
self-referential and executive[.]”
(Searle 1989, p. 552)
He goes on to say that the executiveness is ‘indicated’ by hereby, by which he
appears to mean that hereby makes explicit the executiveness that would not be
explicit in the sentence without hereby:
“Performative speaker meaning includes sentence meaning but goes
beyond it. In the case of the performative utterance, the intention is that
the utterance should constitute the performance of the act named by the
verb. The word ‘hereby’ makes this explicit, and with the addition of this
word, sentence meaning and performative speaker meaning coincide.”
(Searle 1989, p.552)
We could make sense of this if we recall that Gricean speaker meaning is determined
by the speaker’s intentions. So Searle claims that an explicit performative can be
‘executive’ because the speaker intends it to be (which he could make clear by
including hereby). If so, then the speaker can simply intend to make his utterance
manifest his intention to promise.
But it is unclear what guarantees that a speaker who sincerely utters I promise p
has both the intention that his utterance be executive and the intention to promise.
Searle appears to think this has to do with the literal meaning of the utterance.
“The literal meaning of the utterance is such that by that very utterance
the speaker intends to make it the case that he orders me to leave. [. . . ]
Therefore, in making the utterance S manifested an intention to make it
the case by that utterance that he ordered me to leave.”
(Searle 1989, p. 553)
CHAPTER 7. EXPLICIT PERFORMATIVES
206
Again, this begs the question: Presumably, what makes the literal meaning of the
utterance such that by that very utterance the speaker has the right intention is the
word hereby. But is unclear how the word achieves this. Recall that Searle must
ensure that the speaker cannot not have this intention. It seems a speaker could
say I hereby promise p without having that intention, perhaps carefully avoiding
having that intention, so as to not promise what he purports to promise, so that he
cannot be later held responsible.18
Three more quotations. In summing up his argument, Searle says:
“It turns out under investigation that [the question how an explicitly
performative utterance can constitute the performance of the act named
by the main verb] is the same question as how the literal utterance of
these sentences can necessarily manifest the intention to perform those
acts.”
(Searle 1989, p. 555)
On Searle’s construal of illocutionary acts, this is indeed the question, but we can
clarify it a bit by rephrasing slightly: How does it come about that these sentences
necessarily manifest the intention in question? I fail to see how it is answered by:
“You can perform any of these acts by an utterance because the utterance
can be the manifestation (and not just a commitment to the existence)
of the relevant intention. But you can, furthermore, perform them by
a performative utterance because the performative utterance is selfreferential to a verb which contains the notion of the intention which is
being manifested in that very utterance.”
(Searle 1989, p. 556)
18
One possibility is that Searle thinks that, regardless of whether the utterance actually is a
promise (i.e., whether the speaker has the right intentions), any utterance of this form commits the
speaker—but if so, there is nothing left to explain: Then we just can define promising as taking on
the respective commitment, as I have done.
CHAPTER 7. EXPLICIT PERFORMATIVES
207
Both these sentences talk about what you can do, while the puzzle is why it is that
speakers of explicit performatives must be performing the act in question. So how
do these considerations ensure the following?
“The literal utterance of ‘I hereby order you to leave’ is — in virtue of
its literal meaning — a manifestation of the intention to order you to
leave.”
(Searle 1989, p. 556)
I conclude that Searle’s account of explicit performatives is unclear. The only
way I see it could be right is if we take the intentions he talks about to be the
publicized intentions of the speaker, which roughly correspond to the public effective
preferences in the account presented here and in Condoravdi and Lauer (2011). But
once we make this move, Searle’s original argument against assertoric accounts
loses its force: If we understand ‘public intention’ as ‘effective preference the
speaker is committed to’, self-verification follows directly from the assumption that
the declarative explicit performative has the effect that all declarative utterances
have. There is nothing left to explain.
7.6.2
Bach and Harnish (1979, 1992)
Bach and Harnish (1979) (B&H79) analyze explicit performatives as ‘indirect speech
acts’ on a par with Can you pass me a salt? when used in order to get the speaker
to pass the salt (rather than elicit information about his abilities). They provide
the following Gricean inference pattern that an addressee can (be expected to)
go through in order to infer that a speaker’s utterance of I order you to leave
constitutes an order (the version given here is the one from Bach and Harnish
(1992, p. 99)):
(7.41)
a.
He is saying ‘I order you to leave.’
b.
He is stating that he is ordering me to leave.
c.
If his statement is true, then he must be ordering me to leave.
CHAPTER 7. EXPLICIT PERFORMATIVES
d.
208
If he is ordering me to leave, it must be his utterance that constitutes
the order (what else could it be?).
e.
Presumably, he is speaking the truth.
f.
Therefore, in stating that he is ordering me to leave, he is ordering
me to leave.
Bach & Harnish and Searle’s challenge
In Condoravdi and Lauer (2011), we lumped B&H79’s account with other assertoric accounts that Searle (1989) targeted (just as Searle himself did), which we
characterized as accounts that “tr[y] to derive the performative effect by means of
an implicature-like inference that the hearer may draw based on the utterance of
the explicit performative.” (p. 150). And later we say (n. 1): “It should be immediately clear that inference-based accounts cannot meet [Searle’s challenge]. If the
occurrence of the performative effect depends on the hearer drawing an inference,
then such sentences could not be self-verifying, for the hearer may well fail to draw
the inference.”
As Garcı́a-Carpintero (2013, n. 14) points out, that this way of criticizing
B&H79’s account is a bit rash. In Bach and Harnish (1992) (B&H92), the authors
stress that they do not claim that, for an order to occur, it is necessary that the
hearer draw the inference in question. All that is necessary, they insist, is that the
speaker intend the hearer to draw this inference. Whether or not the hearer actually
does will determine whether or not the order is ‘communicatively successful’, not
whether the order actually occurs. So, on B&H92’s account, if a speaker intends to
perform an ordering with his utterance of I order you to leave, but the addressee
fails to draw the inference in (7.41), the speaker has performed an order, but that
order has not been communicatively successful.
It should be immediately obvious, however, that this feature of their analysis
does not get them out of the problem raised by Searle’s challenge: Whether or
not an ordering actually happened still depends on the presence of an intention,
though the intention for B&H92 is slightly different than the one Searle requires
(see below). That means that explicit performatives, on their account, cannot be
CHAPTER 7. EXPLICIT PERFORMATIVES
209
self-verifying, as it is possible to utter I order you to leave without having the
correct illocutionary intention.
B&H92 do not deny this:
“Searle is surely right to insist that for an utterance actually to be a
promise requires a further intention, but it is a mistake to think the
utterance should guarantee the existence of the intention [. . . ]”
(Bach and Harnish 1992, p. 103)
It is unclear then, how it comes about that an utterance of I order you to
leave is sufficient for ordering—and why any such utterance can be described as
an order. It seems that B&H92 want to simply deny that it is (“So a performative utterances is no more self-guaranteeing than a speech act of the same type
made non-performatively”, p. 104). But then, we should find utterances of explicit
performatives that are not performances of the act described by the performative
verbs.
B&H92 appear to think there are such cases, given the fact that there are sentences that have the same surface form as explicit performatives that are not performances of the act named:19
“The trouble is, an unambiguous performative sentence can be used
literally to report some habitual act, in which case one is not performing
the act named by the verb [. . . ]. You can use ‘I promise . . . ’ to report
the sort of promise you regularly make at a certain time or in a certain
situation.”
(Bach and Harnish 1992, p. 98)
19
B&H92 mention another kind of case involving double-channel communication, where one
“describe[s] some collateral act, in which case it is not in the utterance that one is performing the
act.” As examples of such collateral acts they mention signing one’s name or nodding. But doublechannel communication is no argument against a self-verification account like the one defended
here: In case of double-channel communication, there happen to be two witness for the existential
claim made, the utterance and the collateral act, instead of just one.
CHAPTER 7. EXPLICIT PERFORMATIVES
210
But this argument, which is reprised by Garcı́a-Carpintero (2013), rests on a linguistic confusion: If a sentence of the form I promise that . . . can be used both
as an explicit performative and to report one’s promising-habits, the sentence is
not ‘unambiguous’ in the sense that it has the same denotation on both these uses.
On the habitual use, the denotation of the sentence involves some kind of generic
quantification over times, occasions, cases, situations, or something of the kind,
on the performative use, it does not. But then, this is not an argument against the
claim that explicit performatives are self-verifying.
In conclusion, Bach and Harnish’s account does not derive self-verification, and
hence does not explain why an explicitly performative utterance is sufficient for the
act to be performed. At best, their account explains why an explicit performative
can be a performance of the act, not why it must be.
Bach & Harnish’s conception of promising
Despite their failure to meet Searle’s challenge, B&H79’s (and B&H92’s) account of
promising and similar speech acts, and hence their account of explicit performatives
involving such acts, is quite different from Searle’s and the conception of such acts
in the present dissertation.
In a nutshell, the difference is this: On B&H79’s account, promising does not
necessarily create commitments. To see this, consider their definition of promising:
(7.42)
Promising according to B&H79 (p. 50)
In uttering e, S promises H to A if S expresses:
i.
the belief that his utterance obligates him to A,
ii.
the intention to A, and
iii. the intention that H believe that S’s utterance obligates S to A and
that S intends to A.
First, consider the simple case in which the speaker S in fact has all the attitudes
he expresses, and that the addressee H believes that he does: In that case, it will be
true that the speaker believes that he is obligated, but not necessarily that he also is
CHAPTER 7. EXPLICIT PERFORMATIVES
211
obligated.20
Things get worse once we see that, on B&H92’s account, a speaker may express
attitudes that he does not have—he only needs to R-intend that the addressee
takes his utterance as a reason to think that he has it (cf. the discussion B&H79’s
conception of expressing in Section 4.3.1). But that means that, under definition
(7.42), a speaker can be said to have promised if he doesn’t even believe (but
R-intends his addressee to think he believes) that he is obligated to A.
This is apparently intentional:
“Commissives are acts of undertaking obligations, but to undertake an
obligation is not automatically to create one, even if S uses a performative like ‘I promise.”’
(Bach and Harnish 1979, p. 125)
I shall not try and sort out here the difference between ‘undertaking’ an obligation and ‘creating’ an obligation, or how B&H92 intend to account for the fact
that promises do create obligations, at least in many cases. Doerge (2009) contains an attempt to sort through these terminological issues. Instead, I will only
note that I concur with Doerge when he says that “I cannot promise something
without, indeed, actually creating an obligation to do the promised thing.” But
B&H79/B&H92’s account of explicit performatives, at best, can explain why an
explicit performative counts as a promise in the sense of the definition in (7.42).
Their account can explain why an explicit performative can (but does not need
to) express the attitudes listed in i.–iii. The account cannot explain why such an
utterance creates a commitment. But if the creation of a commitment is necessary
for the truth of (7.44), as I believe it is, we are left without an explanation why the
utterance in (7.43) can be reported using the sentence in (7.44).
(7.43)
John: I promise to come.
(7.44)
John promised to come.
20
Also note that even if S has the attitudes in i.–iii., he himself need not think that it is his utterance
that creates the obligation. He only needs to intend that the addressee thinks that.
CHAPTER 7. EXPLICIT PERFORMATIVES
7.7
212
Conclusion
This chapter has shown how the self-verification of explicit performatives arises
straightforwardly from the theory of clause-typing introduced in the previous
chapters, given the right lexical meanings for the verbs that feature in them, which
can be derived from the study of reportative uses of these verbs. The account makes
various welcome predictions about Boolean combinations of explicit performatives,
and explains what the difference is between verbs that can function in explicit
performatives and those that cannot.
Chapter 8
Exclamatives and expressives
I have argued that we should understand the sentential force of declaratives, imperatives and interrogatives in terms of the commitments that utterances of sentences
of these types induce, in virtue of normative conventions of use. Does that mean
all clause types should be subject to such normative conventions?
I think the answer is ‘no’. English features a clause type that I do not think is
conventionally associated with commitment: exclamatives.1 These come in at least
three sub-types: wh-exclamatives, such as (8.1a), nominal exclamatives, such as
(8.1b) and inversion exclamatives, such as (8.1c).
(8.1)
a.
(My,) How high this building is!
b.
(My,) The height of this building!
c.
(Boy,) Is this building high!
I will argue that utterances of exclamatives do not result in speaker-commitments.
Instead, I propose that exclamatives are associated with the kind of non-normative
conventions that Lewis hypothesized for clause-types more generally (cf. Section 4.2).
That is, I claim that there are (at least) two quite different ways in which a
sentences type can be connected to use. Some types are associated with normative
1
Much of the material in this chapter is informed by joint work with Anna Chernilovskaya and
Cleo Condoravdi on exclamatives (Chernilovskaya, Condoravdi and Lauer 2012, in prep).
213
CHAPTER 8. EXCLAMATIVES AND EXPRESSIVES
214
conventions: their force consists in the commitments they induce. Others are associated with Lewis conventions, their force arises differently. From a certain point
of view, this may seem unfortunate. It would be more uniform to assume that all
clause types are associated with the same kind of convention of use. But we also
want to ‘carve nature at its joints’—we want to model things that are alike in a
like fashion, and things that are not, we want to model differently. And I think
that exclamatives work quite differently from the commitment-based declaratives,
interrogatives and imperatives.
In my view, exclamatives are similar to the non-sentential utterances that Kaplan
(1999) discusses, which have been come to be called ‘expressive’. Examples are
Ouch! and Oops! While utterances of these are, on some level, informationally
equivalent to the declaratives in (8.2b) and (8.3b) (the ‘paraphrases’ are Kaplan’s),
they intuitively convey their implication in a different way:
(8.2)
(8.3)
a.
Ouch!
b.
I am in pain.
a.
Oops!
b.
I just witnessed a minor mishap.
In an oft-cited passage, Kaplan articulates his intuition that these items work quite
differently from other expressions usually studied by linguistic semanticists:
“When I think about my own understanding of the words and phrases
of my native language, I find that in some cases I am inclined to say that
I know what they mean, and in other cases it seems more natural to say
that I know how to use them.”
(Kaplan 1999, p. 4)
Another item Kaplan cites is goodbye, about which he writes:
“It is odd to even ask what it means. [. . . ] What one needs to know
about ‘goodbye’ is that it is an expression conventionally used at parting.
CHAPTER 8. EXCLAMATIVES AND EXPRESSIVES
215
Put in terms of the present discussion, what one needs to know about
‘goodbye’ is how it is used.”
(Kaplan 1999, p. 4)
Of course, as discussed in the previous chapters, for any sentence type, speakers
need to know how they are used, and that this is linguistic knowledge. But the
difference between exclamatives and the other sentence types I have discussed can
be articulated, in analogy with what Kaplan says, as follows: While for declaratives,
imperatives and interrogatives, speakers know what their normative effects are, for
expressives, speakers know when they are used.
8.1
8.1.1
Exclamatives as an expressive clause type
The two implications of exclamatives
Focussing on wh-exclamatives for concreteness, we can observe that these have two
implications, which we may call, following Castroviejo Miró (2008), the descriptive
and the expressive implication:
(8.4)
How high this building is!
a.
Descriptive implication: Sp believes that the building is (very) high.
b.
Expressive implication: Sp is surpised/amazed/struck/awed/. . . by the
height of the building.
Chernilovskaya et al. (2012) demonstrate that neither content behaves as if it is
asserted in the way the content of declaratives is asserted (an observation that had
previously been articulated, but not sufficiently supported), but that just raises the
question how either of these contents is conveyed.
Here, I want to focus on (8.4b), the expressive implication, to the exclusion of the
descriptive implication.2 The expressive implication is a lot like what is conveyed
2
Zanuttini and Portner (2003) take the descriptive content to be a semantic presupposition while
Rett (2011) treats it as an appropriateness condition in the style of Searle.
CHAPTER 8. EXCLAMATIVES AND EXPRESSIVES
216
by ouch: The utterer of (8.4) conveys that he is in a certain mental state, or has
just undergone a certain mental event. What makes exclamatives different is that
they have compositionally-determined contents. And the attitude expressed by
the exclamative is directed towards that content. It is about that content.
What kind of content? Various kinds of denotations have been proposed in
the literature on exclamatives, such as sets of propositions (Zanuttini and Portner
2003, Gutiérrez-Rexach 1996)), propositions (Castroviejo Miró 2008) and properties
of degrees (Rett 2008, 2011).
English exclamatives are restricted in a way that exclamatives in other languages
are not. They have to obey what Rett (2008) calls the Degree Restriction: They
must be about a scalar value, i.e., a degree, number or amount. This is so even if
the exclamative does not contain overt degree morphology, as in (8.5). In such a
case, a scalar dimension has to be fixed contextually.
(8.5)
What ideas she came up with in one afternoon!
Sp is surprised/struck/. . . at the number, ingeniousness, stupidity,. . . of the questions she came up with.
Consequently, English wh-exclamatives can only be formed using those wh-words
that can have degree readings—how and what.3 So regardless of what we take
the compositionally-determined denotation of English wh-exclamatives to be, we
need to assume that this denotation involves degrees—if it is a proposition, it must
be a proposition about degrees, if it is a question-denotation, it must be the kind of
denotation that degree questions have, and so forth. For concreteness, I will follow
Rett (2008, 2011) and assume that exclamatives denote properties of degrees.4
Then exclamatives convey that the speaker has a certain attitude, or just underwent a certain mental event, that is directed to the property of degrees. For ease
3
Exclamatives in other languages do not have this restriction (Chernilovskaya and Nouwen
2012): Dutch, German, Russian and Modern Greek allow who- and which-exclamatives, which get
non-scalar readings.
4
Much like discussed for imperatives and interrogatives in Section 6.4, the conventions of use
I will propose here could be reformulated to accommodate different denotation types. It is noteworthy that Rett (2011, p. 431) herself derives a proposition from the degree property on the
illocutionary level.
CHAPTER 8. EXCLAMATIVES AND EXPRESSIVES
217
of talking, I will refer to this attitude as ‘surprise’, even though the precise nature
of the attitude is much less specified: There are uses in which what is expressed is
not surprise (or that the speaker’s expectations got subverted), such as (8.6).
(8.6)
[A and B are spending a relaxed Sunday afternoon in B’s garden, enjoying
the weather. Suddenly:]
A: What a beautiful garden you have!
Intuitively, what A conveys with his utterance is not that he had not expected the
garden to be so beautiful (we can imagine he has known the garden for a long
time), but instead that he was suddenly struck by an appreciation for its beauty.5
But ‘surprise’ is close enough in many cases, and it is much more convenient to
have a single name for what is expressed by an exclamative.
8.1.2
The nature of the expressive implication
So, an exclamative like (8.7) conveys that the speaker is surprised about the (maximal)6 degree of height the building possesses. But how does it convey that?
(8.7)
How tall this building is!
Rett (2011) employs a Searle-style ‘counts-as’ rule that says that an utterance of an
exclamative counts as the expression of surprise. While this sounds correct, it is
not clear what it tells us about exclamatives with respect to a theory of language
use. What does ‘expressing’ amount to?
Commitment to be surprised?
Motivated by a desire for uniformity, we might just think that exclamatives commit
their speakers to being surprised, just as declaratives commit their speakers to
5
In Chernilovskaya, Condoravdi and Lauer (in prep.), we circumscribe this ‘attitude’ more
generally as one that (i) is directed towards a content; (ii) is intersubjectively accessible; and (iii) is
constituted by a mental event, rather than a mental state.
6
It is an open question whether the orientation to the maximal degree is semantically encoded,
or arises pragmatically.
CHAPTER 8. EXCLAMATIVES AND EXPRESSIVES
218
beliefs and imperatives commit their speakers to preferences. That is, we could
hypothesize a convention like (8.8):
(8.8)
Exclamative Convention (normative version)
When a speaker utters a wh-exclamative with content ϕ, he thereby commits himself to act as though he is surprised by ϕ.
This may sound sensible—surely, if one is surprised, that may affect the way one
acts (one might gasp, for example, or utter an exclamative). But it starts to look a
little odd once we spell out what ‘act as though’ means:
(8.9)
Exclamative Convention (normative version, explicated)
When a speaker utters a wh-exclamative with content ϕ, he thereby commits himself to choose his actions in such a way that is consistent with his
being surprised about ϕ.
In the way I have been talking about action choice, it applies to intentional action,
and it is about how one picks actions based on one’s preferences and beliefs. It
is not quite clear how ‘being surprised’ would figure into such intentional action
choice. In itself, this is not a strong argument against (8.8)—perhaps it just indicates
that we have to adjust the notion of action choice in some way. But I think there
is also something conceptually off about the idea that exclamatives commit their
speakers to an attitude such as surprise.
Here is why: Surprise is not a stable attitude. If I am surprised today that
Ariana Huffington used to be a member of the Republican party, this does not
mean that I will be surprised tomorrow—indeed, it is quite likely that I will get
used to the thought, and cease to be surprised. And this is generally true about
most instances of being surprised. There may be exceptions—a fact so baffling that
it never ceases to surprise me—but those exceptions are rare. Belief and (effective)
preference are different: While we are aware that our beliefs and preferences can
change, most of the time, we expect—at least for a large subclass of our beliefs and
preferences—that they will not, at least not in the forseeable future.
CHAPTER 8. EXCLAMATIVES AND EXPRESSIVES
219
At the same time, a crucial fact about commitments is that they are stable. That
is part of their raison d’être, or at least why they are useful: If I am committed to
something at time t, I will be committed to that thing at any later time t1 , until the
commitment gets explicitly rescinded.
It is no mystery, then, why beliefs and preferences are something an agent
might want to commit himself to: Not only are commitments good ways to inform
others about one’s beliefs and preferences, they also enable one’s audience to draw
useful inferences about one’s (likely) future actions. At the same time, taking on a
commitment to a belief or preference that one actually has will usually be harmless,
because one just commits oneself to do things that one would want to do anyway.
If the beliefs and preferences I commit myself to are ones that I both actually have
and expect to persist, then the constraints on my future actions simply constrain
my actions to do what I expect to be doing anyway.
Things are different, though, if I commit myself to an attitude I expect not to
persist—in that case, I commit myself to act in a way that I do not plan on anyway.
So committing to attitudes that are generally unstable would often be a bad thing to
do. That is to say, if exclamatives and other expressives were to commit the speaker
to such unstable attitudes, we would expect that they are used rather rarely, if at
all.
Commitments to a belief about being surprised?
So saying exclamatives ‘express surprise’ in the sense that they commit the speaker
to act surprised is a non-starter.7 The considerations in Section 6.2 suggest a
commit-based alternative. Perhaps exclamatives simply commit the speaker to a
belief that he is surprised—given that surprise is not a persistent attitude, the fact
that the doxastic commitment will persist is harmless: The speaker will remain
7
Maybe we could retreat, and say that the commitments induced by utterances of exclamatives
are short-lived, and self-destruct after a short time? We can take on such short-lived commitments
with declaratives, provided we are explicit about it: I am not convinced p is true, but for the next
part of the talk I will suppose it is. Perhaps exclamatives are simply special in that they always
create such temporary commitments. That might work, but it runs the risk of taking all content out
of the thesis that exclamatives create commitments. If these commitments are always short-lived,
how can we even tell whether such a commitment exists?
CHAPTER 8. EXCLAMATIVES AND EXPRESSIVES
220
committed to the belief that he was surprised at the utterance time, but that will
not commit him to the belief that he is surprised at later times.
(8.10)
Exclamative Convention (commitment-to-belief version)
When a speaker utters a wh-exclamative with content φ at t, he thereby
becomes committed to act as though he believes that he is/was surprised
about φ at t.
This version does not have the problem that the previous version had. Moreover,
it allows us to maintain a certain uniformity in the clause type system and it allows
us to explain why it is usually reasonable to infer that the speaker of an exclamative
is surprised: We only need to assume that he would commit to the belief only if he
had it, and that he is well-informed about his own mental state.
A first reason to be uncomfortable with this way of accounting for exclamatives
is that they would make exclamatives equivalent to claims about ones own mental
state. This does not quite accord with intuition—there is an intuitive difference
between (8.11) and (8.12):
(8.11)
How high this building is!
(8.12)
I am surprised that this building is so high.
And we find related differences that indicate that the expressive implication is not
claimed in the way the content of exclamatives is claimed (cf. Chernilovskaya et al.
2012). For example, claims can generally be challenged as dishonest, even if they
concern the internal state of the speaker, which is not possible with exclamatives:
(8.13)
(8.14)
a.
I am surprised / hungry / in pain.
b.
I don’t believe you.
a.
How high high that building is!
b. #I don’t believe you.
CHAPTER 8. EXCLAMATIVES AND EXPRESSIVES
221
More importantly, the commitment-to-belief account predicts that one can exclaim
by means of an explicit performative. Note that exclaim can be used to describe
many, if not all, utterances of exclamatives: An utterance of How high this building
is! can be reported by saying He exclaimed about the height of the building.
On the commitment-to-belief analysis of exclamatives, this strongly suggests that
(8.15a) can be paraphrased as (8.15b).
(8.15)
a.
John exclaimed about the height of the building
b.
John made an utterance that committed him to a belief that he is
surprised about the height of the building.
But then, (8.16a) should be self-verifying, and hence it should be generally possible
to report it with (8.16b), which does not work.
(8.16)
a.
John: I (hereby) exclaim about the height of this building.
b.
John exclaimed about the height of the building.
Relatedly, and I think tellingly, it is hardly ever appropriate to describe the utterance
of an exclamative with (8.17a) or (8.17b). If we have reason to doubt that the speaker
actually was surprised, it would be much more natural to instead say (8.18):
(8.17)
(8.18)
a.
He claimed that he was surprised that . . .
b.
He said he was surprised that . . .
He acted all surprised that . . .
Taking this at face value, we could say the following: Uttering an exclamative is a
way of acting surprised (as opposed to a way to commit oneself to being surprised),
while uttering a declarative is not just a way to act as if one has a belief, (but also a
way to commit oneself to a belief).
CHAPTER 8. EXCLAMATIVES AND EXPRESSIVES
222
Exclamatives as associated with a Lewis convention
If this is the right way to think about exclamatives, then exclamatives are not governed by a normative convention of the kind I proposed for declaratives, imperatives
and interrogatives, but instead by a Lewis convention like the one in (8.19).
(8.19)
Exclamative Convention (Lewis version)
Speakers of the speech community (generally) utter a wh-exclamative
with content φ only if they are surprised (struck/awed/. . . ) by φ.
That is, a competent speaker of English knows that there is a regularity in behavior
in the speech community to the effect of (8.19). To treat exclamatives in our dynamic
pragmatics, then, we would assume a general (default) belief that a speaker will
utter an exclamative only if he is in fact surprised. With this, we will have for any
belief state BAd , for any exclamative ¡ϕ!:
(8.20)
BA rutterpu, S, p¡ϕ!qqs S is surprised about ϕ.
Generalizing the convention
The exclamative convention formulated above only applies to wh-exclamatives like
(8.21a), but the other types of exclamatives—nominal exclamatives like (8.21b) and
inversion exclamatives like (8.21c)—appear to have the same kind of conventional
use.
(8.21)
a.
How high this building is!
b.
The height of that building!
c.
Is that building high!
Given the morpho-syntactic differences between the three kinds of exclamatives,
we might be tempted to simply state two additional conventions with exactly the
same content, as in (8.22) and (8.23).
(8.22)
Nominal Exclamative Convention
CHAPTER 8. EXCLAMATIVES AND EXPRESSIVES
223
Speakers of the speech community (generally) utter a nominal exclamative
with content ϕ only if they are surprised (struck/awed/. . . ) by ϕ.
(8.23)
Inversion Exclamative Convention
Speakers of the speech community (generally) utter an inversion exclamative with content ϕ only if they are surprised (struck/awed/. . . ) by
ϕ.
This duplication is rather unattractive. As discussed in Section 6.4, if we find a
morpho-syntactically heterogeneous class of sentences that all share the same use, it
is preferable to assume that all these sentences share the same denotation type, and
let the convention reference this type. If we follow Rett (2011) in assuming all three
sub-types denote properties of degrees, we can assume the following convention:
(8.24)
Degree Property Convention
Speakers of the speech community (generally) utter an matrix sentence
with a property of degrees ϕ only if they are surprised (struck/awed/. . . )
by ϕ.
For this to work, it must be the case that (a) there are no other expressions that denote
properties of degrees and that can occur unembedded (unless these expressions also
share the same use) and (b) to the extent that exclamatives can occur embedded,
the predicates that embed them have to be compatible with the hypothesized
denotation.8
8
Whether or not exclamative clauses occur embedded is an open question—Grimshaw (1979)
assumed they do, and that predicates such as surprise semantically select for exclamatives:
(i)
a.
b.
I am surprised at/by how tall this building is.
I am surprised at/by the height of this building.
But this assumption has been questioned, most forcefully by Rett (2011, Section 4.2). This debate
shows that it is by no means always obvious whether a certain clause type embeds, and consequently,
whether we can find semantic reasons for preferring a certain denotation type over another.
CHAPTER 8. EXCLAMATIVES AND EXPRESSIVES
8.2
224
More expressive meanings
I have proposed to view exclamatives as associated with a Lewis convention, rather
than a normative convention. In this section, I want to briefly mention a few other
items that can be viewed in this way—or that at least can be plausibly assumed not
to be connected to commitment.
8.2.1
Non-sentential expressions
The first candidate is, of course, Kaplan’s expressives like ouch and oops. For
these, it is tempting to directly state a convention that governs their use:
(8.25)
Ouch convention
Speakers of the speech community (generally) utter ouch only if they are
in pain.
(8.26)
Oops convention
Speakers of the speech community (generally) utter ouch only if they just
witnessed a minor mishap.
We can then capture Kaplan’s intuition that, for ouch, a competent speaker does
not know what it means, but only how it is used: We simply assume that vouchw is
undefined. Given that we can state the relevant convention of use without making
reference to its meaning or content, it would be superfluous to specify a denotation
for ouch. All that a competent speaker has to know about ouch is that it is governed
by the Ouch convention.
8.2.2
Non-at-issue meanings
Throughout most of this dissertation, I have talked as if sentences have only one
semantic value—roughly, what is usually called the ‘at-issue’ content. But of course,
sentences can have other kinds of contents, such as semantic presuppositions and
conventional implicatures (CIs) in the sense of Grice (1975) and Potts (2005).
CHAPTER 8. EXCLAMATIVES AND EXPRESSIVES
225
For each kind of content the question arises separately what its conventional
effect is. Properly construed, the Declarative Convention that I have proposed
specifies that, when a speaker utters a declarative sentence, he becomes committed
to its at-issue content. It leaves unspecified what happens with other kinds of
content.
Presuppositions. In the current set-up, it seems natural to suppose that when a
speaker utters a sentence with semantic presupposition p, he acts as though he is
already committed to p. Of course, not all theories of presupposition represent the
presuppositions of a sentence as a separate semantic value—the dynamic satisfaction semantics of the kind in Beaver (2001), for example, has only one semantic
value which encodes both presuppositional content and at-issue content, viz., a
partial update function. In the present set-up, such a function is best viewed as
an update function defined on commitment states (rather than belief states)—the
Declarative Convention can then be reconstrued as saying that when a declarative is uttered, the denotation of the uttered sentence gets applied to his doxastic
commitments.
Conventional implicatures of supplements. For other kinds of content, we must
ask, in every particular case, what their effect is. The kinds of CIs discussed by Potts
fall into two broad classes: supplements and expressives. Supplements are things
like appositive relative clauses (as in (8.27)) and nominal appositives (as in (8.28)).
For these, I agree with Potts that they share their main dynamic effect with main
clause assertions—i.e., I think supplements create doxastic speaker commitments.
(8.27)
Mary, who you met last night, is a painter.
(8.28)
Lance Armstrong, a cancer survivor, won the Tour de France.
Integrated expressives.
Another class of items that Potts (2005) discusses under
the label ‘CI’ are expressive items which are similar to ouch and oops, but which
are, or appear to be, syntactically integrated into a another clause. Potts (2007)
CHAPTER 8. EXCLAMATIVES AND EXPRESSIVES
226
details a number of characteristics that these items have:
(8.29)
a.
Independence: Expressive content contributes a dimension of meaning that is separate from the regular descriptive content.
b.
Nondisplaceability: Expressives predicate something of the utterance situation.
c.
Perspective dependence: Expressive content is evaluated from a particular perspective. In general, the perspective is the speakers, but
there can be deviations if conditions are right.
d.
Descriptive ineffability: Speakers are never fully satisfied when they
paraphrase expressive content using descriptive, i.e., nonexpressive,
terms.
e.
Immediacy: Like performatives, expressives achieve their intended
act simply by being uttered; they do not offer content so much as
inflict it.
f.
Repeatability: If a speaker repeatedly uses an expressive item, the
effect is generally one of strengthening the emotive content, rather
than one of redundancy.
Supplements (arguably) also have (8.29a-c) but (8.29d-f) are particular to expressive
items.
For some of these, in particular the subclass Potts (2005) calls these Isolated
CIs, the treatment suggested for ouch and oops seems entirely appropriate. An
example is the expressive use of fucking, illustrated in (8.30) and (8.31).
(8.30)
Thats fantastic fucking news!
(Potts 2005, p. 65, (3.40b))
{ Sp is in a heightened emotional state.
(8.31)
[status message on a social networking site]
dear america. get a fucking handle on gun laws ASAP.
{ Sp is in a heightened emotional state.
CHAPTER 8. EXCLAMATIVES AND EXPRESSIVES
227
For fucking, whose contribution does not interact with the content of its host clause
in any way, we can simply assume that it is semantically vacuous (i.e., vfuckingw is
the identity function) and assume that fucking is directly associated with its own
convention of use:
(8.32)
fucking convention
Speakers of the speech community (generally) utter fucking only if they
they are in a heightened emotional state.
But not all integrated expressives can be treated in such a simple manner (not all
expressives are ‘isolated’ in Potts’ sense). Take epithets like bastard, as in Kaplan
(1999)’s (8.33).
(8.33)
That bastard Kaplan got promoted to tenure.
{ Sp dislikes Kaplan.
Bastard is different from fucking in that it does not just signal an attitude of
the speaker, but it signals an attitude towards something, and this something is
determined by the expression bastard syntactically combines with: (8.33) does not
signal that the speaker does not like someone, it signals that the speaker dislikes
Kaplan.
What’s more, vKaplanw plays two roles in the interpretation of (8.33): It serves
as an argument both for the expressive contribution of bastard, but also for the
main clause predicate got promoted to tenure. An account of such non-isolated
expressive items requires a limited interaction between the specification of the use
and the system of semantic composition. Potts (2007) offers a dynamic treatment of
such items that can be viewed as consistent with the way I have been talking about
expressive items here. The expressive denotations directly manipulate a feature
of the context, the expressive index, which keeps track of evaluative attitudes that
speakers have expressed. Potts stresses this feature of his account:
“Conceptually, it is important to keep in mind that [the expressive denotations] are quite fundamentally different from the usual denotations
CHAPTER 8. EXCLAMATIVES AND EXPRESSIVES
228
in semantics, in that they have access to the context parameter, which
normal denotations cannot manipulate. In this specific sense, they are
metalogical. I regard this move as necessary to doing justice to the
immediacy property.”
(Potts 2007, p. 188)
If we give Potts’ expressive indexes an epistemic construal (i.e., we take them
to represent what interlocutors believe about each other’s attitudes), we can see
his dynamic treatment as analogous to the Krifka (to appear)-style variant of the
current proposal as sketched in Section 6.5. The semantic value is a context change
potential that directly updates the epistemic state of the interlocutors.
8.2.3
Expressive vs. prescriptive conventions
I mentioned already that Kaplan (1999) takes expressions like goodbye to be of the
same kind as ouch and oops. They certainly are similar in that there is an intuition
that a competent speaker needs to know about them how and when they are used,
rather than knowing what they mean.
Goodbye is different from things like ouch and fucking, however, in that intuitively, it does not express any kind of attitude. That does not mean that we cannot
treat it by means of a Lewis convention (such as that in (8.34)), but we could also
treat it in terms of prescriptive rules such as those in (8.35) and (8.36).
(8.34)
goodbye convention (Lewis version)
Speakers in the community utter goodbye only when parting.
(8.35)
goodbye convention (Normative precondition)
One must not: Utter goodbye except when parting.
(8.36)
goodbye convention (Prescriptive rule)
One should: Utter goodbye (or . . . ) when parting, and not otherwise.
It is not quite clear how to decide which one is more appropriate. (8.36) seems
attractive because it might allow us to capture the intuition that if an agent departs
CHAPTER 8. EXCLAMATIVES AND EXPRESSIVES
229
without acknowledging this by means of an utterance like goodbye, he will often
be felt as ‘having done something wrong’—and this is kind of essential to the
meaning/use of goodbye. The very function of the expression is to satisfy this
requirement.
The same is true for things like honorific markers in languages like Japanese or
the formal/familiar versions of pronouns in Romance and Germanic languages. If
we analyze these in terms of prescriptive rules, rather then commitments or Lewis
conventions, we directly capture the sense that a speaker who misuses them has
committed a faux pas.9
If these considerations are on the right track, there are three ways an expression
can have a conventionally-specified use: (i) in virtue of a normative convention
specifying commitments, (ii) in virtue of a Lewis conventions specifying when
typically is used and (iii) in virtue of a prescriptive rule specifying when it should
and should not be used.
8.2.4
Determining the correct convention type
Every type of matrix sentence must have a conventionally specified use, or so I
have argued. In addition, every type of content a sentence can have must have a
conventionally specified use, and these uses may vary with the type of (non-matrix)
expression that contributes the additional content.
And there are at least two, and quite possibly three, different possibilities for
the conventional specification of use. In each case where we want to specify the
conventional use of an expression (or one of its meanings), we must hence ask
ourselves: How is it specified? By means of a Lewis convention? A normative
convention inducing a commitment? A prescriptive rule?
It would be advantageous to have a set of diagnostics, at least rough and
ready ones, that can generally be applied to answer this question, but I am afraid
9
But of course, this is not the only way to capture this intuition. Potts and Kawahara (2004) and
Potts (2005, 2007) offer treatments of these items that propose that they express attitudes, just as
fucking and bastard do. Such treatments can capture the intuition that a misuse of these items is a
faux pas by hypothesizing that there are social rules that mandate or prohibit the expression of the
attitudes involved.
CHAPTER 8. EXCLAMATIVES AND EXPRESSIVES
230
I cannot offer such a set at present. In the discussion in this chapter and the
previous ones, I have made arguments for specific cases, but most of these do not
generalize readily. A stringent requirement for retraction is a good reason to favor
the thesis that a given expression type (or an implication thereof) is governed by a
normative convention, rather than a Lewis convention—but for this to be a strong
argument, the retraction requirement has to be more stringent than those due to
communicative reasons. In the case of declaratives, the phenomenon of loose talk
(Section 4.7) provides a nice test case, but for many other types of expressions, it is
not clear that we can find similar phenomena.
Potts (2007)’s properties of expressives may be taken as a guide in the other
direction: If the effect of an expression is ‘immediate’ and ‘descriptively ineffable’
that may serve as an indication that it is associated with a Lewis convention—but do
note that the creation of a commitment, in the present theory, is just as ‘immediate’
as the expression of an attitude. Similarly, immediacy and descriptive ineffability
arguably are also properties (at least possibly) of items that are associated with
prescriptive rules—Potts argues that honorifics and formal/familiar pronouns have
these properties, but his arguments are not incompatible with the intuition I just
articulated that they are associated with prescriptive rules.
In some cases, we may appeal to introspection: Someone who uses German
du where Sie would have been appropriate intuitively has committed a social
transgression—suggesting that a prescriptive rule is at play. Someone who says
ouch even though he is not in pain is not felt to have committed such a transgression,
though if he consistently does so, we might start to wonder whether he knows how
that word is used properly. But what about someone who says goodbye, before
animatedly continuing the conversation?
In some cases, it may well not matter, and, in line with what I said about the
denotational type of imperatives in Section 6.4.3, maybe there are cases where there
is no fact of the matter whether a given item is governed by a Lewis convention or
by a normative rule—some speakers take it to be one, others take it to be the other,
but due to the nature and function of the item, it simply never makes a difference.
In other cases, there will be difference in how the item is used, and so it will be
CHAPTER 8. EXCLAMATIVES AND EXPRESSIVES
231
relevant to correctly identify how the use condition in question is conventionally
determined. Whether or not this is necessary will depend, of course, on our
explanatory goals in a particular case.
8.3
Conclusion
Much of this chapter has been quite tentative. It was aimed at establishing three
points:
(i) There is good reason to think that there are clause types and other expressions
that are unlike declaratives, imperatives and interrogatives, in that they do
not, by linguistic convention, induce commitments for the speaker.
(ii) We hence need to ask, for any particular semantic implication that an expression has, how this implication is related to use.
(iii) While I have proposed that there are two distinct ways this can happen,
commitment-inducing normative conventions and Lewis conventions, there
may well be other types, such as prescriptive rules.
And even though I have pointed out that the question how the conventional effect
is determined may be unanswerable in particular cases, a theory of language use
will require an answer in many others. It is not an accident that declaratives
and imperatives are often used in planning joint action and to exchange factual
information, while exclamatives are usually used to emote, and honorifics and
formal/familiar pronouns are used to negotiate social relationships. I believe there
is much systematicity to be uncovered once we know which distinctions to pay
attention to, and what the space of theoretical possibilities is.
Chapter 9
Conversational implicatures
The present chapter is concerned with Conversational Implicatures, the topic
that Grice (1975) was mainly concerned with when he laid the foundations of the
approach to pragmatics that has come to be called ‘Gricean’. The previous chapters
have applied a broadly Gricean perspective to a range of issues concerning the
conventional constraints on force that are conditioned by the type of a clause. These
chapters illustrate one of the main points this dissertation aims to make, namely
that such a perspective is fruitful beyond the study of conversational implicatures.
At the same time, an understanding of the conventional force of sentences is
a prerequisite for studying the kinds of pragmatic inferences that are the concern
of Grice’s theory of implicature. Grice’s theory starts from the idea that a speaker
uttered a (declarative) sentence in order to convey information, that is, in order to
make the hearer believe that the truth-conditional content of the sentence is true.
It hence makes sense that we have fixed a theory of clause-typing before we delve
into the study of implicature.
The present chapter has two aims: On the one hand, it shows how ‘optimizationbased’ theories of implicatures fit into the framework of dynamic pragmatics. On
the other hand, it shows that such optimization-based theories make surprising,
and correct, predictions about the nature of conversational implicatures that are at
odds with what (Neo-)Griceans have long believed.
232
CHAPTER 9. CONVERSATIONAL IMPLICATURES
233
By ‘optimization-based’ analyses of implicatures, I mean any analysis that construes conversational implicatures as inferences the addressee can draw by assuming that the speaker chooses his utterance so as to best satisfy a set of constraints—
be they a set of conversational maxims, a more over-arching cooperative principle, preferences of the speaker, or a combination of these (Grice’s theory was
optimization-based in this sense, and I submit that every theory that deserves
the name ‘Gricean’ must be). The last decade in particular has seen a variety
of proposals about how this optimization can be modeled formally, drawing on
ideas in optimality-, decision- and game-theory (Blutner 2000, Parikh 2001, Blutner
2002, van Rooij 2004, Benz and van Rooij 2007, Jäger 2007, Franke 2009, Jäger
and Ebert 2009, Franke 2011, Jäger 2012, a. o.). I do not want to discuss the
differences of these approaches, nor do I wish to present an alternative to them,
or exhaustively discuss how they fit into the existent literature on conversational
implicatures (Franke (2009), in particular Ch. 1, does a good job of concisely situating these formal analyses within the landscape of Neo-Gricean and Post-Gricean
approaches to implicatures, and Jäger (2012) provides a useful summary of the further development of game-theoretical models since). Instead, I want to show how
the conception of these optimization-based models fits naturally into the present
framework.1 To this end, I offer some preliminary considerations in Section 9.1 and
then show in Section 9.2 how some classic cases of implicatures can be modeled in
the present set-up using the simplified conception of the decision procedure Opt
developed in Chapter 5.
1
Though the underlying conception fits naturally, I do not want to claim that integrating any of
these analyses into the present framework is trivial: All of them make quite particular assumptions
about the representation of beliefs and preferences that the current model does not make.
The decision- and game-theoretic models, in particular, assume a probabilistic representation
of graded belief, combined with a representation of preferences by means of real-valued utility
functions. It is worth noting that until very recently, the probabilistic aspect of these representations
has not been exploited in any essential way. This may make us wonder whether the probabilistic
apparatus is necessary to account at least for standard cases of implicatures.
In the context of implicature-like inferences in reference resolution, Degen and Franke (2012),
however, have recently employed a variant of Franke (2009, 2011)’s ‘iterated best response’ model
to derive and test probabilistic predictions, and Frank and Goodman (2012) have proposed a closely
related model that does the same. These results indicate that the added fine-grained representational
structure that such models assume may indeed be necessary.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
234
In Section 9.3, I will show that the conception of pragmatic reasoning that underlies the framework of dynamic pragmatics—and indeed, any optimization-based
analysis of implicatures, once construed appropriately—predicts the existence of
conversational implicature with quite non-standard properties. In particular, it
predicts that there are implicatures that are neither optional nor cancelable, which
yet are entirely Gricean in nature. This has important consequences, as these properties have often been used as more or less definitive tests to determine whether
an observed implication of a sentence is an implicature or not.
A note on terminology: I use the term ‘conversational implicature’ (which I will
abbreviate to ‘implicature’ throughout, as conventional implicatures do not play a
role in what is to follow) in an utterance-centered way: An utterance implicates p if
it gives rise to the pragmatic inference that p, or if it at least licenses that inference,
in its context of use. This use of the term ‘implicate’ is at variance with that of
Bach (2006), who insists that implicatures are not inferences. Bach’s conception is
speaker-centered, in that it is speakers, not utterances that implicate things, and
in that a speaker only can be said to implicate something if he intends to convey
it. Though Bach claims my use is based on a misconception, I do not think we
actually differ on any substantive issue. It is likely that Bach agrees that what I
call ‘implicature’ works the way I say it does—he simply objects to my calling the
inferences in question ‘implicatures’. Bach acknowledges that my use of the term
‘implicature’ is common2 and it is useful for my purposes, so I see no reason to
defer to Bach’s usage.3 In Section 9.2.5, I will briefly return to the issue whether
there is a crucial difference, in the present perspective, between implications of
utterances that are intended by their speakers and those that are not.
2
To pick a random example from a well-known textbook on pragmatics: “For implicatures are not
semantic inferences, but rather inferences based on both the content of what has been said and some
specific assumptions about the co-operative nature of ordinary verbal interaction.” (Levinson 1983,
p. 104)
3
The terminological difference boils down to this: An implicature-according-Bach is an
implicature-according-Lauer that the speaker intends to convey.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
9.1
235
Preliminaries
This section provides some necessary preliminary considerations for accounting for
implicatures in the present system. Section 9.1.1 discusses the role (or lack thereof)
that maxims of conversation play in an analysis of implicatures as it is conceived of
here. Section 9.1.2 discusses the role of alternative utterances in implicature derivations. Section 9.1.3, finally, introduces a second kind of preference (in addition to
the outcome preferences our system can model already), and introduces the way I will
use optimality-theoretic tableaus to display the outcome of the decision procedure
Opt in the rest of this chapter.
9.1.1
Maxims and preferences
Grice (1975) derived implicatures on the basis of his cooperative principle.
(9.1)
Cooperative Principle
(Grice 1975, p. 45)
Make your conversational contribution such as is required, at the stage at
which it occurs, by the accepted purpose or direction of the talk exchange
in which you are engaged.
More specifically, Grice proposed a set of Maxims of Conversations, “the following
of which will, in general, yield results in accordance with the Cooperative Principle”
(p. 45–46):
(9.2)
a.
Maxim of Quality
Try to make your contribution one that is true.
b.
(i)
Do not say what you believe to be false.
(ii)
Do not say that for which you lack adequate evidence.
Maxim of Quantity
(i)
Make your contribution as informative as is required (for the
current purposes of the exchange).
(ii)
Do not make your contribution more informative than is required.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
c.
236
Maxim of Relation
Be relevant.
d.
Maxim of Manner
Be perspicuous.
(i)
Avoid obscurity of expression.
(ii)
Avoid ambiguity.
(iii) Be brief (avoid unnecessary prolixity).
(iv) Be orderly.
The idea was, roughly, that interlocutors take each other to follow these maxims
as best as they can—i.e., that they obey the maxims in choosing which utterances
to make. Implicatures then arise as assumptions that are necessary to justify the
speaker’s utterance in light of the maxims, or, where the maxims are in conflict or
ostentatiously violated, in light of the more general cooperative principle.
While subsequent authors, in particular in the ‘Neo-Gricean’ tradition, often
were concerned with revising, extending and clarifying these maxims (early examples are Atlas and Levinson (1981) and Horn (1984)), recent elaborations of Grice’s
theory, in particular those employing variants of decision- and game-theory, have
largely dispensed with direct appeals to Maxims of Conversation and the Cooperative Principle. Instead, a speaker’s utterance choice is simply understood, just as
any action choice, as the result of the speaker attempting to realize his preferences,
based on his beliefs. The preferences in question may, in a particular case, be motivated by general principles of cooperative behavior, but they also may arise from
other sources.
This is the view that I will adopt here. In general, I think it is more useful to
think of the Gricean maxims as speaker preferences which just happen to be very
common, either because they follow from basic assumptions of cooperativity, or
because they are due to our make-up as social beings.4 Implicature derivation
should then be done on the basis of the assumption (appropriate in some and
perhaps most contexts, but not others) that these preferences are in place, instead of
4
This is arguably quite close to Grice’s view: Recall that he characterized the maxims as a
principles that, if followed, will yield behavior that accords with the cooperative principle.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
237
directly appealing to the maxims. This has two advantages: (i) Not every preference
that figures in implicature calculations must be justified as a general principle of
human interaction and (ii) our models extend more readily to situations in which
there is a significant conflict of interest—i.e., where agents may have (selfish)
preferences that defeat some principles of cooperative behavior.
9.1.2
Alternative utterances
Up to now, I have been silent how the set of alternative actions available to an agent
at a given time is determined. I simply have assumed that this is determined by the
structure of our models (in particular, by the set of available alternative branches
of the branching-time model). I will continue to do so in the present chapter, but
seeing as implicature calculation depends quite heavily on the reasoning about
particular alternative utterances, I want to briefly address the question of how we
should think about alternative utterance actions in the present context.
Preconditions on actions
In a model like our current one, it is tempting to think about the set of actionalternatives among which an agent decides as all those actions that are ‘executable’
by the agent at the time. In general, actions come with preconditions. If I am
currently in my office, for example, I simply cannot perform the action of walking
from my house to the local coffee shop, because the preconditions for this action
are not satisfied. So it is tempting to think of the set of actions an agent decides
between in a world w at a time t as just the set of all actions whose preconditions
are met at t in w.
On some level, this may be correct, but in the case of utterances, it would leave
the set of alternative actions extremely unconstrained: In most situations, an agent
can, in principle, say just about anything. So it is more attractive to think of the set
of alternative utterances the agent chooses between as a subset of those utterances
he could, in principle, make. The agent does not decide between all alternative
utterances, but only those that are salient in the given context. Which actions these
CHAPTER 9. CONVERSATIONAL IMPLICATURES
238
are will depend, in part, on what has happened in the current discourse so far and
what the current topic of conversation is. If the agent has just been asked whether
he mopped the floor, for example, the utterances I mopped the floor and I did not
mop the floor (or simply yes and no) can naturally be assumed to be salient, but
they need not be the only ones.
The importance of alternative utterances
In the following, I will assume that the set of salient alternative utterances is
extrinsically given, but I want to stress that this assumption is not at all innocent.
In particular cases (and depending on the choice of the decision procedure Opt and
the speaker preferences that can be assumed), it can make a significant difference
which alternative utterances are considered. Take for example the classic example
of a scalar implicature in (9.3).
(9.3)
Some students came to the party.
{ Not all students came to the party.
We can summarize the way this implicature is derived classically as follows (this
is basically what Franke (2009) calls ‘naive scalar reasoning’):
(9.4)
The speaker uttered (9.3). Alternatively, he could have uttered All students
came to the party, which would have been relevant information, and is
informationally stronger than (entails) (9.3). Hence I can conclude that the
speaker does not take this stronger alternative to be true.
But it has long been recognized (at least since Atlas and Levinson (1981)) that
this way of accounting for the implicature in (9.3) does not work in quite this
fashion if we assume that Some but not all students came to the party is one of the
alternatives that the addressee takes into account. For if it is, he should analogously
reason as in (9.5), deriving the opposite of the implicature to be explained.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
(9.5)
239
The speaker uttered (9.3). Alternatively, he could have uttered Some but
not all students came to the party, which would have been relevant information, and is informationally stronger (entails) (9.3). Hence I can conclude
that the speaker does not take this stronger alternative to be true.
In the more recent literature, this problem has come to be called the ‘symmetry
problem’,5 for obvious reasons. One way to solve it is to simply deny that the some
but not all alternative is taken into account when reasoning about the speaker’s
utterance choice—but this assumption needs to be motivated, seeing as there is no
obvious reason to exclude it categorically from consideration.6
Alternatives considered by the hearer
Crucially, what matters for implicature calculation is what the addressee takes the
action alternatives to be (and what the speaker believes the addressee takes the
alternatives to be, and so on), not necessarily the alternatives that the speaker
actually considers. We can probably safely assume that he considers all the possible
utterances that he thinks the addressee expects, but he may consider additional
ones.
An important wrinkle is that the set of alternative utterances that the addressee
considers may well depend on the utterance that is actually performed. For the
speaker, it intuitively makes sense that the set of alternative utterances he takes into
account is fixed ex ante, i.e., before he makes his utterance. But for the addressee,
this is not necessarily the case. It may well be that the very utterance he observes
makes salient certain alternatives he would otherwise not have taken into account
when reasoning about the speaker’s choice of utterance.
5
The name was coined in unpublished 1997 class notes by Kai von Fintel and Irene Heim.
There are other ways to deal with this problem—e.g., by assuming that there is a ‘manner’
preference favoring the form some over the form some but not all (which is longer, and more
complex), but that there is no such preference between some and all. Whether or not this solves
the problem depends on the way this preference is represented and the precise Opt assumed,
and may depend on other factors—e.g., in Franke’s IBR model, if some but not all is included
in the alternatives, an additional restriction on belief-revision strategies is necessary to derive the
implicature, in addition to the assumption that some but not all is ‘more costly’ than some and all.
(see Franke (2009, p. 77–89) for discussion).
6
CHAPTER 9. CONVERSATIONAL IMPLICATURES
240
This may explain why certain implicatures arise fairly reliably—in particular
the ones that Grice called ‘generalized conversational implicatures’, such as the
some{not all implicature illustrated in (9.3). The idea is this: Whenever an
addressee observes an utterance containing some, then, as a matter of cognitive
fact, this automatically brings to mind the possibility of uttering the same sentence
with all in place of some, so if the other prerequisites for the implicature are met,
the hearer will infer the scalar implicature.7
This is a way to capture the idea that these ‘generalized’ implicatures arise because they rely on ‘conventionalized’ or ‘lexicalized’ scales of linguistic expressions
(Horn 1972, Gazdar 1979), in a Gricean way: The implicature itself is not conventionalized in any way, but there is a strong association between the members of the
scale which ensures that an utterance containing one member of the scale makes
the other members of the scale salient.8
It is tempting to also assume that this idea of ‘automatic salience of alternatives’
can also help us solve the symmetry problem: Perhaps the only alternative expressions that are taken into account are those that are suggested by the utterance that
is actually performed. This seems to be what Neo-Griceans like Horn (1972) assumed, and it is also the solution favored by Franke (2009, 2011) in a game-theoretic
setting. It is not clear, however, whether this can work as a general solution. It
is one thing to suppose that some automatically makes salient all, but does not
make some but not all salient.9 It is quite another to assume that a certain form is
7
If this explanation is on the right track, we would expect that we can detect this ‘automatic
activation’ of alternative utterances using psycholinguistic methods (e.g., by testing for priming
effects). To my knowledge, this has not been done in any systematic manner for the alternatives
involved in classic examples of ‘generalized’ implicatures.
8
Of course, this does not mean that such a lexical association is required—in a given context,
alternative expressions can also be salient for other reasons, for example, because a certain question
was asked. Hirschberg (1985) was the first to stress the importance of contextually-triggered
alternative expressions.
9
Over the years, there have been various proposals for how the set of ‘salient alternative expressions’ is constructed. The classic proposal by Atlas and Levinson (1981) relies on the fact that
some and all are lexicalized, while some but not all is not. Horn (1989) requires that the items on
a scale (which in turn are used to derive the salient alternative expressions) have the same monotonicity properties, ruling out a scale that contains both some and some but not all. In recent years,
Katzir (2007) and Fox and Katzir (2011) have proposed that the alternatives relevant to implicature
calculations are determined with reference to the syntactic structure of the uttered expression, in
CHAPTER 9. CONVERSATIONAL IMPLICATURES
241
generally not considered at all. Considerations of salience can explain why certain
forms are always considered whenever a certain other form has been uttered, but it
is harder to see how such considerations can generally exclude certain alternatives
from consideration.
Be that as it may, in the following I will simply assume, as is common practice,10
that appropriate alternative sets are given and illustrate how, given this assumption,
implicatures can be modeled in the system of dynamic pragmatics.
9.1.3
Two types of preferences and a visual representation of Opt
In the rest of this chapter, it will be useful to have a succinct way to represent
the interlocutors’ preferences and the way they interact in decision making. Even
though the definition of Opt I have adopted in Section 5.2.2 (which, again, is
adopted here only for the sake of concreteness) is rather simple, it will be useful to
have a way to quickly visualize the predictions that are made.
Luckily, there is a familiar visual representation that we can adopt without
much ado, given the way Opt is defined: Tableaus as used in Optimality Theory
(Prince and Smolensky 1993). For a given w, t such that Agtpw, tq “ i, we can let
the set of candidates be Actpw, tq and the set of constraints EPi pw, tq. Constraint
evaluation is done as follows: Candidate action a violates a constraint/preference c
iff Bi,t,w ras 2 c.11
particular ruling out from consideration any expression that is syntactically more complex than
the uttered expression—just as Atlas and Levinson, Katzir and Fox thus build a preference for
simple expressions into the alternative-selection process, while classically, such preferences (or the
corresponding Maxim of Brevity) have been taken to play a role reasoning about given alternatives.
10
For example, Sauerland (2004, p. 373): “I should mention that there is one important question
that I have nothing to say about here: Namely, the question where quantitative scales come from. I
shall here simply take quantitative scales for granted and use them to account for the implicatures
of sentences.” Similarly, Franke (2009, p. 127): “[t]his issue is strictly speaking orthogonal to the
concerns of [game-theoretic pragmatics], which is a theory of reasoning about alternative messages
and not a theory of alternatives as such.”
11
It is noteworthy that this version of OT is unidirectional and models things from a production
perspective, much like work in OT syntax (Bresnan 2000). OT-theoretic work in semantics (Hendriks
and de Hoop 2001) instead has tended to take a comprehension perspective, where the candidates
are possible interpretations of an observed form. Work in OT pragmatics, instead, has taken a dual
perspective with BiOT (Blutner 1998, Blutner 2000).
CHAPTER 9. CONVERSATIONAL IMPLICATURES
242
This takes care of the preferences that we have represented so far as propositions
which characterize properties of outcomes that the agent prefers. In the context of
implicature calculation, however, we will need to be able to represent a second kind
of preference, which we may call action preferences. These represent preferences
between actions the agent has independently from the outcome of the actions. In
game-theoretic treatments, such preferences are usually modeled as costs that attach
to actions. I will use a similar representation here and model action preferences
as functions f that map actions to integers, with the interpretation that a is better
than a1 according to f if f paq ă f pa1 q. The sets EPi pw, tq thus no longer contain only
propositions, but also costs functions. In the OT representations, the number of
violation marks an action a incurs on such a constraint will simply be f paq.12
Given the simplified language our agents speak, I adopt here a single, simple
action preference that models a preference for shorter or less complex expressions:
(9.6)
Minimize
a.
If e is not of the form utterp¨, ¨, ¨q, then Minimizepeq “ 0.
b.
If e is of the form utterp¨, ¨, ϕq, then Minimizepeq is the number of symbols in ϕ.
For more realistic object languages, it is an open question what counts as ‘shorter’
or ‘simpler’: We could count phones, syllables, morphemes, words or maximal
syntactic projections, or there could be a more complex, structure-based measure of
simplicity or shortness. Which one is adequate is, ultimately, an empirical question.
The advantage of our simple, propositional object language is that we can side-step
this issue here. I will assume that Minimize is universally present in the speaker’s
effective preference structure, i.e., that every speaker strives to minimize the length
of his utterances, and that all agents take each other to minimize the length of
their utterances. In general, however, this preference will be ranked below the
more important outcome preferences. That is, agents strive to minimize the length
of their utterances, but they will opt for a shorter form only if everything else is
12
In classical OT terms, these action preferences are thus treated essentially as markedness constraints.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
243
equal, i.e., if a shorter form and a longer form are otherwise equally well-suited
to achieve the agent’s goals. It is here where the lexicographicness of admissible
decision procedures (cf. Section 5.2.2) comes into play: This property ensures that
lower-ranked preferences are indeed treated as such ceteris paribus preferences,
and never win out over higher-ranked ones.
9.2
Some classical implicatures
In this section, I will show how some classical implicatures can be derived in the
present system. Again, the purpose is not to provide an exhaustive account of these
implicatures, but rather to show how optimization-based accounts of implicatures
fit into the current system. For concreteness, I use the simple version of Opt defined
in Section 5.2.2, but this should not be taken as a proposal competing with existing
ones. Instead, as before, I want to abstract away from the question what the correct
Opt is, and instead focus on how, given a suitable Opt, we can model implicatures
in the present system.
Before looking at some examples of classical implicatures, it is worth noting that
we have seen implicature-like reasoning in the previous chapters, even though the
inferences in question usually are not discussed as such. One example is the
reasoning in Section 6.1.3 for deriving ‘advice’ uses of imperatives:
(9.7)
[Context: Doctor to patient, in a discussion about how to cure patient’s
current illness.]
Take these pills for a week.
{ The doctor believes taking the pill is the best way/one of the best ways
to cure the patients current illness.
The reasoning presented there is quite similar, in many ways, to classical relevance
implicatures like Grice (1975)’s (9.8):
(9.8)
[Context: Ad is standing next to his obviously immobilized car.]
Ad: I am out of petrol.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
244
Sp: There is a garage round the comer.
{ Sp has no reason to believe that the garage is closed, out of petrol to sell,
etc.
In both cases, the addressee can reason to a conclusion about the speaker’s beliefs,
based on contextual assumptions about the speaker’s preferences and the utterance
that took place. It is a great advantage of having a pragmatic theory that incorporates a suitable theory of clause-typing that these two inferences can be shown to
arise in essentially the same way.
Contextual assumptions
I have already introduced the preference Minimize for shorter forms, which I take to
be universally present. Throughout, I will also assume the following high-ranked
preference, named after the Gricean maxims it approximates, for any $ ϕ P L$ :13
(9.9)
Quality
”
ı
λv v pbSp pϕq Ñ Sp ϕ
‘The speaker is committed to believe ϕ only if he actually believes ϕ.’
However, this assumption is crucially different from the assumption that Minimize
is universally present: It is an assumption that only makes sense in certain contexts,
while I want to think of Minimize as a preference that speakers truly take each other
to always have (though it will be frequently defeated by higher-ranked preferences).
In previous chapters, we have seen one case where a preference like (9.9) is not
present (or at least is outranked by a preference that defeats it): Loose talk.14
(9.9) may also be absent (or outranked by a contradictory preference) in cases of
substantial conflict of interest between the interlocutors, i.e., where the speaker
13
I once again suppress temporal parameters throughout for the sake of readability. For the same
reason, I will often refer to utterpSp, Ad, ϕq as utterpϕq, and will write Optpwq for the outcome of Opt
in w at the time t under discussion, i.e., Optpwq “ OptpBAgtpw,tq,w,t , EPAgtpw,tq pw, tq, Actpw, tqq.
14
In cases of loose talk, there will generally be other preferences that make the speaker ‘stay close
to the truth’, such as a general, highly-ranked preference for not violating one’s commitments (i.e.,
not being AtFault), together with a possibly lower-ranked preference against retracting existing
commitments.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
245
cannot be assumed to be fully cooperative. But here, we focus on cases where there
is no looseness involved, and where speaker are taken to be fully cooperative—
that is, the cases of cooperative information exchange that Gricean theory has
traditionally focused on.
I also assume throughout that the speaker’s pre-utterance belief state satisfies
the following:15
(9.10)
BSp rutterpSp, Ad, ϕqs Ad ϕ
‘If the speaker utters ϕ, the addressee will come to believe ϕ.’
9.2.1
A ‘relevance’ implicature
We start with a case that could be viewed as a complete abstraction of a relevance
implicature, such as the petrol example in (9.8) above. The level of abstraction will
be so high that the modeling is not of much interest in itself, but it serves as useful
warm-up to more interesting cases.
I introduce another outcome preference. (9.11) captures the notion that p is
‘relevant’: If p is true, then the speaker wants the addressee to know this.
(9.11)
Inform p
λv rv p Ñ Ad ps
‘If p is true, then the addressee believes p.’
While we have assumed that the addressee knows that the speaker has the preferences Quality and Minimize, we assume that the addressee does not know whether
p is ‘relevant’, i.e., whether the speaker has Inform p as a preference, but that he
assumes that the speaker has no other preferences.16
15
This is actually automatically ensured if the speaker believes that the addressee believes that
Quality outranks all other preferences of the speaker, as will be the case in the examples we examine
in this section.
16
The assumption that the speaker has no other preferences is quite artificial, but it is less problematic to assume that his other preferences are orthogonal to the present utterance choice. As discussed
in Section 6.1.3, in full generality, we also want to allow for a notion of salience of preferences, so
that even possibly-interfering preferences can be ignored when not salient.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
246
This means there are three kinds of worlds in BAd :
(9.12)
a.
Worlds vp`rel such that Sp p is true and
EPSp pvp`rel , tq “ tQuality, Inform p, Minimizeu
b.
Worlds vp`
EPSp pv
c.
rel
rel , tq
Worlds v
p
such that Sp p is true and
“ tQuality, Minimizeu
where the speaker does not believe p to be true.
I assume that ActSp pw, tq contains four actions:17
(9.13)
ActSp pw, tq “
tK, utterpSp, Ad, pq, utterpSp, Ad, qq, utterpSp, Ad, p ^ qq, utterpSp, Ad, p _ qqu
The outcome of Opt at vp`rel -type worlds is shown in Fig. 9.1.18 The only optimal
action is utterpSp, pq. By contrast, the outcome at vp`
rel -type
worlds is shown in
Fig. 9.2. In these worlds, the null utterance K is optimal. And clearly, in v
p
type
worlds, uttering p cannot be optimal, given Quality. That means we have the
following:
(9.14)
a.
Optpvp`rel q “ tutterppqu
b.
Optpvp`
c.
utterppq < Optpv p q
relq
“ tKu
But then, updating with utterpSp, Ad, pq will remove all worlds that are not of type
vp`rel and hence:
(9.15)
17
@w P Ba rutterppqs : Inform p P EPSp pwq
K is the null action (or any other non-utterance action), utterpSp, Ad, pq the one that will be
performed, utterpSp, Ad, qq another one whose content is known to be irrelevant, utterpSp, Ad, p ^
qq stands for utterances of anything that is longer than p, but entails it, utterpSp, Ad, p _ qq is
representative of utterances that are longer than p but do not entail it.
18
With respect to this and the following tableaus, recall that Minimize is an integer-valued action
preference that is treated like a markedness constraint. In Fig. 9.1, I also mark Quality with a ? for
candidate q, as it is irrelevant whether or not it is violated (which depends on whether or not the
speaker believes q), given that uttering q violates Inform p.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
vp`rel
K
+
utterppq
utterpqq
utterpp ^ qq
utterpp _ qq
247
Quality Inform p
*
?
Minimize
*
*
***
***
*
*
Figure 9.1: Decision in worlds vp`rel where the speaker believes p and takes it to be
‘relevant’
Input
Quality
+
K
utterppq
utterpqq
?
utterpp ^ qq
utterpp _ qq
Figure 9.2: Decision in worlds vp`
take it to be ‘relevant’
rel
Minimize
*
*
***
***
where the speaker believes p and does not
That is, from observing utterppq, the addressee learns that the speaker takes p to be
relevant.
9.2.2
A scalar implicature
Scalar implicatures constitute a more interesting case. The example we are going
to look at is:
(9.16)
Ad: Do you know the current address for John? I need to send him a letter.
Sp: He is in Europe.
{ Sp does not know where in Europe John is.
Let e stand for the proposition that John is in Europe and p for the proposition that
John is in Paris (and assume that for all worlds, p ñ e, capturing that p is stronger
than e). Given Ad’s question (assuming that Sp is cooperative), we can assume that
throughout Ad’s information state, Sp’s preferences are given in (9.17):
CHAPTER 9. CONVERSATIONAL IMPLICATURES
vp
Quality Inform p
K
*
+ utterppq
utterpeq
*
Inform e
*
248
Minimize
*
*
Figure 9.3: Decision in worlds vp where the speaker believes p
ve
Quality Inform p
K
utterppq
*
+ utterpeq
Inform e
*
Minimize
*
*
Figure 9.4: Decision in worlds ve where the speaker believes e, but not p
(9.17)
tQuality, Inform p, Inform e, Minimizeu
And assume that Sp’s action choices are those in (9.18) (omitting utterances whose
content is known to be irrelevant, as those will always be dominated by K):
(9.18)
utterppq, utterpeq, K
Figures Figs. 9.3 to 9.5 show the outcome of the decision procedure at worlds at
which Sp knows that John is in Paris, worlds in which he only knows that John is
in Europe, and worlds where he does not even know that. The result is in (9.19).
(9.19)
a.
Optpvp q “ utterppq
b.
Optpve q “ utterpeq
c.
Optpv
e q
v
Sp e
+
“K
K
utterppq
utterpeq
Quality Inform p
*
*
Figure 9.5: Decision in worlds v
Inform e
Minimize
*
*
e
where the speaker believes neither e nor p
CHAPTER 9. CONVERSATIONAL IMPLICATURES
vp` rel Quality Inform e
K
*
+ utterppq
+ utterpeq
Figure 9.6: Decision in worlds vp`
assumption that p is not relevant
rel
249
Minimize
*
*
where the speaker believes e and p, on the
But then, assuming all worlds in BAd are of one of the three types:
(9.20)
BAd rutterpeqs Sp p
This is a simple version of a scalar implicature: p is stronger than e, so if the speaker
utters e, the addressee can conclude that the speaker does not know whether p.19
Of course, this inference is optional. It does not always arise, but it instead
depends on the addressee making the correct contextual assumptions. The crucial
assumption here is that the speaker has the preferences Inform p and Inform e—
i.e., that p and e are both relevant and that the speaker prefers to communicate
the stronger of the two (p) if he believes it to be true. That is, the addressee must
assume that the speaker takes the information provided by p, but not by e, to be
relevant in the context of conversation.
If the addressee’s belief state includes worlds in which this is not given, i.e., in
which Inform p is not in place, the only decision that changes is that for vp -type
worlds (recall that in all those worlds vp`
rel
Sp,t pp ^ eq since p Ñ e is taken to
be common geographical knowledge). The decision in those worlds is shown in
Fig. 9.6, and we have the following
(9.21)
19
a.
Optpvp`rel q “ tutterppqu
b.
Optpve`
c.
Optpve q “ utterpeq
d.
Optpv
e q
rel q
“ tutterppq, utterpequ
“K
Alternative implicatures are possible, if the speaker has (or might have, according to what
the addressee knows) other preferences that prevent him from uttering the stronger alternative.
Cf. Section 9.3.3 below.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
But then, updating with utterpeq will let both vp`
rel
250
and ve worlds survive. So
we do not derive the scalar implicature in the addressee’s information state:
BAd rutterpeqs 2
(9.22)
Sp ppq
This is of course correct: Scalar implicatures only arise if the stronger alternative is
relevant:
[Context: Ad and Sp are in the US]
(9.23)
Ad: Is John in town?
Sp: He is in Europe.
6{ Sp does not know where in Europe John is.
9.2.3
The ‘epistemic step’
The scalar implicatures we have derived so far were rather weak:
(9.24)
Scalar implicature
BAd rutterpSp, eqs Sp p
This ‘epistemic’ implicature is completely adequate in a case like (9.25).
(9.25)
John is in Europe
{ Sp does not know where in Europe John is.
But in other familiar cases of scalar implicatures, what is perceived is something
stronger:
(9.26)
Some people came to the party.
{ (a) Sp does not know that all students came to the party.
{ (b) Sp believes that not all students came to the party.
{ (c) Not all students came to the party.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
251
But the stronger implications in fact arise only if (9.26) is uttered in the right context:
If the speaker is taken to be (honest and) well-informed about which students were
at the party, we derive the strongest implication (9.26c). If, instead, the speaker
is taken to only believe he is well-informed about which students came, we only
derive the intermediate one in (9.26b).
Gazdar (1979) derived (9.26b) directly as an implicature (in addition to the
weaker (9.26a)), but it seems conceptually more attractive to derive (9.26b) and
(9.26c) on the basis of (9.26a), in contexts where this is appropriate, as proposed by
Horn (1989). We can derive these additional implications as contextual strengthenings of the epistemic implicature (Sauerland (2004) calls this the ‘epistemic step’).
It works much like the inference to the truth of the content in Chapter 2. We
characterize the following two contextual conditions:20
(9.27)
a.
‘Speaker has an opinion about p’
Sp ppq _ Sp p pq
b.
‘Speaker is an expert on p’
p ô Sp ppq
(9.27a) gives rise to the ‘intermediate’ implicature that the speaker believes the
negation of the stronger alternative, while (9.27b) gives rise to the strong implicature
that the stronger alternative is false.
Fact 4. Let BAd as in the previous section, and further BAd (9.27a). Then
BAd rutterpSp, eqs Sp p pq
Fact 5. Let BAd as in the previous section, and further BAd (9.27b). Then
BAd rutterpSp, eqs 20
p
Both Sauerland (2004) and Russell (2006) derive the ‘epistemic step’ in essentially the same way
as I do here.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
9.2.4
252
Intended implicatures
The implicatures we have derived so far were not (necessarily) intended inferences.
We did not assume (and did not need to assume), that the speaker preferred to
communicate the content of the implicature to the hearer. But of course, implicatures can be intended to be communicated. They can even be the main motivation
for an utterance. The system as introduced so far captures this.
Firstly, observe that speakers can be aware of the addressee-inferences their
utterances will trigger. The following fact captures this for the strong ‘factual’
implicature from the previous section.
Fact 6. Let BSp be such that for all v P BSp , BAd,v satisfies the assumptions made in
Section 9.2.2 and in addition BAd,v (9.27b). Then
BSp rutterpeqs Ad p pq
Now consider a situation in which the speaker has the preferences the addressee
took him to have in Section 9.2.2, but he also has the another one Inform
p, with
21
the obvious interpretation. His preferences are:
(9.28)
Quality ą Inform p ą Inform
p ą Inform e ą Minimize
The crucial case to investigate is that of worlds in which the speaker knows e ^ p.
Fig. 9.7 shows the tableau on the assumption that the speaker does not believe that
the addressee will draw the scalar implicature and Fig. 9.8 shows the decision in
case he does believe the implicature will be drawn.22
So if, but only if, Sp believes that Ad’s information state is such that he will
draw the implicature, he will prefer uttering e over the longer e ^ p. However, if
he takes it to be possible that the addressee will not draw the implicature, he will
21
The precise ranking of the Inform-preferences is not essential—indeed, we could assume that
all preferences are unranked, with the exception of Minimize, which needs to be subordinate to the
outcome preferences.
22
As discussed in Section 9.1.2 above, the simple implementation of implicature-reasoning given
here depends on the assumption that (the speaker believes that) the addressee does not take into
account the possibility of the utterance e ^ p.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
vi e^ p
Quality Inform p
K
utterppq
*
+ utterpe ^ pq
utterpeq
Inform
*
*
253
p
Inform e
*
*
*
Minimize
*
****
*
Figure 9.7: Decision in worlds vi e^ p where the speaker knows that e is true and p
is false, and believes the implicature will not be drawn
vi p^e
Quality Inform p
K
utterppq
*
utterpe ^ pq
+
utterpeq
Inform
*
*
p
Inform e
*
*
Minimize
*
****
*
Figure 9.8: Decision in worlds vi p^e where the speaker knows that e is true and p
is false, and believes the implicature will be drawn
instead opt for the longer expression. This arguably accords with intuition.
These are but the first two steps in the kind of back-and-forth reasoning that
has played a role in a number of recent accounts of pragmatic phenomena. Franke
(2009, 2011)’s Iterated Best Response model and its variants and precursors (e.g.
Jäger and Ebert (2009), Degen and Franke (2012)) explicitly define an iterated
reasoning chain alternating between the speaker and hearer perspectives to derive implicatures. Recent work by Frank and Goodman (2012) employs a very
similar model that makes quantitative predictions which match observed frequencies of implicature calculation in referent-selection tasks, and Vogel, Potts and
Jurafsky (2013) use Decentralized Partially Observable Markov Decision Process
(Dec-POMDP) models to capture the same kind of iterative reasoning. They apply
their model to implement artificial agents that perform implicature-reasoning in
the course of solving a collaborative task.
An advantage of the explicitly-stepwise nature of these theories is that they
can model bounded rationality, i.e., they can allow that whether or not a certain
inference is drawn depends on the reasoning capabilities of the interlocutors. This
CHAPTER 9. CONVERSATIONAL IMPLICATURES
254
becomes particularly useful in accounting for some inferences in cases of deceptive
language use (as argued by Franke (2008), Franke, de Jager and van Rooij (2012)),
as a speaker can be taken to assume that he can ‘outsmart’ the addressee.
The framework developed in this thesis is not in conflict with these approaches,
but an abstraction over them. It tries to get by with rather minimal assumptions
about the optimization procedure (such as lexicographicness), while the mentioned
approaches generally take a rather specific procedure and apply it to a particular
phenomenon, perhaps refining the procedure in the process. The focus here is not
on the details of the optimization procedure, but rather on the way the belief that
such an optimization procedure is followed results in pragmatic inferences on the
part of the addressee.
It is noteworthy that the initial scalar inference on the part of the addressee
required little pragmatic sophistication (a point that also clearly emerges in the
accounts of Franke and Frank and Goodman, where simple scalar inferences like
this generally arise on the first iteration step). All that the addressee had to attribute
to the speaker was a preference to not say things he does not believe, a preference
to inform and a preference to minimize effort.
9.2.5
Unintended implicatures
The scalar implicatures in Section 9.2.2 and the strengthening in Section 9.2.3 were
derived without assuming an intention (or effective preference) of the speaker.
We did not assume that (the addressee thinks that) the speaker wants to inform
the addressee of the truth of the implicature. Only in Section 9.2.4, did we add
additional assumptions to this effect. But the basic scalar inference arose without
it.
This is empirically adequate, as Horn (2006, p. 37) notes when discussing Ariel
(2004)’s example in (9.29): “it must be recognized that in [some] cases we DO
implicate what harms, if not defeats, our local purpose.”
(9.29)
The majority decided for peace. Me too.
(bumper sticker, originally Hebrew, spotted 4.2002)
CHAPTER 9. CONVERSATIONAL IMPLICATURES
255
Crucially, someone who reads (9.29) is likely—and intuitively warranted—to infer
that not everyone decided for peace. As Ariel points out, in this case, the inference
in fact weakens the speaker’s rhetorical point—so we can safely assume that he
did not intend to communicate its content. The case is exactly parallel to the scalar
implicature we derived before: p is the stronger statement that everyone decided for
peace, e is the (asserted) weaker statement that the majority decided for peace. We
predict the implicature to arise assuming only the preferences Quality, Inform p,
and Minimize. Crucially, what we do not need to assume is a preference Inform p.
Indeed, even assuming the opposite preference would not change the outcome of
the implicature calculation—as long as Quality, the preference for not committing
to falsehoods, is ranked above the preference for not informing the addressee of p.
This, again, accords with intuition. The ‘speaker’ of (9.29) could avoid informing
his addressee that not everyone decided for peace—by uttering Everyone decided
for peace. But he can only do so if he is willing to commit to a belief in the content
of this stronger sentence.
From the perspective of a theory of language use, there does not seem to be
much of a difference between such scalar inferences that are unintended and those
that are intended. If the speaker has the right kind of preference, we may infer, in
addition to the scalar inference, that he intended to communicate its content; but
that does not impact the way the inference arises in the first place.
Some terminological hardliners (e.g., Bach (2006)) have insisted that a (potential)
inference must be intended in order to count as an implicature, at least according
to how Grice himself used the term. And while Grice’s characterization of ‘conversational implicature’, quoted below, does not mention the term ‘intention’, it uses
‘implicate’ in the definiens. Grice does not give a definition of the simpler term,
but the way he introduces it makes clear that he took implicatures to be part of
what is meant by a speaker. And speaker meaning, for Grice, was what the speaker
intended to communicate. So perhaps Bach is right that Grice generally thought
that implicatures must be intended.
“I am now in a position to characterize the notion of conversational
implicature. A man who, by (in, when) saying (or making as if to
CHAPTER 9. CONVERSATIONAL IMPLICATURES
256
say) that p has implicated that q, may be said to have conversationally
implicated that q, provided that (1) he is to be presumed to be observing
the conversational maxims, or at least the Cooperative Principle; (2)
the supposition that he is aware that, or thinks that, q is required in
order to make his saying or making as if to say p (or doing so in those
terms) consistent with this presumption; and (3) the speaker thinks (and
would expect the hearer to think that the speaker thinks) that it is within
the competence of the hearer to work out, or grasp intuitively, that the
supposition mentioned in (2) is required.”
(Grice 1975, p. 49–50)
And yet, Grice concludes his ‘general pattern for working out of a conversational
implicature’ as follows:
“he intends me to think, or is at least willing to allow me to think, that
q, and so he has implicated that q.”
(Grice 1975, p. 50)
So it seems that Grice agreed that, at least from the addressee’s perspective, there
is not much of a difference between inferences that the speaker intends and those
that he merely tolerates.
In any case, I shall continue to use ‘implicature’ in the broad sense in which it
applies to (potential) pragmatic inferences, regardless of whether the inference is
intended by the speaker.
9.3
Need a Reason: Mandatory Gricean implicatures
The previous sections are mainly intended to show how optimization-based analyses of implicatures fit into the current system, by illustrating how a particular
optimization procedure (i.e., the Opt-function defined in Section 5.2.2) can be used
to derive certain standard implicatures.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
257
In the rest of this chapter, I will argue that this conception of pragmatic inference
makes some surprising predictions concerning the properties of implicatures. In
particular, I will argue that properly pragmatic inferences can lack certain properties that they are frequently taken to have by necessity. This result is of considerable
significance, as these—putatively necessary—properties have often been used as
tests to distinguish pragmatic inferences from semantic implications of sentences.
In particular, I will argue that pragmatic inferences can be mandatory, in the
sense that they arise on every use of a given form. And I will argue that any
optimization-based account of pragmatic inference, once conceptualized in the
right way, predicts the existence of mandatory implicatures. The perspective of
the framework of dynamic pragmatics lets us see that this prediction is not at all
conceptually suspect. And once we know what to look for, it is not hard to find
empirical support for the idea that mandatory implicatures indeed exist.
9.3.1
Optionality and cancelability
In contrast with the claim that I want to establish here—that there are implicatures
that are mandatory—it is frequently supposed that all pragmatic inferences are by
necessity both optional and cancelable.
Optionality and cancelability are about the relationship between a linguistic
expression and a (potential) implication that utterances of the expression have.23
For present purposes, the expressions in questions will always be sentences. An
implication i of a sentence S is optional if there can be sincere utterances of S that
do not give rise to i; and it is cancelable if a sincere utterance of S is compatible with
a denial of i even in contexts in which i would normally arise.
These two properties are often grouped under the label cancelability, e.g. by
Horn (1984):
23
In full generality, it would be perhaps more appropriate to say that optionality and cancelability
are properties of the relationship between utterance types and their (potential) implications. Generally, though, the relevant utterance type can be usefully identified by the linguistic expression that
is uttered.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
258
“[. . . ] the aforementioned inference goes through in unmarked contexts, but it may be cancelled—explicitly [. . . ] or implicitly (by establishing the appropriate context, [. . . ]).”
(Horn 1984, p. 13)
For what follows, it will be convenient to have a separate label for the two
properties—and I prefer the term ‘optionality’, as I do not subscribe to Grice’s
view that (some) implicatures arise by default unless the the context prevents
them—as formulations such as ‘the context cancels the implicature’ suggest.
Many implicatures are optional and cancelable, of course. We saw in the last
section how standard scalar implicatures depend on the stronger alternative expression being relevant, and hence will not arise in a context in which the information
provided by this alternative is not relevant. And they are cancelable—an utterance
of (9.30) is compatible with an explicit denial of the scalar inference, as in (9.31):
(9.30)
Some students came to the party.
{ Sp does not know that all students came to the party.
{ Not all students came to the party.
(9.31)
Some students came to the party. In fact, all of them did.
Hirschberg (1985) pointed out that it is actually strange to speak of the implicature
being ‘canceled’ in (9.31), on the Gricean assumption that implicatures are part of
speaker meaning, and hence must be intended: Presumably, the speaker of (9.31)
did not intend to convey that not all students came to the party when he uttered
the first sentence. But then, if an implicature is present only if the speaker intended
the addressee to recognize it, then there is no implicature to be canceled in (9.31).
Of course, there still is a (potential) pragmatic inference that the hearer may have
drawn when the speaker uttered the first sentence. And it is this inference that
the speaker intends to prevent, or ‘call off’ with his second sentence. So speaking
of implicatures being canceled already presupposes a view on which implicatures
need not be intended.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
259
But even setting this issue aside, there is a tension in talking about cancelation of
pragmatic inferences in optimization-based theories: By assumption, speakers only
make utterances that best satisfy their goals and preferences. Now, it seems that
in (9.31), we must assume that the speaker wants to avoid making the addressee
believe that not all students came. If he would not mind the addressee forming
this belief, why did he utter the second sentence? But then, if he chose his previous
utterance optimally, should he not have opted for All students came to the party
in the first place? It seems that optimization-based theories predict that cancelation
never happens.24
Mayol and Castroviejo (2013) articulate this problem for the traditional Gricean
account (which, after all, is an optimization-based account):
“Why do [cancellations] exist at all if, within the Gricean program, they
should be viewed as uncooperative? That is, how can we make sense of
a discourse where the same speaker first utters a weaker statement and,
immediately after, a stronger one? Why did he not utter the stronger
statement to begin with?”
(Mayol and Castroviejo 2013, p. 85)
Mayol and Castroviejo (2013) also note that, despite the wide-spread belief (cf. below) that cancelability is a necessary property of conversational implicatures, and
the fact that cancelability is often used as a test to determine whether a given
implication is an implicature, cancelation “has hardly been studied itself”.
Optimization-based theories can avoid predicting the complete absence of cancelations if they somehow can explain why optimization yields a different outcome
for the first and second sentence in (9.31).25 There are multiple ways of doing
24
I am grateful for Chris Potts (p.c.) for drawing my attention to this fact.
There is an alternative route to go: Assume that speakers do not always make optimal utterances,
even though their audiences believe they do. On this construal, cancelation triggers belief revision
on the part of the addressee, making him conclude that the speaker’s first utterance was not optimal
after all—this is a possible way to construe such cases, but it makes one wonder why hearers persist
in their belief that speakers make optimal utterance choices after they have heard them cancel
implicatures multiple times.
25
CHAPTER 9. CONVERSATIONAL IMPLICATURES
260
so, which all involve the notions of salience, awareness or attention. One way to
account for the cancelation in (9.31) is to assume that, when uttering the first sentence, the speaker was unaware that (the addressee might think that) the stronger
alternative All students came to the party is relevant, but that this occurred to him
after uttering the sentence.26 As we saw, the addressee will only infer the implicature if the stronger alternative is supposed to be relevant. So, when the speaker
made his first utterance, his choice was optimal, because he did not think that the
implicature would arise. If he then becomes aware of the (potential) relevance of
the stronger alternative, and hence of the possibility that the addressee drew the
implicature, we understand why his second utterance is optimal, too.
Similarly, we could assume that, in choosing his first utterance, the speaker
simply did not take into account the possibility that he could also have uttered the
stronger alternative, but that this occurred to him after the fact. This is slightly different than the assumption that he considered the stronger alternative, but deemed
it irrelevant. In the latter case, he failed to take into account one of his preferences;
in the former, he failed to take into account an action-alternative altogether. But
the result will be the same: As the stronger alternative did not occur to him when
making the first utterance, that utterance was optimal, while after it occurred to
him, it was optimal to utter the stronger alternative, canceling the implicature.
If this is on the right track, then we can fully account for cancelation only once
we have a representation of salience (or awareness) and its dynamics. But this
limitation is no hindrance for our present concerns. For note that in any case, the
assumption must be that the speaker’s conception of the context shifted between
his two utterances: While making his first utterance, the speaker took himself
to be in a context in which his utterance would not give rise to the implicature;
while making his second utterance, he took himself to be in a context in which
the implicature did arise (or in which it is at least possible that it did arise). That
means that optionality is a necessary condition for cancelability. An implicature i
26
This possibility seems to be more or less what Mayol and Castroviejo (2013) have in mind when
they articulate their central constraint on implicature cancelation (p. 88): Canceling is only possible
if the Question Under Discussion (QUD in the sense of Roberts 1996) that the cancelation addresses
is different from the one that the previous utterance addressed.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
261
of a sentence S is cancelable only if we can fix contextual assumptions such that an
utterance of S does not give rise to i. As I will argue that there are implicatures that
are non-optional, it follows immediately that they are also non-cancelable.
Grice himself thought that being cancelable is a conceptually necessary property
of implicatures as he understood them:27
“Finally, we can now show that, conversational implicature being what
it is, it must possess certain features: 1. [. . . ] it follows that a [. . . ]
conversational implicature can be canceled in a particular case.”
(Grice 1975, p. 57)
And we find little disagreement in the literature. Unsurprisingly, we find statements to the same effect in standard textbooks, e.g., the classic Levinson (1983) and
the more recent Kadmon (2001):
“But the implicature, like all implicatures, is defeasible [. . . ]”
(Levinson 1983, p. 138)
“It is uncontroversial that conversational implicatures are defeasible.”
(Kadmon 2001, p. 6, emphasis in original)
And we find it echoed in the current literature, where it is used to argue against
the pragmatic status of certain inferences (cf. Section 9.3.5):
“Within the Gricean theory, scalar implicatures are pragmatic inferences.
Hence, they have a weak status: they are optional, cancellable, and
suspendable.”
(Magri 2009, p. 253)
27
I elide Grice’s reasoning here for dramatic effect, and will return to it below, in Section 9.3.4.
The strong statement in this quote is quite uncharacteristic for the for the hedge-prone Grice. Later
in Grice (1978), he was more cautious (emphasis mine): “I think that all conversational implicatures
are cancelable.”
CHAPTER 9. CONVERSATIONAL IMPLICATURES
262
So the idea that conversational implicatures are always optional and cancelable
was there from the start, it is asserted without hedging in standard textbooks, and
it is used, as a critical test, without arguing for it. I think it is safe to say that the
idea is wide-spread and generally accepted.
And yet I shall argue that it is mistaken: A Gricean conception of pragmatic
inference is quite compatible with such inferences being mandatory.28 In the following sections I will argue that there is a clearly circumscribable class of implicatures
that are neither optional nor cancelable, and hence are mandatory. I call these
‘Need a Reason’ (NaR) implicatures, for reasons that will become clear shortly. I in
Section 9.3.2, I will discuss a simple example of an implicature of just this kind, and
I will show how this implicature comes out as mandatory in the present framework
in Section 9.3.3. Subsequently generalize the conditions under which this happens
in Section 9.3.4.
9.3.2
The ignorance implicature of disjunction
In the present section, I want to argue that there is an implication, which typically
is assumed to be an implicature, which is indeed neither optional nor cancelable:
The ‘ignorance’ implication of unembedded disjunction. When a speaker utters a
sentence containing an unembedded disjunction, such as (9.32a), his audience is
often licensed to conclude that the speaker does not know which of the disjuncts is
true.
(9.32)
28
a.
John is in London or he is in Paris.
b.
Speaker does not know that John is in London.
c.
Speaker does not know that John is in Paris.
Hirschberg (1985, p. 24–32) quite clearly saw that optionality and cancelability do not follow
at all from the way Grice defined conversational implicatures, and hence proposed to require that
implicatures must be cancelable by definition.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
263
This implication is quite robust, at least in cooperative contexts. It is so robust,
in fact, that Zimmermann (2000) took it to be an entailment.29 But, as I will show,
we can explain the robustness of this implication perfectly well in purely Gricean
terms, on the assumption that it is a conversational implicature.
If one accepts that the implication is an implicature, it may seem like a gardenvariety case of a scalar implicature: There are alternative, stronger sentences (viz.,
John is in London and John is in Paris) that the speaker could have uttered instead,
and so we conclude that he does not have sufficient evidence for the propositions
expressed by those sentences. Isn’t this just like a speaker uttering John is in
Europe, where the audience can infer, due to the existence of an alternative, stronger
statement (John is in London) that he does not know that the stronger statement
is true?
The cases are indeed similar, but the implicature of disjunction is different in
that it cannot be straightforwardly called off, at least not in the way this is usually
done with scalar implicatures—namely, by asserting the alternative that triggered
the implicature:
(9.33) ??John is in London or he is in Paris. In fact, he is in Paris.
Intuitively, (9.33) is odd because if the speaker knows that John is in Paris, we are
left to wonder why he uttered the disjunction in the first place. Even though I have
called it the ‘ignorance’ implicature of disjunction, the content of the implicature is
not what is given in (9.32), but something more general. That is why it is possible
to conjoin (9.32a) with an assertion that the speaker knows which disjunct is true:
(9.34)
John is in London or he is in Paris. And I know which one.
Grice (1978) assumes that (9.34) is an instance of implicature cancelation. Instead, I
want to view the second sentence as excluding a possible strengthening of a more
general implicature. Note that, when (9.34) is uttered, the audience will still need to
infer a reason for why the speaker uttered the disjunction rather than just asserting
29
Zimmermann was largely concerned with embedded uses of disjunctions, but his account predicts
that (9.32a) entails both (9.32b) and (9.32c).
CHAPTER 9. CONVERSATIONAL IMPLICATURES
264
one of the disjuncts—a natural one, in this case, is that the speaker is unwilling to
share this information. In full generality, then, we can characterize the implicature
as follows:
(9.35)
John is in London or he is in Paris.
{ The speaker had a reason for not uttering either disjunct.
Why would or trigger such an implicature? Eckardt (2007) puts it like this:30
“In using a disjunction, the speaker necessarily has to mention two properties which are usually more specific. These properties are presented
as salient and relevant. The simpler sentences are salient alternative
utterances in context. The hearer hence will look for a reason why the
speaker chose a more complex expression in order to give less information.”
(Eckardt 2007)
In a nutshell, the reasoning is this: There is a general preference for uttering shorter
or less complex sentences. This preference can be overridden if the speaker has
reasons to utter the longer sentence. In many cases, a speaker who chooses to make
a longer utterance (or multiple utterances instead of one) will be motivated by his
desire to convey more information. However, when uttering an unembedded
disjunction instead of one of its disjuncts this cannot be his motivation, as the
disjunction conveys less information than the shorter alternatives. So the speaker
must have had another reason for not uttering either of the disjuncts.
It is clear, then, why the implicature cannot be canceled by asserting one of
the disjuncts: If the speaker is willing and able to assert one of the disjuncts, why
did he utter the longer, more complex disjunction in the first place? And we can
also see why the implicature is not optional: If the preference for a shorter form is
always in place (though it may be overridden), then whenever a speaker utters a
30
Like Zimmermann (2000), Eckardt (2007) is mainly concerned with embedded uses of disjunctions, but her reasoning applies also to the unembedded uses discussed here.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
265
disjunction, he must have had a reason to avoid the otherwise-preferred shorter
disjuncts.
It is useful to consider circumstances in which the disjuncts themselves are are
not more relevant than the disjunction, i.e., contexts in which the information which
disjunct is true does not need to be conveyed. Suppose that all that matters, in a
given context, is whether John is in the USA so that Mary can call him. Suppose
further that Sue knows that John is in London. Intuitively, we would expect Sue to
utter (9.36a) and not (9.36b):
(9.36)
a.
John is in London.
b.
John is in London or Paris.
This is so despite the fact that the information that (9.36b) provides is sufficient,
given Mary’s informational needs. If Sue were to utter (9.36b), Mary would likely
and appropriately infer that Sue does not know where John is (or that she has a
reason not to share this information), even though, by assumption, this information
is not relevant.
The case is quite different with a run-off-the-mill scalar implicature. Suppose,
in the same context, Sue instead utters (9.37).
(9.37)
John is in Europe.
Even though (9.37), in an appropriate context, would implicate that Sue does not
know where in Europe John is, in our present context, it does not, because it is
simply not relevant where in Europe John is. Unlike the disjunctive utterance,
(9.37) is not significantly longer, and not more syntactically complex than more
specific alternative utterances.
In summary, when a speaker utters a disjunction p _ q, the following facts
conspire to trigger an uncancelable, non-optional implicature:
1. There is a ceteris paribus preference for shorter, less complex forms, hence,
everything else being equal, uttering p and uttering q is preferrable to uttering
p _ q.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
266
2. p _ q is asymmetrically entailed by p and by q, hence an utterance of p conveys
the information that p _ q is true.
3. Because of 1. and 2., if the speaker wanted to convey p _ q, and nothing
prevented him from asserting p, he would have done so.
4. The speaker just uttered p _ q instead of p.
It follows that there is something that prevented the speaker from uttering p. In
a cooperative dialogue, often the only plausible reason is that the speaker does
not know that p is true, which is why in many contexts, the addressee can infer
ignorance.
9.3.3
The NaR implicature of disjunction in dynamic pragmatics
I now show how the NaR implicature of disjunction arises in the framework of
dynamic pragmatics. We need at least the following alternative utterances:31
(9.38)
K, utterppq, utterpqq, utterpp _ qq
It is of course crucial that p and q are considered as utterance alternatives. With
Eckardt (2007), I think it is very natural to assume that both disjuncts are made
salient by the utterance of the disjunction.32
In terms of preferences, we already have encountered the crucial preference for
shorter linguistic forms Minimize, as well as the outcome preferences Quality and
Inform ¨. Here I shall assume throughout that Inform p _ q is present, i.e., that
if p _ q, the speaker prefers that the addressee know this. If we assumed also a
preference Inform p, then we would be in a variant of the scalar implicature scenario
considered in Section 9.2.2.33 Instead, we assume such a preference is absent. As
31
I use the same notational simplifications as used in Section 9.2.
Interestingly, this is poses some problems for classical theories which, like Gazdar (1979)’s,
derive alternatives by a simple replacement of logical operators in the uttered sentence. Sauerland
(2004) offers a technical fix for this problem, but it is tempting to simply give up the assumption
that alternatives are derived by simple replacement.
33
The only difference would be that the weaker expression now incurs more violations on Minimize, but that would not change decision outcomes.
32
CHAPTER 9. CONVERSATIONAL IMPLICATURES
267
discussed in the previous section, we still should predict that the implicature arises.
And we do.
Deriving ignorance
In order to derive the ignorance implication (as opposed to more the general implication that the speaker ‘has a reason’), assume that Ad believes that the speaker
has no other (relevant) preferences, i.e.,
(9.39)
For all v P BAd : EPSp,v “ Quality ą Inform p _ q ą Minimize
In worlds in which Sp does not believe p _ q to be true, the optimal action is
obviously K, as it is the only one that does not violate Quality. There are four
logical possibilities for worlds in which Sp believes p _ q to be true, which differ in
terms of which of p and q the speaker believes:
Sp p
v
p^ q
vp^
v
Sp q
False False Fig. 9.9
q
True
False Fig. 9.10
p^q
False
True
Fig. 9.11
True
True
Fig. 9.12
vp^q
The outcome of Opt in each of these types of worlds is illustrated in Figs. 9.9 to 9.12,
and summarized in (9.40).
(9.40)
“ tutterpp _ qqu
a.
Optpv
b.
Optpvp^
c.
Optpv
d.
Optpvp^q q “ tutterppq, utterpqqu
p^ q q
q q
“ tutterppqu
p^q q
“ tutterpqqu
As a result BAd rutterpp _ qqs will only contain worlds of type v
p^ q —that
is, after
observing the utterance of p _ q, the addressee will believe that the speaker believes
neither p nor q to be true. That is the ignorance implicature.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
w
p^ q
K
utterppq
utterpqq
+ utterpp _ qq
Quality Inform p _ q
*
*
*
Figure 9.9: Decision in worlds v
p^ q
w
p^q
K
utterppq
+
utterpqq
utterpp _ qq
w
+
+
p^q
K
utterppq
utterpqq
utterpp _ qq
q
*
*
***
p^q
Minimize
*
*
***
where the speaker knows p but not q
Quality Inform p _ q
*
*
Figure 9.11: Decision in worlds v
Minimize
where the speaker knows neither p nor q
wp^ q
Quality Inform p _ q
K
*
+
utterppq
utterpqq
*
utterpp _ qq
Figure 9.10: Decision in worlds vp^
268
Minimize
*
*
***
where the speaker knows q but not p
Quality Inform p _ q
*
Minimize
*
*
***
Figure 9.12: Decision in worlds vp^q where the speaker knows both p and q
CHAPTER 9. CONVERSATIONAL IMPLICATURES
269
Alternative implicature: Unwillingness to inform
Ignorance is not the only reason why a speaker could opt for the disjunction instead
of uttering either disjunct. He might also be unwilling to divulge which of p and
q is true (while still communicating that one of them is true). We can account for
this, too.
Let us consider an alternative addressee belief state B1Ad in which the addressee
is certain that the speaker believes either p or q, but not both, i.e.,
(9.41)
BAd Sp pp ^ qq ^ pSp ppq _ Sp pqqq
However, Ad does not yet have a belief as to which of p and q the speaker believes,
and he takes it to be possible that Sp has the additional preference
Inform p/q,
which is satisfied if the addressee believes neither p nor q:
Inform p/q
(9.42)
λv rv Ad p ^ Ad qqs
If the addressee is uncertain about (9.42), then there are four relevantly different
worlds in B1Ad (recall that we assume (9.41)):
vp^epp
Infq
vp
vq^epp
Infq
vq
Sp p
Sp q
epp Inform p/qq
True
False
True
Fig. 9.13
True
False
False
Fig. 9.14
False
True
True
False
True
False
Figures 9.13 and 9.14 shows the outcome of Opt in worlds in which the speaker
believes that p (the other two cases are symmetric). (9.43) summarizes the decision
in the four kinds of worlds:
(9.43)
“ tutterpp _ qqu
a.
Optpvp^epp
b.
Optpvp q “ tutterppqu
c.
Optpvq^epp
d.
Optpvq q “ tutterpqqu
Infq q
Infq q
“ tutterpp _ qqu
CHAPTER 9. CONVERSATIONAL IMPLICATURES
vp^epp Infq Quality
K
utterppq
utterpqq
*
+ utterpp _ qq
270
Inform p/q Inform p _ q Minimize
*
*
*
*
***
Figure 9.13: Decision in worlds where the speaker believes p and has a preference
against revealing that
vp
Quality Inform p _ q
K
*
+
utterppq
utterpqq
*
utterpp _ qq
Minimize
*
*
***
Figure 9.14: Decision in worlds where the speaker believes p and does not have a
preference against revealing that
But then, updating B1Ad with utterpp _ qq will only leave worlds in which the speaker
has the preference Inform p/q. Thus we predict that if the addressee believes that
the speaker knows which of p and q is true, but is uncertain whether the speaker
is willing to share this information, observing an utterance of p _ q will let him
conclude that the speaker prefers to withhold this information.
The general implicature
Of course, it may also be that the addressee’s information state is a combination of
the two we have just looked at. In that case, the addressee will remain uncertain as
to whether the speaker does not know, or does not want to reveal, which of p and
q is true. Again, this accords with intuition.
In general, as long as we assume that the preference Minimize is present, an
update with utterpp _ qq will only allow worlds v to survive such that
(i) in v, the speaker has a preference C;
(ii) in v, C is ranked higher then Minimize;
CHAPTER 9. CONVERSATIONAL IMPLICATURES
271
(iii) in v, Sp believes that utterppq violates C but utterpp _ qq does not.
That is, updating with utterpS, p _ qq removes all those worlds in which the speaker
does not have a reason for not uttering p (and symmetrically for q). In the two
cases we looked at, this defeating preference was either Quality (leading to the
ignorance implicature) or
Inform p/q (leading to an implicature that the speaker
wants to withhold information). Of course, in the general case, these need not be
the only two possible reasons. Here, as in the case of scalar implicatures in general,
any kind of preference could lead the speaker to avoid the stronger alternative. Just
as Grice’s original theory has been amended by assuming, for example, a set of
politeness maxims (Leech 1983, Brown and Levinson 1987), so in the present system
we can assume speaker preferences that are motivated by politeness considerations.
In this case, the reason an addressee may infer that the speaker used a disjunction
for reasons of politeness, even though the speaker believes one of the disjuncts to
be true and wishes the addressee believed the same. This is plausible for utterances
of sentences like (9.44).
(9.44)
9.3.4
Either you made a mistake or I did.
Generalizating NaR
We can now generalize the conditions under which NaR implicatures arise. Before
doing so, I want to briefly note that the prediction of mandatory implicatures does
not hinge on the particular choice of decision procedure made here. Indeed, any
optimization-based analysis of implicatures potentially predicts such implicatures.
Here is why: As an analysis of implicatures, such a theory should allow us to
characterize, for a given expression S and implicature i, the set CS{i of contexts
in which the implicature arises. At the same time, being an optimization-based
theory, it allows us to characterize the set COptpSq of contexts in which S is optimal.
Further, any such theory involves the assumption, on some level, that speakers
utter S only in contexts that are in COptpSq .34 But then, the theory will predict i to be
34
At the very least, it seems, any such theory must attribute the belief that speakers only make
CHAPTER 9. CONVERSATIONAL IMPLICATURES
272
a mandatory implicature of S just in case (9.45) holds.
(9.45)
COptpSq Ď CS{i
That is, unless this situation is somehow excluded, any optimization-based theory
has the potential to predict mandatory implicatures. Further, any optimizationbased theory that assumes a general ceteris paribus preference for shorter or simpler forms will predict the NaR implicature of disjunction (provided that disjunctions count as longer or more complex in the relevant sense).
This is true, in particular, of the optimality and game-theoretic approaches
referenced in the introduction to this chapter. These generally predict mandatory
implicatures, though if their authors point out this fact, they seem to view it as
a problem (perhaps due to the pervasive belief that all implicatures should be
optional and cancelable).
As a concrete example, consider the game-theoretic account of Franke (2009,
2011). Franke (2011) presents the account in two stages. In the first stage, his gamemodels encode the assumption that the speaker is perfectly informed about the state
of the world. As Franke discusses (p. 50–51), in these ‘base-level’ games, his analysis predicts that p_q is not used.35 It is only in the second stage, where he introduces
‘epistemically-lifted’ game models that can represent speaker-uncertainty, that his
model predicts p _ q to be used, and to implicate uncertainty.36 Of course, I think
that this prediction is perfectly adequate It captures the fact that the ‘ignorance’
implicature is mandatory.
In order to predict a mandatory implicature on any theory, including the one
‘optimal’ utterances to addressees, even if it allows for speakers to make non-‘optimal’ utterances.
It is conceivable to have an optimization-based theory where this belief itself is an assumption that
is only in place in some contexts—but then, this hypothesized theory of implicature really is only
a theory of implicatures for contexts in which this assumption is in place. And for such contexts, the
theory potentially predicts mandatory implicatures.
35
Technically, this prediction arises only if p ^ q is in the set of alternatives—without this alternative, his base-model in fact predicts, problematically, that p _ q is interpreted as p ^ q.
36
Franke limits attention to situation in which (a) the addressee needs to know the state of the
world exactly and (b) the preferences of the speaker and the hearer are perfectly aligned and (c)
the preferences are mutually known. Presumably, if this assumption is lifted, Franke’s also could
predict alternative NaR implicatures, e.g., that the speaker prefers to withhold information.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
273
used in the present chapter, a number of additional prerequisites need to be in
place. We can generalize from the case of disjunction as follows:
(9.46)
An expression e will give rise to a NaR implicature if:
a.
There is alternative expression e1 which is informationally stronger
than e.
e1 need not logically entail e, as it does in the case of disjunction. It
can also be informationally stronger because of entrenched worldknowledge, for example.
b.
e1 is salient whenever e is uttered.
In the case of disjunction, e1 is actually a constituent of e, but this need
not be the case, of course: All that is necessary is that observing an
utterance of e makes e1 salient.
c.
There is a linguistic preference for uttering e1 rather than e, all else
being equal.
In the case of disjunction, this preference plausible has to do with
length or complexity, but this is not a necessity: Any preference
between linguistic expressions will do.
For the implicature to be truly mandatory, of course, the preference in (9.46c) must
be assumed to be universally present. It is here where an account of implicatures
that relies on maxims of conversation and a cooperative principle (instead of a more
general notion of a preference) may make different predictions. NaR implicatures
are quite similar to classical implicatures that arise on the basis of the Maxim
of Manner, which after all also involves prescriptions about how things should
be expressed. But the Gricean maxims, as classically understood, depend on
the assumption of cooperativity. And indeed, that is why Grice thought that all
implicatures must be cancelable. The quote I gave initially elided his reasoning:
“Since, to assume the presence of a conversational implicature, we have
to assume that at least the Cooperative Principle is being observed,
and since it is possible to opt out of the observation of this principle,
CHAPTER 9. CONVERSATIONAL IMPLICATURES
274
it follows that a [. . . ] conversational implicature can be canceled in a
particular case.”
(Grice 1975)
In a preference-based analysis of implicatures, we can construe the preference in
(9.46c) instead as a selfish preference, which is in place even in cases in which the
speaker cannot be taken to be cooperative. In this case, we predict truly mandatory
implicatures. Without such an assumption of a selfish preference, we still predict
near-mandatory implicatures, which arise in all contexts in which it is a given that
the speaker is cooperative.
9.3.5
More NaR implicatures?
As pointed out in Section 9.3.1, the existence of NaR implicatures challenges widelyheld pragmatic orthodoxy. Optionality and cancelability have long been used
as a test for implicature-hood: In order to determine whether a given observed
implication of a sentence is an implicature or not, it is checked whether the sentence
can be uttered and, at the same time, the implication can be coherently denied. If
this is not possible, it is concluded that the implication cannot be a conversational
implicature, but must instead be semantic in kind. Indeed, Sadock (1978) called
this ‘the best of the tests’ for implicature-hood. Once we realize that conversational
implicatures can be mandatory, that means that using cancelability as a test is
problematic: If it is used at all, it has to be used with great care.37
On the plus side, this means that phenomena that may otherwise have seemed
outside of the reach of pragmatic theory can possibly be accounted for in pragmatic
terms after all, with the usual benefit of simplifying and streamlining semantic
analyses. In this section, I want to briefly mention three possible cases that have
37
Hirschberg (1985)’s move to require cancelability as a property of implicatures by definition, of
course, cures the symptom but not the illness: We can then use cancelability as a test for implicaturehood, but if a given implication is not cancelable, we still cannot conclude that it must be semantic
in nature—because it could still be a non-implicature pragmatic implication.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
275
been discussed in the literature in which NaR reasoning, or something close to it,
might be at play.
Huitink and Spenader (2004)’s cancelation-resistant implicatures
The first case involves implicatures that are not mandatory in the sense that NaR
implicatures are mandatory, but I want to discuss them for two reasons. On the
one hand, these have been described as involving failures of cancelability, and I
want to highlight the difference between the two kinds of non-cancelability. On
the other hand, I think that NaR-style reasoning may be a better way to account for
the robustness of these implicatures than the analysis that has been proposed.
Huitink and Spenader (2004) discuss cases were classical examples of implicatures are very difficult to cancel (Weiner (2006) discusses a similar case):
(9.47)
[In a letter of reference for a philosopher]
Mr. X’s command of English is excellent and his attendance at tutorials
has been regular. He is a brilliant philosopher.
(9.48)
Miss X produced a series of sounds that corresponded closely with the
score of “Home Sweet Home”. She has a beautiful voice.
The first sentence sentence in each of these examples is a classical example from
Grice (1975): The first sentence in (9.47) is taken to implicate that Mr. X is no good
in philosophy, while the first sentence in (9.48) is taken to implicate that Miss X sang
badly. The second sentence in each case is the negation of the putative implicature.
As Huitink and Spenader note, it is actually quite difficult to take these sentences
as cancelations: The much more natural interpretation is to re-interpret them as
ironic statements and maintain the implicature.
Now, of course, even if it were impossible to interpret these examples as cancelations, this would not show much: At best, it shows that there are implicatures
that cannot be canceled in every context in which they arise—obviously Mr X’s
command of English is excellent does not implicate Mr X is a bad philosopher in
every context in which it is used. The NaR implicature of disjunction is mandatory
CHAPTER 9. CONVERSATIONAL IMPLICATURES
276
in a much stronger sense: There is no context where it is absent or can be canceled.
Huitink and Spenader propose that the implicatures in these cases are ‘semiconventionalized’, by which they mean that there is a conventional link between
a certain type of context and the implicature. While something like that might be
plausible for (9.47), I find it hard to see how the same could be true for (9.48). In any
case, NaR reasoning—of a more contextual sort than in the case of disjunction—
can potentially help us to understand why the implicature is difficult to cancel in
these contexts: In both cases (and the other cases Huitink and Spenader discuss),
it is quite difficult to see why a speaker would say or write the first sentence if he
believes the second.
Romero and Han (2004): High-negation polar questions
Romero and Han (2004) develop an an account of high-negation polar questions
(HNPQs), such as (9.49).
(9.49)
Doesn’t John drink?
{ The speaker believes or at least expects that John drinks.
HNPQs are an intricate topic which I do not wish to discuss here. I mention the
issue because Romero and Han claim that the implication in (9.49) is an implicature,
despite being ‘strong and uncancelable’—and they aim to account for it in a way
that is very close to the NaR reasoning we have seen so far.
Han and Romero propose that HNPQs contain a semantic operator that turns
the question into a ‘meta-conversational move’ (essentially, a move having to do
with common-ground management). Then they propose the following pragmatic
principle (p. 629):
(9.50)
Principle of Economy:
Do not use a meta-conversational move unless necessary (to resolve epistemic conflict or to ensure Quality).
CHAPTER 9. CONVERSATIONAL IMPLICATURES
277
It should be obvious how similar reasoning on the basis of such a principle is to
NaR reasoning. Essentially, the principle says that a speaker must have a reason
for making a ‘meta-conversational move’. The similarity is not merely superficial:
Romero and Han take (9.50) to be an inviolable pragmatic principle (for otherwise,
it could not explain why the epistemic implicature in (9.49) is uncancelable), but in
my terms, we can reconstruct it by assuming a ceteris paribus preference against
making ‘meta-conversational moves’. So, Romero and Han (2004)’s analysis of
HNPQs fits very naturally into the general NaR picture.
Infelicity via implicature
The final case I want to discuss is perhaps the most promising one: NaR implicatures can be used to explain why certain sentences are infelicitous: If a sentence S
that has a NaR implicature i is uttered in a context in which i is known to be false,
it will be perceived infelicitous, because the addressee cannot make sense of the
speaker’s utterance choice.
This is impossible with implicatures that are optional: If S only optionally
implicates i, and S is used in a context in which i is known to be false, then i should
simply be absent, instead of rendering the utterance infelicitous.
For this reason, a set of phenomena have recently been argued to involve grammaticalized scalar inferences (e.g., á la Chierchia, Fox and Spector (2012)),38 on the
grounds that the expressions triggering these inferences are infelicitous if the inference is known to be false (Magri 2009, Magri 2011, Ivlieva 2012). The argument
has the general form in (9.51).
(9.51)
a.
Sentences s of a given type are infelicitous, often to the point of
appearing ungrammatical.
b.
Comparison with similar (felicitous) sentences suggests that s triggers
a scalar inference i.
38
These authors maintain the use of the term ‘implicature’ for the grammaticalized inferences
they hypothesize. I refer to them as ‘grammaticalized scalar inferences’, such inferences do not fit
the usual conception of a scalar implicature.
CHAPTER 9. CONVERSATIONAL IMPLICATURES
c.
278
Based on logical entailment and/or entrenched world-knowledge,
speakers know that i is incompatible with the truth conditions of
s.
d.
The infelicity is hence due to the fact that the utterance of s triggers
an inference that is known to be false if s is true.
e.
Because Gricean implicatures are always optional, i cannot be a
Gricean implicature. Instead, i must be an inference arising as part of
the semantic meaning of s.
We can accept steps (a-d) but deny (e) if i is a NaR implicature. The observed
infelicity can then be explained in terms of pragmatic theory alone.
To illustrate: Magri (2009) examines individual-level predictes (i-predictes),
which can be characterized as predicates that denote permanent properties. They
contrast with stage-level predicates (s-predicates), which denote temporary properties. Since i-predicates behave differently from s-predicates in a number of ways,
it had been proposed (e.g., by Kratzer (1995), Chierchia (1995)) that the two kinds
of predicates differ in terms of their grammatical properties. Magri proposes that
both kinds of predicates are grammatically the same, and aims to explain the observed differences in terms of obligatory scalar inferences. One such difference is
that i-predicates are generally infelicitous (or ‘odd’) with various kinds of temporal
modification, for example, temporal frame adverbials such as this month:
(9.52)
(9.53)
a.
John is of noble birth.
i-predicate
b.
John is in in financial trouble.
s-predicate
a. #John is of noble birth this month.
i-predicate
b.
s-predicate
John is in financial trouble this month.
Magri’s explanation for the infelicity of (9.53a) goes roughly as follows: The sentences in (9.53) give rise to the obligatory scalar inferences in (9.54). However,
due to world knowledge, it will generally be common ground that if John is of
noble birth this month, he is always of noble birth. But then, the scalar inference
(9.54a) triggered by (9.53a) is contextually incompatible with the asserted content
CHAPTER 9. CONVERSATIONAL IMPLICATURES
279
of the sentence. The utterance gives rise to contradictory implications, and hence
is infelicitous.
(9.54)
a.
John is not of noble birth at times before or after this month.
b.
John is not in financial trouble at times before or after this month.
For this explanation to work, the scalar inference cannot be optional: If it were,
(9.53a) should be felicitous, because the offending inference should simply be absent. Magri concludes that the scalar inference hence cannot be a Gricean implicature, and instead proposes to derive it within Chierchia et al. (2012)’s grammatical
theory of implicature.
Magri’s account is attractive, as it obviates the need to represent the distinction
between i- and s-predicates in the grammar. The existence of Gricean NaR implicatures has the potential of simplifying the explanation even further, predicting the
observed contrasts based on a uniform semantic representation and pragmatic reasoning. The idea, essentially, is this: Given entrenched the world-knowledge that
‘being of noble birth’ is a permanent property, the (longer, hence dispreferred) sentence John is of noble birth this month is informationally equivalent to the (shorter,
hence preferred) sentence John is of noble birth. But then, there cannot be a reason to utter the dispreferred form instead of the preferred one. Consequently, the
utterance of the dispreferred form is perceived as odd.
Of course, for this explanation to be convincing, we have to establish that the
prerequisites for NaR implicatures are met. We have to establish that John is of
noble birth is preferred (due to length, complexity, or other considerations) to John
is of noble birth this month, and that any utterance of the latter sentence makes
the former sentence salient as an alternative. But both of these seem to be sensible
assumptions, at least at first glance.
It is worth stressing what the success of such an account would mean: It would
mean that we can explain why a given expression is never felicitous (at least as
we maintain the required world-knowledge assumptions), on entirely Gricean
grounds. Such explanations have long been thought to be beyond the reach of
Gricean theory, which was construed, after all, as a theory concerned with weak,
CHAPTER 9. CONVERSATIONAL IMPLICATURES
280
defeasible, cancelable inferences. NaR reasoning shows that this conception is
incorrect, and it flows from the general conviction that formed the starting point
of this dissertation: Gricean reasoning is operational on every occasion of language
use.
Chapter 10
Outlook
This dissertation has developed a formal framework for pragmatics that is faithful
to the Gricean conception of pragmatic inference as interlocutors’ reasoning about
utterance choice. It has emphasized the usefulness of such a Gricean perspective for
formal semantics and pragmatics beyond its classical field of application, i.e., the
study of non-entailed inferences. The framework incorporates an articulated theory
of sentential force, and allows fine-grained modeling of the interaction between
semantic content, force, interactional reasoning and contextual assumptions.
In the present chapter, I want to reprise some of the open questions that were
raised, discuss some of the shortcomings of the current framework and simplifications that were made, and make some brief remarks about how these issues could
be addressed.
10.1
The role of intentions in pragmatic theory
The framework of dynamic pragmatics does not represent intentions. As I pointed
out in Chapter 1, this may seem surprising for a pragmatic theory that calls itself
‘Gricean’, considering the central role Grice saw for intentions in determining
speaker meaning.
281
CHAPTER 10. OUTLOOK
282
Intentions vs. effective preferences
One reason for this choice was that ‘intention’, at least as it is understood in
philosophy, is a rather intricate and complicated concept. I have opted to use a
novel technical term, ‘effective preferences’, to talk about non-epistemic attitudes
that influence action choice. Looking at things this way, the concept of effective
preference may be viewed as ‘intention lite’.
However, it is not clear that everything that I want to consider an effective
preference—a preference that shapes the speaker’s action choice—can sensibly be
called an intention. One example are ‘linguistic preferences’ such as the preference
for shorter, simpler forms that has played a considerable role in Chapter 9. While
it makes sense to assume that the utterance choices of speakers are affected by such
a preference, it is not quite clear if it makes sense to call it an intention. I therefore
think it is preferable to view of intentions as just one of the various preferential
attitudes that can figure as effective preferences of an agent.1
The relevance of intentions
At the same time, a number of the analyses presented in the course of this dissertation did not even assume an effective preference in cases where intentions are
commonly assumed to play a crucial role. The analysis of how addressees come
to believe in the truth of the content (Chapters 2 and 5) did not assume that (the
addressee believes that) the speaker has an effective preference to communicate the
content of his utterance. Similarly, according to the analysis presented in Chapter 7,
it does not matter whether a speaker who utters I hereby promise to be there has
an effective preference for committing himself (or for being there): His utterance
will be a promise (and commit him) regardless. Finally, in Chapter 9, I showed
how an implicature can arise without assuming that (the addressee believes that)
the speaker effectively prefers to communicate it.
1
Though perhaps, intentions are given a greater weight than other such attitudes. This would go
some way to explain the ‘special role’ intentions (as opposed to mere desires) play in our decision
making according to Bratman (1987), as well as some of the properties Bratman requires intentions
to have—viz., that they be consistent.
CHAPTER 10. OUTLOOK
283
In each of these cases, of course, a speaker could have such an effective preference, and, in certain context, it may be relevant whether he does and whether the
addressee recognizes that he does. But the basic phenomena seem to be the same
regardless of whether such an effective preference is present.
Perhaps the investigation of other phenomena will reveal that intentions play a
more central role in them. In the meantime, I take it to be an open question whether
intentions, and their recognition, figure in a central way in our understanding of
everyday language use.
10.2
Ambiguity and underspecification
At the very outset of this dissertation, Chapter 2, I made a radically idealizing
assumption. I assumed that agents observe each other’s utterances perfectly, and
went on to ignore all kinds of ambiguity and underspecification, essentially assuming (in modeling speakers of a simple propositional language) that agents
utter disambiguated logical forms, with all contextual parameters filled in.
Ultimately, we want to lift this assumption. This is not only because we know
that, as a matter of fact, natural languages are rife with underspecification and
ambiguity (at multiple levels of linguistic description). It also is likely that in
resolving such ambiguity and underspecification, agents employ the same kind
of Gricean reasoning as I have described in the course of this dissertation. Two
options for treating ambiguity and underspecification in the current framework
naturally suggest themselves:
Ambiguous utterance events.
The first option is to change the nature of the
utterance events we have assumed, but hold on to the assumption that utterance
events are perfectly observed. Instead of having events utterpi, i1 , ϕq, where ϕ is
mapped to a unique interpretation (e.g., a set of possible worlds), we would have
utterpi, i1 , Sq, where S is mapped to a set of interpretations (e.g., a set of sets of
possible worlds).
This would require an adjustment in our conventions of use. Recall that the
CHAPTER 10. OUTLOOK
284
utterances of many sentences create speaker commitments to beliefs or preferences.
We would have to specify what it means to be committed to a belief in a possiblyambiguous sentence, rather than a proposition.
Imperfect perception of utterance events. The second option is to maintain that
utterance events are individuated by the disambiguated logical form, but give
up the assumption that events are perfectly perceived. For a sentence S that is
ambiguous between two readings ϕ1 and ϕ2 , we would have two utterance types:
(10.1)
a.
utterpi, i1 , S, ϕ1 q
b.
utterpi, i1 , S, ϕ2 q
Then we have to replace the PAL constraint to reflect that, if (10.1a) happens, a
hearer learns that either (10.1a) or (10.1b) happened.
A promising option for finding a suitable new constraint on belief change is to
turn to generalizations of the models for Public Announcement Logic that inspired
our PAL constraint. In particular, the ‘event models’ of Baltag, Moss and Solecki
(1998)2 allow for just this kind of selective indistinguishability of events. van
Benthem (2011, Chapter 4) gives an accessible introduction to the basic framework,
which has been employed to interpret various systems of multi-agent Dynamic
Epistemic Logic (DEL).
Transition Preference Pragmatics. Either of these options, if implemented appropriately, would lead to a model that is quite close to a combination of the dynamic
pragmatics developed here and the Transition Preference Pragmatics (TPP) of
Beaver (2002).
Beaver employs a dynamic system in which updates can be indeterminate, and
hence an addressee may need to employ pragmatic reasoning in order to identify the correct transition between contexts. Beaver himself employs this system
2
Baltag et al. (1998) called these models ‘action models’, but in the more recent literature, the
name ‘event models’ has become standard.
CHAPTER 10. OUTLOOK
285
mainly for modeling ambiguity and underspecification, but it has also been employed by Davis (2011) to model the effect of sentence-final particles and intonation
in Japanese—both devices that affect the conventional effect of the uttered sentence.
As Davis otherwise uses a conception of the basic effects of the main clause types
that is very similar to the one in this dissertation, integrating something like TPP
into the current framework offers the possibility of unifying his theoretical perspective with the one taken in this dissertation.
10.3
A question of commitment
Any treatment of ambiguity and underspecification in the current framework raises
a theoretical question: How does the commitment induced by an ambiguous utterance get determined? If we allow for ambiguous utterance events, this question
manifests itself in the issue of how the conventions of use should be adjusted to
take into account that utterance events do not determine a single interpretation. If
instead we maintain that utterance events determine their contents (but allow for
hearer-uncertainty about which utterance happened), the question manifests itself
in the issue of how the event that actually happens in a world is determined.
In either case, the question comes down to this: When a speaker utters (10.2),
what determines whether he becomes committed to the belief that Cleo is German,
or the belief that Hanne is German, or . . . ?
(10.2)
She is German.
In Gricean terms, this question amounts to how ‘what is said’ gets determined—in
this example, what determines who she refers to? The classical answer involves
appeal to speaker intentions: She refers to whichever individual the speaker intended to refer to. Perhaps intentions will make their great entrance here; perhaps
this is where intentions are crucial in the current view of things. They determine
what the speaker refers to with an underspecified referential expression, and how
ambiguities get resolved. Thereby, intentions determine what the speaker gets
CHAPTER 10. OUTLOOK
286
committed to by his utterance.
But once again, I am sceptical that it is a speaker’s private intentions that are the
deciding factor. Conceptually, private intentions just do not mesh with the very
public nature of commitments. A speaker who utters (10.2) cannot freely choose
what he gets committed to by having a private intention to refer to some arbitrary
(female) individual, without publicizing this intention somehow, and he cannot
avoid becoming committed at all by secretly failing to have a referential intention.
I leave this as an open question. The current framework brings out sharply
that this is a question, and one that is distinct, in principle, from the question
how audiences figure out what the speaker wanted to communicate. The question
What did the speaker really, truly refer to? may seem like an idle one in cases in
which the speaker had a certain interpretation in mind, and the addressee correctly
identified this interpretation—but in the current set-up, this question has a very
practical bite to it: The answer determines what the speaker becomes committed
to, and hence, for example, what kind of future utterances the speaker can make
without retraction.
10.4
Belief revision, salience, and awareness
Throughout this thesis, I have often made rather strong assumptions about agents’
beliefs. For example, I have occasionally assumed that an addressee believes,
with certainty, that the speaker has a certain effective preference. This is often
implausible in practice, which makes it tempting to add a representation of graded
belief (e.g., subjective probability) to the model. This is a possibility, but there
are also a number of related issues whose solution may solve the problem of
unreasonably strong assumptions about belief at the same time:
There is a tension between two properties of the framework as introduced in
this dissertation: (i) The framework does not properly model belief revision, as
such revision is essentially unconstrained; (ii) the framework does not incorporate
a treatment of salience, attention, or limited awareness of possibilities.
It is defensible to ignore the issue of belief revision, if one conceptualizes the
CHAPTER 10. OUTLOOK
287
modeled beliefs in an appropriate way. We can say that the i -operator models
very strongly-held beliefs, which agents form cautiously and rarely have occasion
to give up. Then we can maintain that revision of these beliefs happens so rarely
that we can often ignore the issue in practice.
Conversely, we can make do without an explicit model of salience or limited
awareness if we assume that beliefs are very context-dependent. To model a
situation in which an agent only takes salient possibilities into account, we simply
can assume that the speaker believes that the salient possibilities are the only ones.
But if we do this, belief revision is a common occurrence, and we need to model it
faithfully.
One way to resolve this tension is to extend the current system in such a way
that it can account for agents being unaware of possibilities which their beliefs, in
principle, do not exclude. A particularly promising approach is the model of Franke
and de Jager (2011).3 They employ a standard Kripke model as a specification of
‘background beliefs’, which gets ‘filtered’ through a state of limited awareness.
The result of the filtering operation is again a standard Kripke model, but one in
which certain possibilities are ignored. Modalities and decision procedures are
then defined in the usual way on this ‘foreground’ model.
This kind of approach is promising for two reasons: Firstly, it means that we
can deal with issues such as salience without changing the basic set-up of our
models—we simply add awareness states and specify their dynamics. Secondly,
explicitly modeling (lack of) awareness to possibilities is a very intuitive way to
model the effect of utterances that do not actually provide novel information, but
instead serve as reminders that make a certain possibility salient, as in Franke and
de Jager’s (10.3), and similar examples that they discuss.
(10.3)
[A is looking for his keys.]
B: Could they be in the car?
3
As Franke and de Jager (2011, Section 5), discuss, their model is inspired by models of ‘unawareness’ in computer science and rational choice theory (e.g., Fagin and Halpern (1988), Feinberg
(2004)), but differs in the details, in part because Franke and de Jager are mainly interested in a kind
of unawareness that is easily overturned—a property that is also desirable for our purposes here.
CHAPTER 10. OUTLOOK
288
So amending the model with a treatment of unawareness (or similar models of
‘default’ beliefs, such as Veltman’s (1996) treatment of defaults in update semantics)
will allow us to apply the current framework to more phenomena. At the same
time, doing so could remedy, in one fell swoop, the two shortcomings of the system
mentioned above: The lack of a proper model of belief revision becomes defensible,
as we can conceptualize the ‘background’ beliefs as ‘cautious’ beliefs that rarely
need to be revised. At the same time, we can model instances in which only a
limited range of possibilities is taken into account (e.g., the assumptions about
utterance alternatives discussed in Section 9.1.2), without making unreasonable
assumptions about what the agents believe to be possible.
10.5
Conclusion
I have only scratched the surface of the range of phenomena that can be illuminated
by systematically taking a Gricean perspective in the way the system of dynamic
pragmatics does. In addition to putting forth particular analyses, and independently from the particulars of the formal framework, I hope this dissertation serves
as an advertisement for this kind of perspective on language use. Thinking about
language use—every occasion of language use—as an instance of purposive human
behavior sheds new light on old questions, makes us see new generalizations, and
raises new questions.
Appendix A
The basic system
This appendix contains the basic definitions for the system set up in Chapter 2.
Definition 1 (Prop syntax). Given a set of proposition letters P, the language Prop is the
smallest set such that:
(i) P Ď Prop.
(ii) If ϕ P Prop then so is
ϕ.
(iii) If ϕ1 , ϕ2 P Prop, then so are pϕ1 ^ ϕ2 q, pϕ1 _ ϕ2 q, pϕ1 Ñ ϕ2 q, pϕ1 Ð ϕ2 q and
pϕ1 Ø ϕ2 q.
Definition 2 (Prop interpretation). Given an a set of worlds W and an interpretation
function I : P ÞÑ ℘pWq, v¨w
• For p P P : vpw
Prop
Prop
is defined as:
“ Ippq.
0
8Prop 0 8Prop
Prop
• pφ ^ ψq
“ φ
X vψw .
•
0
8Prop
0 8Prop
“ Wz φ
.
φ
For other formulas, v¨w
Prop
is defined using the usual equivalences.
Definition 3 (Pragmatic Language: Syntax). For a given object language Prop, we
define PProp (the pragmatic language for Prop) as follows: Let
289
APPENDIX A. THE BASIC SYSTEM
290
(a) Ic , Tc , Ec be sets of constant symbols (the ‘individual constants’, the ‘temporal constants’
and the ‘event constants’, respectively).
(b) Iv , Tv , Ev be sets of variable symbols (the ‘individual variables’, the ‘temporal variables’
and the ‘event constants’, respectively).
(c) P be a set of predicate symbols and ar a function P ÞÑ N.
such that Ic , Tc , Iv , Tv , Ec , Ev , P are pairwise disjoint. Then PProp is the smallest set such
that:
1. If ϕ P Prop, then ϕ P PProp .
2. If φ, ψ P PProp , so are
φ, pφ ^ ψq, pφ _ ψq, (φ Ñ ψq, pφ Ð ψq and pφ Ø ψq.
3. If φ P PProp and i P It Y Iv , t P Tc Y Tv , then i,t φ P PProp .
4. If φ P PProp and t P Tc Y Tv , then Ct φ P PProp .
5. If φ P PProp and x P Iv Y Tv Y Ev , then Dx : φ P PProp .
6. If P P P and arpPq “ n and a1 , . . . , an P Ic Y Iv Y L and t P Tc Y Tv , then
Pt pa1 , . . . , an q P PProp .
7. If ϕ P L, i, i1 P I, e P Ec Y Ev , then
a. uttere pi, i1 , pϕqq P PProp .
b. uttere pi, pϕqq P PProp .
c. uttere piq P PProp .
Definition 4 (T ˆ W frames). A T ˆ W frame is a quadruple hW, T, ă, «i, where W and
T are non-empty sets, ă is a transitive relation on T which is also irreflexive and linear,
and « is a 3-place relation on T ˆ W ˆ W, such that (1) for all t, «t is an equivalence
relation and (2) for all w1 , w2 P W and t, t1 P T, if w1 «t w2 and t1 ă t then w1 «t1 w2 .
(Thomason 1984, Definition 6, p. 216)
APPENDIX A. THE BASIC SYSTEM
291
Definition 5 (N ˆ W frames). A N ˆ W frame is a T ˆ W frame hW, T, ă, «i where
T “ N and ă has its usual interpretation.
Definition 6 (PProp frames). A PProp frame is a W, «, Ind, Ag, R , where hW, «i is a
N ˆ W frame, Ind a domain of individuals, Ag Ď Ind a set of agents, and R a function
Ag ˆ N ÞÑ ℘pW ˆ Wq that takes agent, time pairs into relations between worlds, such that
• Ri,t is transitive, serial, and Euclidean for all i, t;
• historicity: If w1 «t w2 , then w1 Ri,t v iff w2 Ri,t v.
• no fore-belief: If v1 «t v2 , then wRi,t v1 iff wRi,t v2 .
Definition 7 (Transitive Closure). For a given PProp -frame W, «, Ind, Ag, R , let R :
N ÞÑ W ˆ W be the function such that for every t P N, Rptq is the smallest relation such
that:
(i)
Ť
iPInd
Ri,t Ď Rptq.
(ii) Rptq is transitive.
Definition 8 (Event types). Let E be a set of event classes, Ind a set of individuals, and
L a language. Then the set of event types based on E, Ind, L is defined as:
EvpE, Ind, Lq :“ tPpi1 , . . . , in q | P P E & i1 , . . . in P Ind Y Lu
Definition 9 (PProp models). Given a language L, a model M for PProp is a tuple
hF , I, E, Hapi where
(i) F “ W, «, Ind, Ag, R is a PProp frame
(ii) I is an interpretation function such that
a. Iptq P N if t P Tc .
b. Ipiq P Ind if i P Ic .
c. Ipeq P T ˆ T if e P Ec .
APPENDIX A. THE BASIC SYSTEM
292
d. IpPq P W ÞÑ T ÞÑ pInd Y Propqn if P P P and arpPq “ n.
e. Ippq P ℘pWq if p is a proposition letter of Prop.
(iii) E is a set of event classes such that utter P E.
(iv) Hap : W ˆ pT ˆ Tq ÞÑ EvpE, Ind, Lq such that
a. for any w, Hapw pt, t1 q is defined only if t1 “ t ` 1.
b. if w1 «t w2 then for all t1 , t2 ď t : Hapw1 pt1 , t2 q “ Hapw2 pt1 , t2 q (historicity).
c. if w1 «t w2 and Hapw1 pt, t ` 1q “ Hapw2 pt, t ` 1q then w1 «t`1 w2 (determinism).
Additionally, we require that I respects «, that is that for P P P: IpPqpw, tq “ IpPqpw1 , tq
if w «t w1 .
Definition 10 (PProp satisfaction). For a given language Prop, given a
PProp -model W, «, Ind, Ag, R , I, E, Hap , and an assignment g, is the following relation between W and PProp (where I g is the function such that I g pxq “ gpxq if x is a variable,
I g pxq “ Ipxq otherwise):
Prop
1. w g ϕ if ϕ P Prop and w P vϕw .
D
E
2. w g Pt pa1 , . . . , an q iff I g pa1 q, . . . , I g pan q P IpwqpI g ptqqpPq.
3. w g i,t φ if for all v : wRIg piq,Ig ptq v : v g φ.
4. w g Ct φ if for all v : wRpI g ptqqv : v g φ.
5. w g Dx : φ if there is d P Ind Y T Y pT ˆ Tq : w grx{ds φ.
6. a. w g uttere pa, b, pϕqq iff Hapw pI g peqq “ utterpI g paq, I g pbq, ϕq.
b. w g uttere pa, pϕqq iff there is i P Ind : Hapw pIpeqq “ utterpI g paq, i, ϕq.
c. w g uttere paq iff there are i P Ind, ϕ P L : Hapw pI g peqq “ utterpI g paq, i, ϕq.
7. Propositional connectives are interpreted in the usual way.
Bibliography
Ajdukiewicz, K.: 1938, Logiczne Podstawy Nauczania [The logical foundation of teaching], Nasza Ksiȩgarnia, Warsaw/Vilnius.
Ariel, M.: 2004, Most, Language 80(4), 658–706.
Atlas, J. and Levinson, S.: 1981, It-clefts, informativeness and logical form, in P. Cole
(ed.), Radical Pragmatics, Academic Press, New York, NY, pp. 1–61.
Aumann, R.: 1995, Backward induction and common knowledge of rationality,
Games and Economic Behavior 8(1), 6–19.
Aumann, R.: 1996, Reply to Binmore, Games and Economic Behavior 17(1), 138–146.
Austin, J. L.: 1962, How to do things with words, Harvard University Press, Cambridge, MA.
Bach, K.: 2006, The top ten misconceptions about implicature, in B. Birner and
G. Ward (eds), Drawing the Boundaries of Meaning: Neo-Gricean Studies in Pragmatics and Semantics in Honor of Laurence R. Horn, John Benjamins, Amsterdam/New York, NY, pp. 21–30.
Bach, K.: 2012, Saying, meaning and implicating, in K. Allan and K. M. Jaszczolt
(eds), Cambridge Handbook of Pragmatics, Cambridge University Press, Cambridge, UK, pp. 47–68.
Bach, K. and Harnish, R. M.: 1979, Linguistic Communication and Speech Acts, MIT
Press, Cambridge, MA.
293
BIBLIOGRAPHY
294
Bach, K. and Harnish, R. M.: 1992, How performatives really work: A reply to
Searle, Linguistics and Philosophy 15(1), 93–110.
Baltag, A., Moss, L. S. and Solecki, S.: 1998, The logic of public announcements,
common knowledge, and private suspicions, in I. Gilboa (ed.), Proceedings of
Theoretical Aspects of Rationality and Knowledge (TARK) 7, Morgan Kaufmann
Publishers, San Francisco, CA, pp. 43–56.
Beaver, D.: 2001, Presupposition and Assertion in Dynamic Semantics, CSLI Publications, Stanford, CA.
Beaver, D. I.: 2002, Pragmatics, and that’s an order, in D. Barker-Plummer, D. I.
Beaver, J. van Benthem and P. S. di Luzi (eds), Words, Proofs and Diagrams,
CSLI Publications, Stanford, CA, pp. 191–216.
Belnap, N. and Perloff, M.: 1990, Seeing to it that: A canonical form for agentives, in
H. Kyburg, R. Loui and G. Carlson (eds), Knowledge representation and defeasible
reasoning, Vol. 5 of Studies in Cognitive Systems, Kluwer Academic Publishers,
Dordrecht, pp. 167–190.
Benz, A. and van Rooij, R.: 2007, Optimal answers, and what they implicate, Topoi
26(1), 63–78.
Bierwisch, M.: 1980, Semantic structure and illocutionary force, in J. R. Searle,
F. Kiefer and M. Bierwisch (eds), Speech Act Theory and Pragmatics, Reidel,
Dordrecht, pp. 1–35.
Blutner, R.: 1998, Lexical Pragmatics, Journal of Semantics 15(2), 115–162.
Blutner, R.: 2000, Some aspects of optimality in natural language interpretation,
Journal of Semantics 17(3), 189–216.
Blutner, R.: 2002, Lexical semantics and pragmatics, Linguistische Berichte, Sonderheft
10, 27–58.
Boghossian, P. A.: 1989, The rule-following considerations, Mind 98(392), 507–549.
BIBLIOGRAPHY
295
Brandom, R.: 1983, Asserting, Noûs 17(4), 637–650.
Bratman, M. E.: 1987, Intention, Plans, and Practical Reason, Harvard University
Press, Cambridge, MA.
Bresnan, J.: 2000, Optimal syntax, in J. Dekkers, F. van der Leeuw and J. van de
Weijer (eds), Optimality Theory: Phonology, Syntax and Acquisition, Oxford University Press, Oxford, UK, pp. 334–385.
Brown, P. and Levinson, S.: 1987, Politeness: Some universals in language usage,
Cambridge University Press, Cambridge, UK.
Castroviejo Miró, E.: 2008, Deconstructing exclamations, Catalan Journal of Linguistics 7, 41–90.
Chernilovskaya, A., Condoravdi, C. and Lauer, S.: 2012, On the discourse effects of
wh-exclamatives, in N. Arnett and R. Bennett (eds), Proceedings of the 30th West
Coast Conference on Formal Linguistics (WCCFL), Cascadilla Press, Somerville,
MA, pp. 109–119.
Chernilovskaya, A., Condoravdi, C. and Lauer, S.: in prep., The context change
effect of exclamatives. Manuscript, Stanford and Utrecht University.
Chernilovskaya, A. and Nouwen, R.: 2012, On wh-exclamatives and noteworthiness, in M. Aloni, F. Roelofsen, G. W. Sassoon, K. Schulz, N. Kimmelman
and M. Westera (eds), Proceedings of the Amsterdam Colloquium 2011, Springer,
Berlin.
Chierchia, G.: 1995, Individual-level predicates as inherent generics, in G. N. Carlson and F. J. Pelletier (eds), The generic book, University of Chicago Press,
Chicago, IL, pp. 125–175.
Chierchia, G., Fox, D. and Spector, B.: 2012, Scalar implicature as a grammatical
phenomenon, in C. Maienborn, K. Heusinger and P. Portner (eds), Semantics.
An International Handbook of Natural Language Meaning, Vol. 3, De Gruyter
BIBLIOGRAPHY
296
Mouton, pp. 2297–2332.
(Previous versions were circulated under the title The Grammatical View of
Scalar Implicatures and the Relationship between Semantics and Pragmatics).
Chierchia, G. and McConnel-Ginet, S.: 1990, Meaning and Grammar: An introduction
to semantics, MIT Press, Cambridge, MA.
Ciardelli, I. and Roelofsen, F.: 2011, Inquisitive logic, Journal of Philosophical Logic
40(1), 55–94.
Condoravdi, C.: 2002, Temporal interpretation of modals: Modals for the present
and for the past, in D. I. Beaver, S. Kaufmann, B. Clark and L. Casillas (eds),
The Construction of Meaning, CSLI Publications, Stanford, CA, pp. 59–88.
Condoravdi, C. and Lauer, S.: 2009, Performing a wish: Desiderative assertions
and performativity. Talk presented at California Universities Semantics and
Pragmatics (CUSP) 2, UC Santa Cruz, November 2009.
Condoravdi, C. and Lauer, S.: 2011, Performative verbs and performative acts, in
I. Reich, E. Horch and D. Pauly (eds), Sinn and Bedeutung 15: Proceedings of the
2010 Annual Conference of the Gesellschaft für Semantik, Universaar – Saarland
University Press, Saarbrücken, pp. 149–164.
Condoravdi, C. and Lauer, S.: 2012, Imperatives: Meaning and illocutionary force,
in C. Piñón (ed.), Empirical Issues in Syntax and Semantics 9, pp. 37–58.
Condoravdi, C. and Lauer, S.: to appear, Anankastic conditionals are just conditionals, Proceedings of Semantics and Linguistic Theory (SALT) 23, Cornell University,
CLC Publications, Ithaca, NY.
Davis, C.: 2009, Decisions, dynamics, and the Japanese particle yo, Journal of Semantics 26(4), 329–366.
Davis, C.: 2011, Constraining Interpretation: Sentence Final Particles in Japanese, PhD
thesis, University of Massachusetts at Amherst.
BIBLIOGRAPHY
297
Degen, J. and Franke, M.: 2012, Optimal reasoning about referential expressions, in
S. Brown-Schmidt, J. Ginzburg and S. Larsson (eds), Proceedings of SeineDial,
Université Paris–Diderot (Paris 7), Paris Sorbonne-Cité, Paris, France, pp. 2–
11.
Doerge, F.: 2009, A scholarly confusion of tongues, or, is promising an illocutionary
act?, Lodz Papers in Pragmatics 5(1), 53–68.
Douven, I.: 2006, Assertion, knowledge, and rational credibility, Philosophical Review 115(4), 449–485.
Eckardt, R.: 2007, Licensing ‘or’, in S. Uli and P. Stateva (eds), Presupposition and
Implicature in Compositional Semantics, Palgrave Macmillan, New York, NY,
pp. 34–70.
Fagin, R. and Halpern, J.: 1988, Belief, awareness and limited reasoning, Artificial
Intelligence 34(1), 39–76.
Faller, M.: 2002, Semantics and Pragmatics of Evidentials in Cuzco Quechua, PhD thesis,
Stanford University.
Farkas, D. F. and Bruce, K. B.: 2010, On Reacting to Assertions and Polar Questions,
Journal of Semantics 27(1), 81–118.
Farkas, D. F. and Roelofsen, F.: forthcoming, Polar initiatives and polarity particles
in an inquisitive discourse model, Language .
Feinberg, Y.: 2004, Subjective reasoning—Games with unawareness, Research Paper
1875, Stanford Graduate School of Business, Stanford, CA.
Fox, D. and Katzir, R.: 2011, On the characterization of alternatives, Natural Language Semantics 19(1), 87–107.
Frank, M. C. and Goodman, N. D.: 2012, Predicting pragmatic reasoning in language games, Science 336(6084), 998.
BIBLIOGRAPHY
298
Franke, M.: 2008, Meaning and inference in case of conflict, in K. Balogh (ed.),
Proceedings of the 13th ESSLLI Student Session, pp. 65–74.
Franke, M.: 2009, Signal to Act: Game Theory and Pragmatics, PhD thesis, Universiteit
van Amsterdam.
Franke, M.: 2011, Quantity implicatures, exhaustive interpretation, and rational
conversation, Semantics and Pragmatics 4(1), 1–82.
Franke, M. and de Jager, T.: 2011, Now that you mention it: Awareness dynamics
in discourse and decisions, in A. Benz, C. Ebert, G. Jäger and R. Rooij (eds),
Language, Games, and Evolution, Vol. 6207 of Lecture Notes in Computer Science,
Springer, Berlin/Heidelberg, pp. 60–91.
Franke, M., de Jager, T. and van Rooij, R.: 2012, Relevance in cooperation and
conflict, Journal of Logic and Computation 22(1), 23–54.
Garcı́a-Carpintero, M.: 2013, Explicit performatives revisited, Journal of Pragmatics
49, 1–17.
Gazdar, G.: 1979, Pragmatics: Implicature, Presupposition and Logical Form, Academic
Press, New York.
Giddens, A.: 1984, The Constitution of Society, University of California Press, Berkeley, CA.
Ginet, C.: 1979, Performativity, Linguistics and Philosophy 3(2), 245–265.
Ginzburg, J.: 1996, Interrogatives: Questions, facts and dialogue, in S. Lappin
(ed.), The Handbook of Contemporary Semantic Theory, Blackwell Textbooks in
Linguistics, Blackwell, pp. 385–422.
Grice, H. P.: 1957, Meaning, The Philosophical Review 66(3), 377–388.
Grice, H. P.: 1969, Utterer’s meaning and intention, The Philosophical Review
78(2), 147–177.
BIBLIOGRAPHY
299
Grice, H. P.: 1975, Logic and conversation, in P. Cole and J. L. Morgan (eds), Speech
Acts, Vol. 3 of Syntax and Semantics, Academic Press, pp. 41—58.
Grice, H. P.: 1978, Further notes on logic and conversation, in P. Cole (ed.), Syntax
and Semantics 9: Pragmatics, Academic Press, New York.
Grice, H. P.: 1982, Meaning revisited, in N. V. Smith (ed.), Mutual Knowledge,
Academic Press. page numbers refer to the reprint in (Grice 1989).
Grice, H. P.: 1989, Studies in the Way of Words, Harvard University Press.
Grimshaw, J.: 1979, Complement selection and the lexicon, Linguistic Inquiry
10(2), 279–326.
Groenendijk, J. and Roelofsen, F.: 2009, Inquisitive semantics and pragmatics,
in J. M. Larrazabal and L. Zubeldia (eds), Meaning, Content, and Argument:
Proceedings of the ILCLI International Workshop on Semantics, Pragmatics, and
Rhetoric, University of the Basque Country, Publication Service, pp. 41–72.
Groenendijk, J. and Stokhof, M.: 1984, Studies on the Semantics of Questions and the
Pragmatics of Answers, PhD thesis, Universiteit van Amsterdam.
Groenendijk, J. and Stokhof, M.: 1991, Dynamic predicate logic, Linguistics and
Philosophy 14(1), 39–100.
Groenendijk, J., Stokhof, M. and Veltman, F.: 1995, Coreference and modality,
in S. Lappin (ed.), The Handbook of Contemporary Semantic Theory, Blackwell
Publishers, Oxford, UK, pp. 179–213.
Gunlogson, C.: 2003, True to Form: Rising and Falling Declaratives in English, Routledge, New York.
Gunlogson, C.: 2008, A question of commitment, Belgian Journal of Linguistics
22(1), 101–136.
BIBLIOGRAPHY
300
Gutiérrez-Rexach, J.: 1996, The semantics of exclamatives, in E. Garrett and F. Lee
(eds), Semantics at sunset. UCLA working papers in linguistics, University of
California, Los Angeles, Los Angeles, CA, pp. 146–162.
Hamblin, C. L.: 1958, Questions, Australasian Journal of Philosophy 36(3), 159–168.
Hamblin, C. L.: 1971, Mathematical models of dialogue, Theoria 37(2), 130–155.
Han, C.-H.: 2002, Interpreting interrogatives as rhetorical questions, Lingua
112(3), 201–229.
Hansson, S. O.: 1996, What is ceteris paribus preference?, Journal of Philosophical
Logic 25(3), 307–332.
Hare, R. M.: 1968, Wanting: Some pitfalls, in R. Binkley (ed.), Agent, Action and
Reason, Proceedings of the Western Ontario Colloquium.
Hare, R. M.: 1970, Meaning and speech acts, Philosophical Review 79(1), 3–24.
Harnish, R. M.: 2005, Commitments and speech acts, Philosophica 75(1), 11–41.
Hausser, R.: 1978, Surface compositionality and the semantics of mood, in J. Groenendijk and M. Stokhof (eds), Amsterdam Papers in Formal Grammar, Vol. II,
Universiteit van Amsterdam.
Heal, J.: 1974, Explicit performative utterances and statements, The Philosophical
Quarterly 24(95), 106–121.
Hedenius, I.: 1963, Performatives, Theoria 29(2), 115–136.
Heim, I.: 1983, File change semantics and the familiarity theory of definiteness,
in R. Bäuerle, R. Schwarze and A. von Stechow (eds), Meaning, Use and the
Interpretation of Language, Walter de Gruyter, Berlin, pp. 164–190.
Hendriks, P. and de Hoop, H.: 2001, Optimality theoretic semantics, Linguistics and
Philosophy 24(1), 1—32.
BIBLIOGRAPHY
301
Hindriks, F.: 2009, Constitutive rules, language, and ontology, Erkenntnis 71(2), 253–
275.
Hintikka, J.: 1961, Modality and quantification, Theoria 27(3), 119–128.
Hirschberg, J.: 1985, A theory of scalar implicature, PhD thesis, University of Pennsylvania.
Hiz, H.: 1978, Introduction, in H. Hiz (ed.), Questions, Reidel, Dordrecht.
Horn, L. R.: 1972, On the Semantic Properties of Logical Operators in English, PhD
thesis, University of California, Los Angeles.
Horn, L. R.: 1984, Towards a new taxonomy of pragmatic inference: Q-based and
R-based implicature, Meaning, Form, and Use in Context: Linguistic Applications,
Georgetown University Press, Washington, DC, pp. 11–42.
Horn, L. R.: 1989, A Natural History of Negation, Chicago University Press, Chicago,
IL.
Horn, L. R.: 2006, The border wars: A neo-Gricean perspective, in K. von Heusinger
and K. Turner (eds), Where semantics meets pragmatics, Vol. 16 of Current Research
in the Semantics/Pragmatics Interface, Elsevier, Amsterdam/London, pp. 21–48.
Huitink, J. and Spenader, J.: 2004, Cancelation resistant PCis, in R. van der Sandt
and B. Geurts (eds), Proceedings of the ESSLLI 2004 Workshop on Implicature and
Conversational Meaning.
Ivlieva, N.: 2012, Obligatory implicatures and grammaticality, in M. Aloni, V. Kimmelman, F. Roelofsen, G. Sassoon, K. Schulz and M. Westera (eds), Logic,
Language and Meaning. Proceedings of the 18th Amsterdam Colloquium, Revised Selected Papers, Vol. 7218 of Lecture Notes in Computer Science, Springer,
Berlin/Heidelberg, pp. 381–390.
Jäger, G.: 2007, Game dynamics connects semantics and pragmatics, in A.-V. Pietarinen (ed.), Game Theory and Linguistic Meaning, Elsevier, Amsterdam/New York,
pp. 89–102.
BIBLIOGRAPHY
302
Jäger, G.: 2012, Game theory in semantics and pragmatics, in C. Maienborn, P. Portner and K. von Heusinger (eds), Semantics. An International Handbook of Natural
Language Meaning, Vol. 3, De Gruyter Mouton, Berlin, pp. 2487–2516.
Jäger, G. and Ebert, C.: 2009, Pragmatic rationalizability, in A. Riester and T. Solstad
(eds), Proceedings of Sinn und Bedeutung 13, University of Stuttgart, OPUS.
Joshi, A. K.: 1982, Mutual belief in question answering systems, in N. S. Smith (ed.),
Mutual Knowledge, Academic Press, London, pp. 181–197.
Kadmon, N.: 2001, Formal Pragmatics: Semantics, Pragmatics, Presupposition, and
Focus, Blackwell, Malden, MA.
Kamp, H.: 1978, Semantics versus pragmatics, in F. Guenthner and S. J. Schmidt
(eds), Formal Semantics and Pragmatics for Natural Languages, Reidel, Dordrecht,
pp. 255–287.
Kamp, J. A. W.: 1981, A theory of truth and semantic representation, in J. Groenendijk, T. Janssen and M. Stokhof (eds), Formal Methods in the Study of Language, Vol. 135 of Mathematical Centre Tracts, Mathematisch Centrum, Amsterdam, pp. 277–322.
Kanger, S.: 1957, Provability in Logic, Almqvist and Wiksell, Stockholm.
Kaplan, D.: 1999, What is meaning? Explorations in the theory of meaning as use.
Manuscript, Brief Version – Draft # 1. University of California, Los Angeles.
Karttunen, L.: 1977, Syntax and semantics of questions, Linguistics and Philosophy
1(1), 3–44.
Katzir, R.:
2007, Structurally-defined alternatives, Linguistics and Philosophy
30(6), 669–690.
Kaufmann, M.: 2012, Interpreting Imperatives, Springer, Dordrecht/New York.
Kaufmann, S.: 2005, Conditional truth and future reference, Journal of Semantics
22(3), 231–280.
BIBLIOGRAPHY
303
Kaufmann, S. and Schwager, M.: 2009, A uniform analysis of conditional imperatives, in E. Cormany, S. Ito and D. Lutz (eds), Proceedings of Semantics and
Linguistic Theory (SALT) 19, Cornell University, CLC Publications, Ithaca, NY,
pp. 239–256.
Kratzer, A.: 1981, The notional category of modality, in H. J. Eikmeyer and H. Rieser
(eds), Words, Worlds, and Contexts. New Approaches in Word Semantics, de
Gruyter, Berlin, pp. 38–74.
Kratzer, A.: 1995, Stage-level and individual-level predicates, in G. N. Carlson and
F. J. Pelletier (eds), The generic book, The University of Chicago Press, Chicago,
pp. 125–175.
Krifka, M.: 2001a, For a structured meaning account of questions and answers, in
C. Féry and W. Sternefeld (eds), Audiatur Vox Sapientiae: A Festschrift for Arnim
von Stechow, Akademie-Verlag, Berlin, pp. 287–319.
Krifka, M.: 2001b, Quantifying into question acts, Natural Language Semantics
9(1), 1—40.
Krifka, M.: to appear, Embedding speech acts, in T. Roeper and P. Speas (eds),
Recursion in Language and Cognition.
Kripke, S.: 1963, Semantical analysis of modal logic, I. normal modal propositional
calculi, Zeitschrift für Mathematische Logik und Grundlagen der Mathematik 9, 67–
96.
Lackey, J.: 2007, Norms of assertion, Noûs 41(4), 594–626.
Landman, F.: 1986, Towards a Theory of Information: The status of partial objects in
semantics, Foris Publications, Dordrecht/Riverton, MJ.
Lasersohn, P.: 1999, Pragmatic halos, Language 75(3), 522–551.
Lauer, S.: 2012, On the pragmatics of pragmatic slack, in A. Aguilar, R. Nouwen
and A. Chernilovskaya (eds), Proceedings of Sinn and Bedeutung (SuB) 16, MIT
Working papers in Linguistics, Cambridge, MA, pp. 389–401.
BIBLIOGRAPHY
304
Lauer, S. and Condoravdi, C.: 2012, The basic dynamic effect of interrogative utterances. Talk presented at the 13th Texas Linguistics Society (TLS) conference,
University of Texas, Austin, June 2012.
Leech, G. N.: 1983, Principles of Pragmatics, Longman, London/New York.
Lemmon, E. J.: 1962, On sentences verifiable by their use, Analysis 22(4), 86–89.
Levinson, D.: 2003, Probabilistic model-theoretic semantics for want, in R. Young
and Y. Zhou (eds), Proceedings of Semantics and Linguistic Theory (SALT) 13,
Cornell University, CLC Publications, Ithaca, NY, pp. 222–239.
Levinson, S.: 1983, Pragmatics, Cambridge University Press, Cambridge, UK.
Levinson, S.: 2000, Presumptive Meanings, MIT Press, Cambridge, MA.
Lewis, D.: 1969, Convention: A philosophical study, Harvard University Press, Cambridge, MA.
Lewis, D.: 1975, Languages and language, in K. Gunderson (ed.), Language, Mind,
and Knowledge, University of Minnesota Press, Minneapolis, MN, pp. 3–35.
Lewis, D.: 1979, Scorekeeping in a language game, Journal of Philosophical Logic
8(1), 339–359.
Lewis, D. and Lewis, S. R.: 1975, Review of Olson and Paul (1972) , Theoria 41(1), 39–
60.
MacFarlane, J.: 2005, Making sense of relative truth, Proceedings of the Aristotelian
Society 105(1), 305–323.
MacFarlane, J.: 2011, What is assertion?, in J. Brown and H. Cappelen (eds), Assertion, Oxford University Press, Oxford, UK, pp. 79–96.
Magri, G.: 2009, A theory of individual-level predicates based on blind mandatory
scalar implicatures, Natural Language Semantics 17(3), 245–297.
BIBLIOGRAPHY
305
Magri, G.: 2011, Another argument for embedded scalar implicatures based on
oddness in downward entailing contexts, Proceedings of Semantics and Linguistic
Theory (SALT) 20, Cornell University, CLC Publications, Ithaca, NY, pp. 564–
581.
Mastop, R.: 2005, What can you do?, PhD thesis, Universiteit van Amsterdam.
Mayol, L. and Castroviejo, E.: 2013, How to cancel an implicature, Journal of Pragmatics 50(1), 84–104.
Olson, R. and Paul, A.: 1972, Contemporary Philosophy in Scandinavia, Johns Hopkins,
Baltimore/London.
Owens, D.: 2006, Testimony and assertion, Philosophical Studies 130(1), 105–129.
Parikh, P.: 2001, The Use of Language, CSLI Publications, Stanford, CA.
Portner, P.: 2005, The semantics of imperatives within a theory of clause types,
in R. B. Young (ed.), Proceedings of Semantics and Linguistic Theory (SALT) 14,
Vol. 4, Cornell University, CLC Publications, Ithaca, NY, pp. 235–252.
Portner, P.: 2007, Imperatives and modals, Natural Language Semantics 15, 351–383.
Portner, P.: 2012, Permission and choice, in G. Grewendorf and T. E. Zimmermann
(eds), Discourse and Grammar: From Sentence Types to Lexical Categories, Studies
in Generative Grammar, De Gruyter Mouton, Berlin, pp. 43–68.
Potts, C.: 2005, The Logic of Conventional Implicatures, Oxford University Press,
Oxford, UK.
Potts, C.: 2007, The expressive dimension, Theoretical Linguistics 33(2), 165–197.
Potts, C. and Kawahara, S.: 2004, Japanese honorifics as emotive definite descriptions, in K. Watanabe and R. B. Young (eds), Proceedings of Semantics and
Linguistic Theory (SALT) 14, Cornell University, CLC Publications, Ithaca, NY,
pp. 235–254.
BIBLIOGRAPHY
306
Prince, A. and Smolensky, P.: 1993, Optimality theory: Constraint interaction in
generative grammar, Technical Report 2, Rutgers University Center for Cognitive Science.
Ransdell, J.: 1971, Constitutive rules and speech-act analysis, The Journal of Philosophy 68(13), 385–400.
Åqvist, L.: 1965, A new approach to the logical theory of interrogatives, University of
Uppsala.
Åqvist, L.: 1983, On the “tell me truly” approach to the analysis of interrogatives,
in F. Kiefer (ed.), Questions and Answers, Reidel, Dordrecht, pp. 9–14.
Rett, J.: 2008, A degree account of exclamatives, in T. Friedman and S. Ito (eds),
Proceedings of SALT XVIII, Cornell University, Ithaca, NY, pp. 601–618.
Rett, J.: 2011, Exclamatives, degrees and speech acts, Linguistics and Philosophy
34(5), 411–442.
Roberts, C.: 1996, Information structure in discourse: Towards an integrated account of formal pragmatics, Papers in Semantics, OSU Working Papers in Linguistics, Vol. 49, Department of Linguistics, The Ohio State University, Columbus,
OH.
Rohde, H.: 2006, Rhetorical questions as redundant interrogatives, San Diego Linguistics Papers 2(134–168).
Romero, M. and Han, C.-H.: 2004, On negative “yes/no” questions, Linguistics and
Philosophy 27(5), 609–658.
Ross, J. R.: 1970, On declarative sentences, in R. Jacobs and P. Rosenbaum (eds),
Readings in English transformational grammar, Ginn, Waltham, MA, pp. 222–272.
Russell, B.: 2006, Against grammatical computation of scalar implicatures, Journal
of Semantics 23(4), 361–382.
BIBLIOGRAPHY
307
Sadock, J. M.: 1978, On testing for conversational implicature, Syntax and Semantics
9: Pragmatics, Academic Press, New York, pp. 281–297.
Sadock, J. M. and Zwicky, A. M.: 1985, Speech act distinctions in syntax, in T. Shopen
(ed.), Language Typology and Syntactic Description, Vol. I, Cambridge University
Press, Cambridge, UK, pp. 155–196.
Sauerland, U.: 2004, Scalar implicatures in complex sentences, Linguistics and Philosophy 27(3), 367—391.
Schlenker, P.: 2010, Presuppositions and local contexts, Mind 119(474), 377–391.
Schmerling, S.: 1982, How imperatives are special and how they aren’t, in R. Schneider, K. Tuite and R. Chametzky (eds), Papers from the Parasession on Nondeclaratives: Chicago Linguistic Society, Chicago, IL, pp. 202–218.
Schwager, M.: 2006, Interpreting Imperatives, PhD thesis, Johann Wolfgang GoetheUniverstät, Frankfurt am Main.
Searle, J. R.: 1969, Speech Acts: An essay in the philosophy of language, Cambridge
University Press, Cambridge, UK.
Searle, J. R.: 1989, How performatives work, Linguistics and Philosophy 12(5), 535–
558.
Siebel, M.: 2003, Illocutionary acts and attitude expression, Linguistics and Philosophy 26(3), 351–366.
Sperber, D. and Wilson, D.: 1986, Relevance: Communication and Cognition, Blackwell
Publishers, Oxford, UK.
Stalnaker, R.: 1974, Pragmatic presupposition, in M. Munitz and P. Unger (eds),
Semantics and Philosphy, New York University Press, New York, pp. 197–213.
Stalnaker, R.: 1978, Assertion, in P. Cole (ed.), Syntax and Semantics 9: Pragmatics,
New York Academic Press, New York, pp. 315—332.
BIBLIOGRAPHY
308
Stalnaker, R.: 1994, On the evaluation of solution concepts, Theory and Decision
37, 49–74.
Stalnaker, R.: 1996, Knowledge, belief and counterfactual reasoning in games,
Economics and Philosophy 12(2), 133–163.
Stalnaker, R.: 2002, Common ground, Linguistics and Philosophy 25(5-6), 701–721.
Stenius, E.: 1967, Mood and language-game, Synthese 17, 254–274.
Strawson, P. F.: 1964, Intention and convention in speech acts, The Philosophical
Review 73(4), 439–460.
Szabolcsi, A.: 1982, Model theoretic semantics of performatives, in F. Kiefer (ed.),
Hungarian and General Linguistics, John Benjamins, Amsterdam, pp. 515–536.
Thomason, R. H.: 1984, Combinations of tense and modality, in D. Gabbay and
F. Guenther (eds), Handbook of Philosophical Logic: Extensions of Classical Logic,
Reidel, Dordrecht, pp. 135–165.
van Benthem, J.: 2011, Logical Dynamics of Information and Interaction, Cambridge
University Press, Cambridge, UK/New York.
van Benthem, J. and Pacuit, E.: 2006, Thre tree of knowledge in action: Towards a
common perspective, in uido Governatori, I. Hodkinson and Y. Venema (eds),
Advances in Modal Logic, Vol. 6, College Publications, London.
van Ditmarsch, H., van der Hoek, W. and Kooi, B. (eds): 2008, Dynamic Epistemic
Logic, Vol. 337 of Studies in Epistemology, Logic, Methodology and Philosophy of
Science, Springer, Dordrecht.
van Rooij, R.: 2004, Signaling games select Horn strategies, Linguistics and Philosophy 27(4), 493–527.
Veltman, F.: 1996, Defaults in update semantics, Journal of Philosophical Logic
25(3), 221–261.
BIBLIOGRAPHY
309
Vogel, A., Potts, C. and Jurafsky, D.: 2013, Implicatures and nested beliefs in
approximate Decentralized-POMDPs, Proceedings of the 2013 Annual Conference
of the Association for Computational Linguistics, Association for Computational
Linguistics, Stroudsburg, PA, pp. 74–80.
von Savigny, E.: 1988, The social foundations of meaning, Springer, Berlin.
Weiner, M.:
2006, Are all conversational implicatures cancellable?, Analysis
66(2), 127–130.
Williams, B.: 2002, Truth and Truthfulness, Princeton University Press, Princeton,
NJ.
Williamson, T.: 1996, Knowing and asserting, Philosophical Review 105(4), 489–523.
Witek, M.: 2009, Scepticism about reflexive intentions refuted, Lodz Papers in Pragmatics 5(1), 69–83.
Zanuttini, R. and Portner, P.: 2003, Exclamative clauses: At the syntax-semantics
interface, Language 79(1), 39–81.
Zimmermann, M.: 2004, Zum wohl: Diskurspartikeln als Satztypmodifikatoren,
Linguistische Berichte 199, 253–286.
Zimmermann, T. E.: 2000, Free choice disjunction and epistemic possibility, Natural
Language Semantics 8(4), 255–290.
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

advertisement