Literal Meaning

François Recanati
Institut Jean-Nicod
1bis, avenue de Lowendal, 75007 Paris, France
[email protected]

1
 Contents

Introduction

Chapter 1: Two approaches to 'what is said'

1.1 The basic triad
1.2 Minimalism
1.3 Literal truth-conditions vs actual truth-conditions
1.4 A problem for Minimalism
1.5 The availability of what is said
1.6 The availability-based approach
1.7 'Saying' as a pragmatic notion
1.8 Availability vs Minimalism

Chapter 2: Primary pragmatic processes

2.1 Enrichment, loosening, and transfer
2.2 Rejecting the Gricean picture
2.3 Accessibility
2.4 Objections and responses
2.5 Interactive processing
2.6 The role of schemata

Chapter 3: Relevance-theoretic objections

3.1 One or two systems?
3.2 Personal and sub-personal inferences
3.3 Implicature or enrichment?
3.4 Mutual adjustment of explicature and implicature
3.5 Implicated premisses

Chapter 4: The Syncretic View

4.1 Four levels?
4.2 Semantics and pragmatics: the standard picture
4.3 Semantic underdeterminacy

2
4.4 The minimal proposition as 'common denominator'
4.5 Interaction between saturation and optional pragmatic processes
4.6 Taking stock
4.7 Minimal or reflexive?

Chapter 5: Nonliteral uses

5.1 Nonliteral uses as nonminimal departures from literal meaning
5.2 Nonliteral uses and secondary meaning
5.3 Nonminimal departures without secondariness
5.4 The transparency condition
5.5 Varieties of nonliteral meaning
5.6 Conclusion

Chapter 6: From Literalism to Contextualism

6.1 Five positions
6.2 Indexicalism
6.3 Contextualism
6.4 Literalist responses to the contextualist challenge
6.5 Where Indexicalism and Contextualism meet

Chapter 7: Indexicalism and the Binding Fallacy

7.1 Mandatory v. Optional
7.2 Two criteria
7.3 The indexicalist challenge
7.4 Is the Binding Criterion reliable?
7.5 Variadic functions
7.6 The Binding Fallacy
7.7 Conclusion: the failure of Indexicalism

Chapter 8: Circumstances of evaluation

8.1 Modality
8.2 Time and tense
8.3 Situations

3
8.4 Saturation or enrichment?
8.5 Sub-sentential circumstances
8.6 Conclusion

Chapter 9: Contextualism: How far can we go?

9.1 The modulation of sense
9.2 The semantic relevance of modulation
9.3 Four approaches
9.4 Truth-conditional unstability: from Waismann's 'open texture' to
Searle's 'background'
9.5 Ostensive definitions
9.6 Meaning Eliminativism
9.7 Conclusion

4
 Introduction

Around the middle of the twentieth century, there were two opposing
camps within the analytic philosophy of language. The first camp — IDEAL

LANGUAGE PHILOSOPHY, as it was then called — was that of the pioneers,

Frege, Russell, Carnap, Tarski, etc. They were, first and foremost, logicians
studying formal languages and, through them, “language” in general.
They were not originally concerned with natural language, which they
thought defective in various ways;1 yet, in the sixties, some of their
disciples established the relevance of their methods to the detailed study
of natural language (Montague 1974; Davidson 1984). Their efforts gave
rise to contemporary FORMAL SEMANTICS, a very active discipline whose
stunning developments in the last quarter of the twentieth century
changed the face of linguistics.
The other camp was that of so-called ORDINARY LANGUAGE PHILOSOPHERS,

who thought important features of natural language were not revealed but
hidden by the logical approach initiated by Frege and Russell. They
advocated a more descriptive approach and emphasized the pragmatic
nature of natural language as opposed to, say, the formal language of
Principia Mathematica. Their own work2 gave rise to contemporary
pragmatics, a discipline which, like formal semantics, developed
successfully within linguistics in the past forty years.
Central in the ideal language tradition had been the equation of, or
at least the close connection between, the meaning of a (declarative)
sentence and its truth conditions. This truth-conditional approach to

1 There are a few exceptions. The most important one is Reichenbach,

whose insightful “Analysis of conversational language” was published in

1947 as a chapter — the longest — in his Elements of Symbolic Logic.

2 The most influential authors were Austin, Strawson, Grice, and the later

Wittgenstein. Grice is a special case, for he thought the two approaches

were not incompatible but complementary.

5
meaning is perpetuated, to a large extent, in contemporary formal
semantics. A language is viewed as a system of rules or conventions, in
virtue of which certain assemblages of symbols count as well-formed,
meaningful sentences. The meaning of a sentence (or of any complex
symbol) is determined by the meanings of its parts and the way they are
put together. Meaning itself is patterned after reference. The meaning of a
simple symbol is the conventional assignment of a worldly entity to that
symbol: for example, names are assigned objects, monadic predicates are
assigned properties or sets of objects, etc. The meaning of a declarative
sentence, determined by the meanings of its constituents and the way
they are put together, is equated with its truth conditions. For example,
the subject-predicate construction is associated with a semantic rule for
determining the truth conditions of a subject-predicate sentence on the
basis of the meaning assigned to the subject and that assigned to the
predicate.3 On this picture, knowing a language is like knowing a ‘theory’
by means of which one can deductively establish the truth conditions of
any sentence of that language.
This truth-conditional approach to meaning is one of the things
which ordinary language philosophers found quite unpalatable. According
to them, reference and truth cannot be ascribed to linguistic expressions
in abstraction from their use. In vacuo, words do not refer and sentences
do not have truth conditions. Words-world relations are established
through, and indissociable from, the use of language. It is therefore
misleading to construe the meaning of a word as some worldly entity that
it represents or, more generally, as its truth-conditional contribution. The
meaning of a word, insofar as there is such a thing, should rather be
equated with its use-potential or its use-conditions. In any case, what must
be studied primarily is speech: the activity of saying things. Then we will
be in a position to understand language, the instrument we use in speech.

3 The rule might go as follows: “A sentence of the form ‘a is F’ is true if

and only if the entity denoted by ‘a’ belongs to the set of entities denoted

by ‘F’.”

6
Austin’s theory of speech acts and Grice’s theory of speaker’s meaning
were both meant to provide the foundation for a theory of language, or at
least for a theory of linguistic meaning.
Despite the early antagonism I have just described, semantics (the
formal study of meaning and truth conditions) and pragmatics (the study
of language in use) are now conceived of as complementary disciplines,
shedding light on different aspects of language. The heated arguments
between ideal language philosophers and ordinary language philosophers
are almost forgotten. There are two main reasons for the new situation. On
the one hand semanticists, in moving from artificial to natural languages,
have given up Carnap’s idea that the semantic relation between words and
the world can be studied in abstraction from the context of use. That the
Carnapian abstraction is illegitimate given the pervasiveness of context-
sensitivity in natural language is fully acknowledged by those working in
formal semantics. On the other hand those working in pragmatics no
longer hold that ‘meaning is use’. Instructed by Grice, they systematically
draw a distinction between what a given expression means, and what its
use means or conveys, in a particular context (or even in general).
Still, the ongoing debate about the best delimitation of the
respective territories of semantics and pragmatics betrays the persistence
of two recognizable currents or approaches within contemporary
theorizing. According to the dominant position, which I call ‘Literalism’, we
may legitimately ascribe truth-conditional content to natural language
sentences, quite independent of what the speaker who utters this
sentence means. Literalism contrasts with another view, reminiscent of
that held by ordinary language philosophers half a century ago. That other
view, which I call ‘Contextualism’, holds that speech acts are the primary
bearers of content. Only in the context of a speech act does a sentence
express a determinate content.
I say that Literalism is the dominant position because I believe most
philosophers of language and linguists would accept the following
description of the division of labour between semantics and pragmatics :

7
 Semantics deals with the literal meaning of words and
sentences as determined by the rules of the language, while
pragmatics deals with what users of the language mean by their
utterances of words or sentences. To determine 'what the speaker
means' is to answer questions such as: Was John's utterance
intended as a piece of advice or as a threat? By saying that it was
late, did Mary mean that I should have left earlier? Notions such as
that of illocutionary force (Austin) and conversational implicature
(Grice) thus turn out to be the central pragmatic notions. In contrast,
the central semantic notions turn out to be reference and truth. It is
in terms of these notions that one can make explicit what the
conventional significance of most words and expressions consists in.
The meaning of an expression may be insufficient to determine
its referential content: that is so whenever the expression is
indexical or otherwise context-dependent. In such cases, the
meaning of the expression provides a rule which, given a context,
enables the interpreter to determine the content of the expression in
that context. The content thus determined in context by the
conventional meanings of words is their literal content. The literal
content of a complete declarative utterance is 'what is said', or the
proposition expressed, by that utterance.
As Grice emphasized, speaker's meaning is not a matter of
rules but a matter of intentions: what someone means is what he or
she overtly intends (or, as Grice says, 'M-intends') to get across
through his or her utterance. Communication succeeds when the M-
intentions of the speaker are recognized by the hearer. Part of the
evidence used by the hearer in figuring out what the speaker means
is provided by the literal content of the uttered sentence, to which
the hearer has independent access via his knowledge of the
language. In ideal cases of linguistic communication, the speaker
means exactly what she says, and no more is required to understand
the speech act than a correct understanding of the sentence uttered
in performing it. In real life, however, what the speaker means

8
 typically goes beyond, or otherwise diverges from, what the uttered
sentence literally says. In such cases the hearer must rely on
background knowledge to determine what the speaker means —
what her communicative intentions are.

There is much that is correct in this description, but there also is

something which I think must be rejected, namely the contrast between
literal truth-conditions and speaker’s meaning. That contrast commits us
to Literalism, and in this book I want to argue for Contextualism. According
to Contextualism, the contrast between what the speaker means and what
she literally says is illusory, and the notion of ‘what the sentence says’
incoherent. What is said (the truth-conditional content of the utterance) is
nothing but an aspect of speaker’s meaning. That is not to deny that there
is a legitimate contrast to be drawn between what the speaker says and
what he or she merely implies. Both, however, belong to the realm of
‘speaker’s meaning’ and are pragmatic through and through.
I will not only criticize Literalism, and argue for Contextualism, in the
following chapters. I will discuss all sorts of intermediate positions
corresponding to views actually held in the current debate about the
semantics/pragmatics interface. Whether or not one accepts my
arguments, I hope the survey of logical space which I provide will be useful
to those interested in the debate, and will contribute to shaping it in the
years to come.

9
 Chapter 1

Two approaches to 'what is said'

1.1. The basic triad

Anyone who has reflected on the sentence meaning/speaker's meaning

distinction knows that a simple distinction is in fact insufficient. Two
equally important distinctions must be made. First, there is the distinction
between the linguistic meaning of a sentence-type, and what is said (the
proposition expressed) by an utterance of the sentence. For example, the
English sentence 'I am French' has a certain meaning which, qua meaning
of a sentence-type, is not affected by changes in the context of utterance.
This context-independent meaning contrasts with the context-dependent
propositions which the sentence expresses with respect to particular
contexts. Thus 'I am French', said by me, expresses the proposition that I
am French; if you utter the sentence, it expresses a different proposition,
even though its linguistic meaning remains the same across contexts of
use.
Second, there is a no less important distinction between what is
actually said and what is merely 'conveyed' by the utterance. My
utterance of 'I am French' expresses the proposition that I am French, but
there are contexts in which it conveys much more. Suppose that, having
been asked whether I can cook, I reply: 'I am French'. Clearly my utterance
(in this context) provides an affirmative answer to the question. The
meaning of the utterance in such a case includes more than what is
literally said; it also includes what the utterance 'implicates'. 4

4 See Paul Grice, Studies in the Way of Words (Harvard University Press,

1989), p. 24: « I wish to introduce, as terms of art, the verb implicate and

the related nouns implicature (cf implying) and implicatum (cf what is

10
 'What is said' being a term common to both distinctions, we end up
with a triad:

sentence meaning
vs
what is said
vs
what is implicated

The distinguishing characteristic of sentence meaning (the linguistic

meaning of the sentence type) is that it is conventional and context-
independent. Moreover, in general at least, it falls short of constituting a
complete proposition, i.e. something truth-evaluable. In contrast, both
'what is said' and 'what is implicated' are context-dependent and
propositional. The difference between 'what is said' and 'what is
implicated' is that the former is constrained by sentence meaning in a way
in which the implicatures aren't. What is said results from fleshing out the
meaning of the sentence (which is like a semantic 'skeleton') so as to
make it propositional. The propositions one can arrive at through this
process of contextual enrichment or 'fleshing out' are constrained by the
skeleton which serves as input to the process. Thus 'I am French' can
express an indefinite number of propositions, but the propositions in
question all have to be compatible with the semantic potential of the
sentence; this is why the English sentence 'I am French' cannot express
the proposition that kangaroos have tails. There is no such constraint on
the propositions which an utterance of the sentence can communicate
through the mechanism of implicature. Given enough background, an
utterance of 'I am French' might implicate that kangaroos have tails.
What's implicated is implicated by virtue of an inference, and the inference

implied). The point of this maneuver is to avoid having, on each occasion,

to choose beween this or that member of the family of verbs for which

implicate is to do general duty. »

11
chain can (in principle) be as long and involve as many background
assumptions as one wishes.
The basic triad can be mapped back onto the simple sentence
meaning/speaker's meaning distinction by grouping together two of the
three levels. There are two ways to do it, corresponding to two
interpretations for the triad. The 'minimalist' interpretation stresses the
close connection between sentence meaning and what is said; together,
sentence meaning and what is said constitute the literal meaning of the
utterance as opposed to what the speaker means:

{ sentence meaning
literal meaning
what is said

vs

speaker's meaning

The other, 'non-minimalist' interpretation of the triad stresses the

commonality between what is said and what is implicated, both of which
are taken to be pragmatically determined:

sentence meaning


vs
what is said
speaker's meaning
what is implicated

Essential to this interpretation is the claim that 'what is said', though

constrained by the meaning of the sentence, is not as tightly constrained
as is traditionally thought and, in particular, does not obey what I will refer
to as the 'minimalist' constraint.

12
 1.2. Minimalism

As I said above, what distinguishes 'what is said' from the implicatures is

the fact that the former must be « closely related to the conventional
meaning of the words (the sentence) [one] has uttered » (Grice, Studies in
the Way of Words, p. 25). However, this constraint — which I call the
'closeness-to-linguistic-meaning' constraint, or CLM — can be construed
more or less strictly. The two interpretations of the basic triad to be
discussed below differ in how strictly they construe the CLM. What I call
'Minimalism' construes the constraint very strictly; 'what is said', in the
minimalist framework, departs from the conventional meaning of the
sentence (and incorporates contextual elements) only when this is
necessary to 'complete' the meaning of the sentence and make it
propositional. In other words, the distance between sentence meaning and
what is said is kept to a minimum (hence the name 'Minimalism').
The crucial notion here is that of 'saturation'. Saturation is the
process whereby the meaning of the sentence is completed and made
propositional through the contextual assignment of semantic values to the
constituents of the sentence whose interpretation is context-dependent
(and, possibly, through the contextual provision of 'unarticulated'
propositional constituents, if one assumes, as some philosophers do, that
such constituents are sometimes needed to make the sentence fully
propositional). This process takes place whenever the meaning of the
sentence includes something like a 'slot' requiring completion or a 'free
variable' requiring contextual instantiation. 5 Thus an indexical sentence

5 Even when saturation consists in contextually providing a constituent

that is unarticulated in surface syntax (as the implicit argument in 'I

noticed'), it is something in the sentence (here the predicate 'notice',

which arguably denotes a two-place relation) which triggers the search for

the contextual element and makes it obligatory. See my ‘Unarticulated

Constituents’, in Linguistics and Philosophy 25 (2002), pp. 308-13.

13
like 'He is tall' does not express a complete proposition unless a referent
has been contextually assigned to the demonstrative pronoun 'he', which
acts like a free variable in need of contextual instantiation. Genitives
provide another well-known example: an utterance including the phrase
'John's book' does not express a complete proposition unless a particular
relation has been identified as holding between the book and John.
Nominal compounds work the same way: 'burglar nightmare' means
something like 'a nightmare that bears a certain relation R to burglars',
which relation must be contextually identified. Other well-known examples
of saturation include parametric predicates ('small', 'on the left'), definite
null instantiation (i.e. the case where one of the arguments in the
semantic structure of a lexeme, typically a verb, is not syntactically
realized and must be contextually identified, as when someone says: 'I
heard' or 'I noticed'), and so on and so forth.
Whenever saturation is in order, appeal to the context is necessary
for the utterance to express a complete proposition: from a semantic point
of view, saturation is a mandatory contextual process. Other contextual
processes — e.g. the inference process generating implicatures — are
semantically optional in the sense that the aspects of meaning they
generate are dispensable; the utterance would still express a complete
proposition without them. According to Minimalism, those extra
constituents of meaning which are not necessary for propositionality are
external to what is said. The only justification for including some
pragmatically determined constituent of meaning into what is said (as
opposed to what is merely conveyed) is the indispensability of such a
constituent - the fact that the utterance would not express a complete
proposition if the context did not provide such a constituent.

1.3. Literal truth-conditions vs. actual truth-conditions

Consider examples (1)-(6), often discussed in the literature:

(1) I've had breakfast

14
(2) You are not going to die
(3) It's raining
(4) The table is covered with books
(5) Everybody went to Paris
(6) John has two children

In all such cases, as we shall see, the Minimalist constraint implies that
what the utterance literally says is not what intuitively seems to be said.
From a minimalist point of view, the first sentence, 'I've had
breakfast' expresses the proposition that S (the speaker) has had
breakfast before t* (the time of utterance). Strictly speaking this
proposition would be true if the speaker had had breakfast twenty years
ago and never since. This is clearly not what the speaker means (when she
answers to the question 'Do you want something to eat' and replies 'I've
had breakfast'); she means something much more specific, namely that
she's had breakfast on that very day (i.e. the day which includes t*). This
aspect of speaker's meaning, however, has to be construed as external to
what is said and as being merely conveyed, in the same way in which the
utterer of 'I am French' implies, but does not say, that he is a good cook.
That is so because the 'minimal' interpretation, to the effect that the
speaker's life was not entirely breafkastless, is sufficient to make the
utterance propositional. Nothing in the sentence itself forces us to bring in
the implicit reference to a particular time span. Indeed we can easily
imagine contexts in which a speaker would use the same sentence to
assert the minimal proposition and nothing more.
The same thing holds even more clearly for the second example.
Kent Bach, to whom it is due, imagines a child crying because of a minor
cut and her mother uttering (2) in response (Bach 1994b: 134). What is
meant is: 'You're not going to die from that cut'. But literally the utterance
expresses the proposition that the kid will not die tout court — as if he or
she was immortal. The extra element contextually provided (the implicit
reference to the cut) does not correspond to anything in the sentence
itself; nor is it an unarticulated constituent whose contextual provision is

15
necessary to make the utterance fully propositional. Again, we can easily
imagine a context in which the same sentence would be used to
communicate the minimal proposition and nothing more.
What about (3)? John Perry and many others after him have argued
as follows (Perry 1986). Even though nothing in the sentence 'It's raining'
stands for a place, nevertheless it does not express a complete proposition
unless a place is contextually provided. The verb ‘to rain’, Perry says,
denotes a dyadic relation — a relation between times and places. In a
given place, it doesn’t just rain or not, it rains at some times while not
raining at others; similarly, at a given time, it rains in some places while
not raining in others. To evaluate a statement of rain as true or false, Perry
says, we need both a time and a place. Since the statement ‘It is raining’
explicitly gives us only the two-place relation (supplied by the verb) and
the temporal argument (indexically supplied by the present tense), the
relevant locational argument must be contextually supplied for the
utterance to express a complete proposition. If Perry is right, the
contextual provision of the place concerned by the rain is an instance of
saturation, like the assignment of a contextual value to the present tense:
both the place and the time are constituents of what is said, even though,
unlike the time, the place remains unarticulated in surface syntax.
But is Perry right? If really the contextual provision of a place was
mandatory, hence an instance of saturation, every token of 'Its raining'
would be unevaluable unless a place were contextually specified. Yet I
have no difficulty imagining a counterexample, i.e. a context in which ‘It is
raining’ is evaluable even though no particular place is contextually
singled out. In 'Unarticulated constituents' (p. 317) I depicted an imaginary
situation in which

rain has become extremely rare and important, and rain

detectors have been disposed all over the territory (whatever
the territory — possibly the whole Earth). In the imagined
scenario, each detector triggers an alarm bell in the Monitoring
Room when it detects rain. There is a single bell; the location

16
 of the triggering detector is indicated by a light on a board in
the Monitoring Room. After weeks of total drought, the bell
eventually rings in the Monitoring Room. Hearing it, the
weatherman on duty in the adjacent room shouts: ‘It’s raining!’
His utterance is true, iff it is raining (at the time of utterance)
in some place or other.

The fact that one can imagine an utterance of ‘It’s raining’ that is true iff it
is raining (at the time of utterance) in some place or other arguably
establishes the pragmatic nature of the felt necessity to single out a
particular place, in the contexts in which such a necessity is indeed felt.
When a particular place is contextually provided as relevant to the
evaluation of the utterance, it is for pragmatic reasons, not because it is
linguistically required. (Again, if it were linguistically required, in virtue of
semantic properties of the sentence type, it would be required in every
context.) If this is right, then the contextual provision of a place is not an
instance of saturation after all: it's not something that's mandatory. It
follows (by minimalist standards) that the place is not a constituent of
what is strictly and literally said: when I say 'It is raining' (rather than
something more specific like 'It's raining in Paris' or 'It's raining here'),
what I literally say is true iff it's raining somewhere or other. 6 That is
obviously not what I mean, since what I mean involves a particular place.
Appearances notwithstanding, the situation is similar to the case of 'I've
had breakfast', where a restricted time-interval is contextually provided for
pragmatic reasons, without being linguistically mandated.
Examples (4) and (5) are amenable to the same sort of treatment.
According to standard Russellian analysis, a definite description conveys
an implication of uniqueness: hence 'The table is covered with books' is
true iff there is one and only table and it is covered with books. To make
sense of this, we need either to focus on a restricted situation in which
there is indeed a single table, or to expand the predicate 'table' and enrich
it into, say 'table of the living-room' in order to satisfy the uniqueness

6 See Borg 2002 for a defense of that claim.

17
constraint. Either way, the form of enrichment through which we make
sense of the utterance is not linguistically mandated: it is only
pragmatically required. If we don't enrich, what we get is an already
complete proposition (albeit one that is pretty absurd): the proposition that
the only existing table is covered with books. Similarly with example (5):
without enrichment the utterance expresses a proposition that is true iff
every existing person went to Paris. Such a proposition is unlikely to be
true, but that does not make it incomplete. The enrichment process
through which, in context, we reach the proposition actually
communicated (to the effect that everybody in such and such group went
to Paris) is not linguistically but pragmatically required; hence it is not an
instance of saturation, but an optional process of 'free enrichment'. It
follows that, in those examples as much as in the previous ones, the
proposition literally expressed is different from, and more general than,
the proposition actually communicated.

1.4. A problem for Minimalism

In general, the literal truth-conditions posited as part of the Minimalist

analysis turn out to be very different from the intuitive truth-conditions
which untutored conversational participants would ascribe to the
utterance. This divergence between the intuitive truth-conditions of an
utterance and the literal truth-conditions postulated by the theorist is
particularly striking in connection with examples like (6). According to a
fairly standard view (Horn 1972), the proposition literally expressed by (6)
is the proposition that John has at least three children, i.e., no less than
three but possibly more. In certain contexts this corresponds to what the
speaker actually means (as when I say, 'If John has three children he can
benefit from lower rates on public transportation') but in other contexts
what the speaker means is quite different. Suppose for example that I am
asked how many children John has and that I reply by uttering (6). Clearly,
in this context, I mean that John has (exactly) three children — no more
and no less. This is standardly accounted for by saying that the proposition

18
literally expressed, to the effect that John has at least three children,
combines with the 'implicature' that John has no more than three children
(a generalized implicature that is accounted for in terms of the maxim of
quantity);7 as a result of this combination, what is globally communicated
— and what I actually mean — is the proposition that John has exactly
three children. Now this is the only proposition I am conscious of
expressing by my utterance; in particular, I am unaware of having
expressed the 'minimal' proposition that John has at least three children.
To account for this obvious fact, the minimalist claims that we are aware
only of what is globally conveyed or 'communicated' by the utterance.
Analysing this into 'what is literally said' and 'what is implied' is the
linguist's task, not something that is incumbent upon the normal language
user. Figure 1.1 illustrates this widespread conception.

[Figure 1.1 about here]

The problem with this conception is that it lacks generality. Recall

the example I gave earlier — the utterance 'I am French' used to convey
that I am a good cook. In the relevant situation of utterance, both the
speaker and the listener are aware that the speaker says he is French, and
thereby implies he is a good cook. This typical case of implicature is very
different from a case like (6) in which the speaker is not only (like the
hearer) unaware of the proposition literally expressed, but would strongly
deny having said what the minimalist claims was actually said.

7 As Grice puts in one of his early papers, "one should not make a weaker

statement rather than a stronger one unless there is a good reason for so

doing" (Grice 1961: 132). Since the statement that John has (at least)

three children is weaker than the statement that John has n children (for n

> 3), the maxim is obeyed only if John has no more than three children. (If

John has more than three children, the speaker should have made a

stronger statement.) The statement 'John has three children' therefore

implicates that John has no more than three children.

19
 It turns out that there are two sorts of case. On the one hand there
are prototypical cases of implied meaning, in which the participants in the
speech situation are aware both of what is said and of what is implied, and
also of the inferential connection between them. On the other hand, there
are the cases illustrated by (1)-(6). Given his willingness to treat certain
aspects of the intuitive meaning of (1)-(6) as conversational implicatures
external to what is literally said, the minimalist must explain why those
implicatures, unlike the prototypical cases (e.g. the French/cook example),
do not have the property of conscious 'availability'.
The only explanation I have come across in the literature makes use
of Grice's distinction between 'generalized' and 'particularized'
conversational implicatures, i.e. between implicatures which arise 'by
default', without any particular context or special scenario being
necessary, and those which require such specific contexts. In contrast to
the latter, the former are 'hard to distinguish from the semantic content of
linguistic expressions, because such implicatures [are] routinely associated
with linguistic expressions in all ordinary contexts' (Levinson 1983: 127).
Generalized implicatures are unconsciously and automatically generated
and interpreted. They belong to the 'micropragmatic' rather than to the
'macropragmatic' level, in Robin Campbell's typology:

A macropragmatic process is one constituted by a sequence of

explicit inferences governed by principles of rational cooperation. A
micropragmatic process develops as a cryptic [= unconscious] and
heuristic procedure which partially replaces some macropragmatic
process and which defaults to it in the event of breakdown.
(Campbell 1981: 101)

But there are problems with this explanation. According to Horn

(1992), the generalized nature of an implicature does not entail its
conscious unavailability — its 'cryptic' character. In other words, it is
possible for an implicature to be both 'generalized' and intuitively
accessible as an implicature distinct from what is said. Thus Horn insists

20
that the generalized scalar implicature from 'some' to 'not all' is
consciously available (in contrast to that from 'three' to 'exactly three'). A
speaker saying 'Some students came to the meeting' normally implies that
not all students came, and when this is so there is no tendency on the part
of the interpreter to conflate the implicature with what is said. This is
actually debatable, for the 'implicature' at issue can arise at sub-sentential
level (e.g. 'He believes some students came'), and in such cases there are
reasons to doubt that the availability condition is satisfied. Be that as it
may, the 'generalization' of an implicature does not seem to be necessary
for its unconscious character. Many particularized 'bridging' inferences are
automatic and unconscious. To take an example from Robyn Carston, 'He
went to the cliff and jumped' is readily interpreted as saying that the
person referred to jumped over the cliff, even though this is only
contextually suggested.

1.5. The availability of what is said

In earlier writings (Recanati 1989, 1993) I put forward a conception

diametrically opposed to that illustrated by Figure 1.1 above. 'What is
said', I held, is consciously available to the participants in the speech
situation (Figure 1.2).

[Figure 1.2 about here]

In this framework 'what is communicated' is not a distinct level where

'what is said' and 'what is implied' have been merged and integrated into
a unified whole; it is merely a name for the level at which we find both
what is said and what is implied, which level is characterized by conscious
accessibility. On this picture, there are only two basic levels: the bottom
level in which we find both the meaning of the sentence and the
contextual factors which combine with it to yield what is said; and the top
level at which we find both what is said and what is implied, both being
consciously accessible (and accessible as distinct).

21
 The availability of what is said follows from Grice's idea that saying
itself is a variety of nonnatural meaning. One of the distinguishing
characteristics of nonnatural meaning, on Grice's analysis, is its essential
overtness. Nonnatural meaning works by openly letting the addressee
recognize one's primary intention (e.g. the intention to impart a certain
piece of information, or the intention to have the addressee behave in a
certain way), that is, by (openly) expressing that intention so as to make it
graspable. This can be done in all sorts of ways, verbal or nonverbal. Even
if we restrict ourselves to verbal communication, there are many ways in
which we can mean things by uttering words. Saying is one way ; implying
is another.
The view that 'saying' is a variety of nonnatural meaning entails that
what is said (like what is meant in general, including what is implied) must
be available — it must be open to public view. That is so because
nonnatural meaning is essentially a matter of intention-recognition. On
this view what is said by uttering a sentence depends upon, and can
hardly be severed from, the speaker's publicly recognizable intentions.
Hence my 'Availability Principle' (Recanati 1993: 248), according to which
'what is said' must be analysed in conformity to the intuitions shared by
those who fully understand the utterance — typically the speaker and the
hearer, in a normal conversational setting.
I take the conversational participants' intuitions concerning what is
said to be revealed by their views concerning the utterance's truth-
conditions. I assume that whoever fully understands a declarative
utterance knows which state of affairs would possibly constitute a truth-
maker for that utterance, i.e. knows in what sort of circumstance it would
be true. The ability to pair an utterance with a type of situation in this way
is more basic than, and in any case does not presuppose, the ability to
report what is said by using indirect speech; it does not even presuppose
mastery of the notion of 'saying'. Thus the proper way to elicit such
intuitions is not to ask the subjects 'What do you think is said (as opposed
to implied or whatever) by this sentence as uttered in that situation'? 8 I

8 Michael Thau notes that speakers

22
therefore agree with Bach's criticism of the experiments through which
Gibbs and Moise (1997) attempted to support the availability-based
approach.9 However, Bach himself uses what he calls the 'IQ test' to
determine what is said, that is, he ties what is said to indirect speech
reports of what is said (Bach 1994a: 278, 1999, 2001). I find this procedure
most objectionable, and that is not what I mean when I claim that what is
said should be individuated according to the intuitions of normal
interpreters. Thus I strongly disagree with Cappelen and Lepore's
surprising statement:

We ourselves don't see how to elicit intuitions about what-is-said by

an utterance of a sentence without appealing to intuitions about the
accuracy of indirect reports of the form 'He said that...' or 'What he
said is that...' or even 'What was said is that...' (Cappelen and
Lepore, 'On an alleged connection between indirect speech and the
theory of meaning', Mind and Language 12: 278-296, 1997, p.280)

I find this statement surprising, because there obviously is another way of

almost never explicitly think about the distinction between what

they’ve said and what they’ve implicated. So the question of what a

speaker takes himself to have said by some utterance will have to

depend upon the answer he would give if he were asked. And it’s

very likely that in many circumstances there won’t be a single

answer, that the answer will differ depending on how the question is

put. It’s also very likely that the answer will vary from circumstance

to circumstance. (Consciousness and Cognition, Oxford University

Press, 2002, p. 148)

Contrary to what Thau thinks, however, this does not speak against the

availability-based approach. The speaker’s intuitions concerning what is

said need not involve the very notion of what is said.

9 K. Bach, 'You don't say', Synthèse 128: 15-44, 2001, p. 43.

23
eliciting truth-conditional intuitions. One has simply to provide subjects
with scenarios describing situations, or, even better, with — possibly
animated — pictures of situations, and to ask them to evaluate the target
utterance as true or false with respect to the situations in question. That
procedure has been used by several researchers to test speaker's
intuitions about e.g. the truth-conditions of donkey sentences. Thus Bart
Geurts describes his experimental set-up (inspired from earlier work by
Yoon) as follows:

Twenty native speakers of Dutch were asked to judge whether or not

donkey sentences correctly described pictured situations.
Instructions urged subjects to answer either true or false, but they
were also given the option of leaving the matter open in case they
couldn't make up their minds. (B. Geurts, 'Donkey Business',
Linguistics and Philosophy 25: 129-156, 2002, p. 135)

This procedure presupposes that normal interpreters have intuitions

concerning the truth-conditional content of utterances. On my view, those
intuitions correspond to a certain 'level' in the comprehension process — a
level that a proper theory of language understanding must capture. That is
the level of 'what is said' (as opposed to e.g. what is implied).

1.6. The availability-based approach

From a psychological point of view, we can draw a helpful parallel between

understanding what one is told and understanding what one sees. In
vision, the retinal stimuli undergo a complex (multistage) train of
processing which ultimately outputs a conscious perception, with the dual
character noted by Brentano: the subject is aware both of what he sees,
and of the fact that he is seeing it. Although more complex in certain
respects, the situation with language is similar. The auditory signal
undergoes a multistage train of processing which ultimately outputs a
conceptual experience: the subject understands what is said. This is very

24
much like (high-level) perception. If I am told that it is four o'clock, I hear
that it is four o'clock, just as, when I look at my watch, I see that it is four
o'clock. Like the visual experience, the locutionary experience possesses a
dual character: we are aware both of what is said, and of the fact that the
speaker is saying it.
In calling understanding an experience, like perception, I want to
stress its conscious character. Understanding what is said involves
entertaining a mental representation of the subject-matter of the
utterance that is both determinate enough (truth-evaluable) and
consciously available to the subject. This suggests a criterion, distinct from
the Minimalist criterion, for demarcating what is said. Instead of looking at
things from the linguistic side and equating 'what is said' with the minimal
proposition one arrives at through saturation, we can take a more
psychological stance and equate what is said with (the semantic content
of) the conscious output of the complex train of processing which underlies
comprehension.10
To be sure, that output itself is subject to further processing through
e.g. inferential exploitation. Consider, once again, vision. Seeing John's car,
I can infer that he is around. Similarly, hearing that John has had breafkast,
I can infer that he is not hungry and does not need to be fed. Just as what
is seen corresponds to the primary conscious output of visual processing,
not to what can be secondarily derived from it, 'what is said' corresponds
to the primary truth-evaluable representation made available to the
subject (at the personal level) as a result of processing the sentence. It is
therefore minimal in a certain sense, though not (as we shall see) in the
sense of Minimalism.
Accordingly, I distinguish between two sorts of pragmatic process.
The contextual processes which, like saturation, are (subpersonally)

10 As Ian Rumfitt once put it, "what is said in the course of an utterance is

nothing other than what somebody who understands the utterance

understands to be said" ("Content and Context: the Paratactic Theory

Revisited and Revised", Mind 102: 429-453, at p. 439).

25
involved in the determination of what is said I call primary pragmatic
processes. In contrast, secondary pragmatic processes are ordinary
inferential processes taking us from what is said, or rather from the
speaker's saying of what is said, to something that (under standard
assumptions of rationality and cooperativeness) follows from the fact that
the speaker has said what she has said. To the extent that the speaker
overtly intends the hearer to recognize such consequences as following
from her speech act, they form an integral part of what the speaker means
by her utterance. That is, roughly, Grice's theory of 'conversational
implicature'. An essential aspect of that theory is that the hearer must be
able to recognize what is said and to work out the inferential connection
between what is said and what is implied by saying it. Again, it follows that
what is said must be consciously available to the interpreter. It must
satisfy what I call the Availability constraint.
In this framework we solve the difficulty raised in section 1.5. We no
longer have two sorts of case of implicature — the prototypical cases
where the interlocutors are aware of what is said, aware of what is implied,
and aware of the inferential connection between them, and the cases in
which there is no such awareness. Conscious awareness is now a built-in
feature of both what is said and the implicatures. That is so because what
is said is the conscious output of linguistic-cum-pragmatic processing, and
the implicatures correspond to further conscious representations
inferentially derived, at the personal rather than sub-personal level, from
what is said (or, rather, from the speaker's saying what is said). The
alleged cases in which the speech participants themselves are not
distinctly aware of what is said and of what is implied are reclassified: they
are no longer treated as cases of 'implicature', strictly speaking, but as
cases in which a primary pragmatic process operates in the (sub-personal)
determination of what is said.

1.7. 'Saying' as a pragmatic notion

So far I have followed Grice, who construes saying as a variety of meaning.

26
But this pragmatic approach to 'saying' is controversial. Most philosophers
use the notion of 'what is said' (or 'the proposition expressed') in such a
way that it is not a 'pragmatic' notion — having to do with what the
speaker means or with what the hearer understands. What is said is
supposed to be a property of the sentence (with respect to the context at
hand) — a property which it has in virtue of the rules of the language.
Minimalism is closely associated with such a nonpragmatic way of
looking at what is said. In the minimalist framework, saturation is the only
contextual process allowed to affect 'what is said', because it alone is a
bottom-up process, i.e. a process triggered (and made obligatory) by a
linguistic expression in the sentence itself. 11 All other contextual processes
determine aspects of meaning external and additional to what is said.
Take, for example, 'free enrichment' — the process responsible for making
the interpretation of an utterance more specific than its literal
interpretation (as when 'jumped' is contextually understood as 'jumped
over the cliff'). That form of enrichment is 'free' in the sense of not being
linguistically controlled. Thus what triggers the contextual provision of the
relevant temporal restriction in example (1) is not something in the
sentence but simply the fact that the utterance is meant as an answer to a
question about the speaker's present state of hunger (which state can be
causally affected only by a breakfast taken on the same day). While
saturation is a bottom-up, linguistically controlled pragmatic process, free
enrichment is a top-down, pragmatically controlled pragmatic process.
Insofar as it is pragmatically rather than linguistically controlled, free
enrichment is taken to be irrelevant to 'what is said', on the nonpragmatic
construal of what is said.
I will discuss the nonpragmatic construal of what is said in Chapter 4.
For the time being, I'm interested in the pragmatic construal, based on
Grice's idea, and the reasons it provides for rejecting the Minimalist
constraint (§1.8). Before turning to that issue, however, I want to rebut a
couple of objections to the pragmatic construal.

11 As I pointed out in footnote 2, that is true even when saturation consists

in providing a so-called 'unarticulated constituent'.

27
 The first objection is this. If, following Grice, we construe saying as a
variety of meaning, we will be prevented from acknowledging an
important class of cases in which the speaker does not mean what he
says. Irony is a good example of that class of cases. If I say 'John is a fine
friend' ironically, in a context in which it is obvious to everybody that I
think just the opposite, it is clear that I do not mean what I say: I mean the
opposite. Still, I say that John is a fine friend. Grice's construal of saying as
a variety of meaning prevents him from acknowledging that fact.
According to Grice, when I say 'John is a fine friend' in the mentioned
situation, I do not really say that John is a fine friend — I pretend to be
saying it. The pragmatic construal of saying forces Grice to draw a
distinction between 'saying' and 'making as if to say'.
As far as I am concerned, I find Grice's distinction (between genuine
saying and making as if to say) perfectly legitimate, but I can understand
the worries of those who feel that the notion of 'saying' he uses is too
much on the pragmatic side. We certainly need a notion of 'what is said'
which captures the objective content of an utterance irrespective of its
pragmatic force as a serious assertion or as an ironical utterance. Still, I
find the objection superficial, for it is quite easy to actually construct the
desired notion within Grice's own framework. Grice uses 'say' in a strict
sense. In that sense whatever is said must be meant. But we can easily
define a broader sense for 'say':

S says that p, in the broad sense, iff he either says that p (in the
strict sense) or makes as if to say that p (again, in the strict sense of
'say').

I will henceforth use 'say' in that broad sense, which remains within the
confines of the pragmatic construal.
Another objection to the pragmatic construal focusses on the loss of
objectivity that allegedly goes with it. What is said is objective in the sense
that it is possible both for the speaker to make a mistake and say
something other than what he means, and for the hearer to misunderstand

28
what the speaker is saying. Those mistakes are possible, the objector will
argue, because what is said is an objective property of the sentence (in
context). But on the pragmatic construal, it is not clear that this objectivity
can be captured. Imagine the following situation: the speaker wants to say
that Paul is tall, and, mistaking Tim for Paul, says 'He is tall' while pointing
to Tim. The speaker thus inadvertenly says that Tim is tall. Now imagine
that the hearer also mistakes Tim for Paul. Thanks to this lucky mistake, he
grasps what the speaker means, thinking that this is what he has said. The
speaker and the hearer therefore converge on a certain interpretation,
which is not objectively what was said, but which they both (mistakenly)
think is what was said. How, in the framework I have sketched, will it be
possible to dissociate what is actually said from the protagonists' mistaken
apprehension of what is said? Have we not equated what is said with their
understanding of what is said?
We have not. We have equated what is said with what a normal
interpreter would understand as being said, in the context at hand. A
normal interpreter knows which sentence was uttered, knows the meaning
of that sentence, knows the relevant contextual facts (who is being
pointed to, etc.).12 Ordinary users of the language are normal interpreters,
in most situations. They know the relevant facts and have the relevant
abilities. But there are situations (as in the above example) where the
actual users make mistakes and are not normal interpreters. In such
situations their interpretations do not fix what is said. To determine what is
said, we need to look at the interpretation that a normal interpreter would
give. This is objective enough, yet remains within the confines of the
pragmatic construal.

1.8. Availability vs. Minimalism

In the framework I have sketched, there is a basic constraint on what is

said:

12 This is all tacit knowledge, not the sort of 'conscious awareness' I talk

about in connection with secondary pragmatic processes.

29
 Availability :
What is said must be intuitively accessible to the conversational
participants (unless something goes wrong and they do not count as
'normal interpreters').

This constraint leads us to give up Minimalism. That is the price to pay if

we want Availability to be satisfied.
The reason why Availability is incompatible with Minimalism is simple
enough. The aspects of the meaning of (1)-(6) which the minimalist
construes as conversational implicatures are, one may admit, contextual
ingredients in the overall meaning of the utterance. They do not belong to
the conventional meaning of the sentence. The minimalist claims that they
do not belong to 'what is said' either, because they are optional: those
contextual aspects of the meaning of the utterance are not necessary for
the latter to express a complete proposition. But the Availability Constraint
pulls in the other direction. The very fact that the minimal propositions
allegedly expressed are not consciously available shows that it would be a
mistake to equate them with what is said; rather, the Availability
Constraint dictates that the aspects of meaning which Minimalism
construes as external to what is said (e.g. the implicit reference to a place
in (3), or to the cut in (2), or to a time-interval in (1)) are actually
constitutive of what is said, because when we substract them from the
intuitive meaning of the utterance the proposition which results is no
longer something accessible to the participants in the speech situation.
Thus we have two quite distinct phenomena: examples like 'I am French'/'I
am a good cook' involve something which is said and whose saying implies
something else; examples like (1)-(6), in contrast, do not involve the
distinction between what is said and what is implied but a different
distinction between the literal meaning of the sentence and contextual
ingredients entering into the determination of what is said. If we maintain
that those ingredients are indeed 'optional' rather than necessary for
propositionality, this implies that we must give up the minimalist criterion

30
according to which the context contributes to what is said only when this
is necessary for some proposition to be expressed.
According to the view we arrive at, truth-conditional interpretation is
pragmatic to a large extent. Various pragmatic processes come into play in
the very determination of what is said; not merely saturation — the
contextual assignment of values to indexicals and free variables in the
logical form of the utterance — but also free enrichment and other
processes which are not linguistically triggered but are pragmatic through
and through. Figure 1.3 summarizes the contrast between the two
conceptions (Minimalism, and the Availability-based approach).

[Figure 1.3 about here]

According to the Availability-based approach, the crucial distinction

is not between mandatory and optional contextual processes, but between
those that are 'primary' and those that are 'secondary'. Primary pragmatic
processes include not only saturation, but also 'optional' processes such as
free enrichment. Independent evidence for their inclusion in this category
is provided by the fact that, in general, the notion of 'what is said' we
need to capture the input to secondary, inferential processes already
incorporates contextual elements of the optional variety. Consider
examples (1)-(6) once again. In each case we may suppose that the
speaker implies various things by saying what she does. Thus, by saying
that she's had breafkast, the speaker implies that she is not hungry and
does not want to be fed. By saying that the child is not going to die, the
mother implies that the cut is not serious; and so forth. Now those
implicatures can be worked out only if the speaker is recognized as
expressing the (nonminimal) proposition that she's had breakfast that
morning, or that the child won't die from that cut. Clearly, if the speaker
had had breafkast twenty years ago (rather than that very morning),
nothing would follow concerning the speaker's present state of hunger and
her willingness or unwillingness to eat something. The implicature could
not be derived, if what the speaker says was not given the richer,

31
temporally restricted interpretation. If therefore we accept the Gricean
picture, according to which 'what is said' serves as input to the secondary
process of implicature-generation, we must, pace Grice himself,
acknowledge the non-minimal character of what is said. This provides
some support to the Availability-based approach, as against Minimalism.

32
 Chapter 2

Primary pragmatic processes

2.1. Enrichment, loosening, and transfer

Secondary pragmatic processes are 'postpropositional'. They cannot take

place unless some proposition p is considered as having been expressed,
for they proceed by inferentially deriving some further proposition q (the
implicature) from the fact that p has been expressed. In contrast, primary
pragmatic processes are 'pre-propositional': they do not presuppose the
prior identification of some proposition serving as input to the process.
Another difference is the fact that secondary pragmatic processes are
conscious in the sense that normal interpreters are aware both of what is
said and of what is implied and are capable of working out the inferential
connection between them. Primary pragmatic processes are not conscious
in that sense. Normal interpreters need not be aware of the context-
independent meanings of the expressions used, nor of the processes
through which those meanings are enriched or otherwise adjusted to fit
the situation of use. Unless they are linguists or would-be linguists, they
are aware only of the output of the primary processes involved in
contextual adjustment.
Saturation is a primary pragmatic process. If the uttered sentence is
'She is smaller than John's sister', then in order to figure out what is said I
must (at least) determine whom the speaker refers to by the pronoun 'she'
and what the relevant relation is between John and the mentioned sister.
Were saturation a secondary pragmatic process, I would have to proceed
in reverse order, that is, to identify what is said in order to determine
those things.
Beside saturation, which is linguistically mandated (bottom-up),
there are, I claim, other primary pragmatic processes that are optional and
context-driven (top-down). The paradigm case is free enrichment,

33
illustrated by example (1):

(1) Mary took out her key and opened the door

In virtue of a ‘bridging inference’, we naturally understand the second

conjunct as meaning that Mary opened the door with the key mentioned in
the first conjunct ; yet this is not explicitly articulated in the sentence.
Insofar as the bridging inference affects the intuitive truth-conditions of
the utterance, it does so as a result of free enrichment. 13
In typical cases free enrichment consists in making the interpretation
of some expression in the sentence contextually more specific. This
process has sometimes been described in the literature as 'specifization'.
For example the mass term 'rabbit' will be preferentially interpreted as
meaning rabbit fur in the context of 'He wears rabbit' and as meaning
rabbit meat in the context of 'He eats rabbit' (Nunberg and Zaenen
1992).14 This not a matter of selecting a particular value in a finite set; with
a little imagination, one can think of dozens of possible interpretations for
'rabbit' by manipulating the stipulated context of utterance; and there is
no limit to the number of interpretations one can imagine in such a way.
Nor can the process of specifization be construed as linguistically
mandated, that is, as involving a hidden variable. Were it linguistically
mandated (bottom up), it would be mandatory, but it is not: In some
contexts the mass term 'rabbit' means nothing more than RABBIT STUFF
('after the accident, there was rabbit all over the highway').
Can free enrichment be equated with specifization, or are there
instances of free enrichment that are not cases of specifization? The
provision of (optional) unarticulated constituents is supposed to be a case
of free enrichment in which it is not the interpretation of some expression

13 The term ‘bridging inference’ was originally introduced by Herb Clark, a

pioneer of pragmatic studies, in the seventies. Example (2) is discussed in

Carston (1988).

14 The term 'specifization' comes from Bach. Nunberg & Zaenen use

'precision' in the same sense.

34
in the sentence that is enriched, but more globally the interpretation of the
sentence. In most cases, however, what can be done in terms of
unarticulated constituents can also be done in terms of specifization. We
can construe the implicit instrument in the second conjunct or (1) either as
an unarticulated constituent (corresponding to the unexpressed
prepositional phrase 'with the key'), or as an aspect of the interpretation of
the predicate 'open' resulting from specifization (the concept contextually
expressed by 'open' being the specific, ad hoc concept OPEN_WITH_KEY,
rather than the generic concept OPEN).15 The same options are presumably
available for dealing with the 'rabbit fur/meat' example. In such cases, I
will assume that there is a single form of free enrichment, that can be
handled in different frameworks — either in terms of specifization (ad hoc
concepts) or in terms of unarticulated constituents. Which framework we
choose to handle such cases depends upon extraneous considerations.
(For example, if we want to preserve the principle of compositionality,
we'd better opt for the specifization view which spares us the postulation
of unarticulated constituents.) Still, there is a type of case for which I think
we need the notion of unarticulated constituent and cannot make do with
specifization and ad hoc concepts: whenever the alleged unarticulated
constituent is the intended 'circumstance of evaluation', we can't deal with
it in terms of specifization or ad hoc concept. That type of case will be
discussed in Chapter 8.
Another issue regarding enrichment is whether or not it can be
described as 'strengthening' or logical enrichment (Recanati 1993: 261). I
think it can. A predicate has conditions of application, and enrichment
consists in restricting the application of a predicate by contextually
providing further conditions that are not linguistically encoded. Thus 'table'
has such and such conditions of application packed into the concept
TABLE, and through contextual enrichment the further condition

15 The useful notion of 'ad hoc concept', originally introduced by Barsalou

(19xx), now belongs to the toolkit of relevance theory (Sperber and Wilson

1998, Wilson and Sperber 2002, Carston 1998, 2002).

35
IN_THE_LIVING_ROOM is provided, which results in a restricted application.
Thus construed enrichment can account for the (so-called) contextual
restriction of quantifiers and for the interpretation of (so-called)
'incomplete' definite descriptions. ('All the books are on the table', where
a particular set of books and a particular table are in question). 16
The converse of enrichment is loosening (Cohen 19xx, Carston
1996). There is loosening whenever a condition of application packed into
the concept literally expressed by a predicate is contextually dropped so
that the application of the predicate is widened. An example is 'The ATM
swallowed my credit card'. There can be no real swallowing on the part of
an ATM, since ATMs are not living organisms with the right bodily
equipment for swallowing. By relaxing the conditions of application for
'swallow', we construct an ad hoc concept with wider application.

16 Stephen Neale objects to my notion of strengthening:

It is sometimes said that enrichment in Sperber and Wilson’s sense

involves strengthening in that the post-enrichment statement entails

the statement one might associate with the utterance prior to

enrichment. That this cannot be true in general is made clear by

cases involving (non-persistent) determiners like “every”, “no” and

“some”. If, when discussing a dinner party I attended last night, I

say, “everyone drank a lot of wine”, I am saying only that everyone

who attended the dinner party drank a lot of wine, and this does not

entail that everyone (unqualified) drank a lot of wine. (Neale

forthcoming, chapter 1)

But I had already disposed of that objection in Direct Reference:

"Strengthening, it may be argued, can be understood as operating locally.

For example, in the case of 'Every boy came', we might say that it is the

predicate 'boy' that is strengthened into 'boy in the class', rather than the

proposition 'Every boy came' into 'Every boy in the class came'." (Recanati

1993: 262-3)

36
 A third type of primary pragmatic process that is not linguistically
mandated (bottom up) but contextually driven is semantic transfer
(Nunberg 1979, 1995). In transfer the output is neither an enriched nor an
impoverished version of the concept literally expressed by the input
expression. It's a different concept altogether, bearing a systematic
relation to it. Thus 'parked out back' denotes either the property a car has
when it is parked out back, or a different property, namely the property a
car-owner has whenever his or her car has the former property ('I am
parked out back'). Arguably, 'parked out back' literally denotes the former
property, and comes to denote the latter property as a result of transfer
(Nunberg 1995). Similarly, the expression 'ham sandwich' in 'The ham
sandwich left without paying' arguably denotes, through transfer, the
derived property HAM_SANDWICH_ORDERER rather than the linguistically
encoded property HAM_SANDWICH.
A number of problems arise in connection with primary pragmatic
processes. Do we really need the four categories I have mentioned? It may
be argued that we need less categories, or that we need more. In
particular, if we construe enrichment not as logical strengthening but as a
quasi-syntactic process of 'expansion' through which a representation (or a
structured content) is made more complex by the addition of further
constituents, then it may be that we don't need transfer, nor even
loosening, but that we can handle everything in terms of that (little
constrained) notion of expansion. Be that as it may, I will not deal with
those problems, internal to the theory of primary pragmatic processes, in
this chapter. What I am concerned with is a more basic set of issues
pertaining to the very distinction between primary and secondary
pragmatic processes.

2.2. Rejecting the Gricean picture

Even though he construed saying as a variety of meaning nn, Grice

espoused Minimalism. On his view, disambiguation and saturation suffice
to give us the literal interpretation of the utterance — what is literally said.

37
All other pragmatic processes involved in the interpretation of the
utterance are secondary and presuppose the identification of what is said.
Interpretation is construed as a two-step procedure: (i) The interpreter
accesses the literal interpretations of all constituents in the sentence and
uses them to compute the proposition literally expressed, with respect to
the context at hand; (ii) on the basis of this proposition and general
conversational principles he or she infers what the speaker means (which
may be distinct from what is said, i.e. from the proposition literally
expressed).
The picture I have presented also makes interpretation a two-step
procedure, but there is a major difference. The primary pragmatic
processes that are involved in determining what is said include not only
saturation (and disambiguation) but also optional processes such as free
enrichment, loosening and semantic transfer. Those processes take us
from the literal meaning of some constituent (the meaning that is
linguistically encoded, or that which results from saturating the
linguistically encoded meaning) to a derived meaning which may be
richer, poorer, or involve some kind of transfer. I hold that, for such
processes to take place, there is no need to antecedently compute the
proposition literally expressed. That is why I take those processes to be
'primary', like saturation. What I am disputing therefore is the claim that
the process of semantic composition which consists in putting together the
semantic values of the parts to determine the semantic value of the whole
begins by paying attention only to literal semantic values (as delivered
through disambiguation and saturation), and turns to derived values only
after the literal semantic value of the whole (the proposition literally
expressed) has been computed.
There is a simple argument purporting to show that the Gricean
picture must be right. Enrichment, loosening and transfer all take as input
the literal meaning of some expression; hence they cannot take place
unless that literal meaning has been accessed. This seems to support the
Gricean picture, according to which we process the literal interpretation
first, and move on to the derived interpretation only when this is required

38
to make sense of the speaker's utterance.
This argument is fallacious (Recanati 1994b, 1995). I admit that the
literal interpretation must 'come first', insofar as the derived interpretation
is derived from it through enrichment, loosening, or transfer. But I want to
resist the conclusion that the literal interpretation must be 'processed'
first, in the sense that is relevant to the debate. Or, to put it another way:
What I am rejecting is not the claim that the literal interpretation of the
constituent is accessed before the derived interpretation - that I take to be
obvious - but the claim that a similar priority holds at the level of the
complete sentence; that is, I reject the claim that the process of semantic
composition begins by paying attention only to literal semantic values, and
turns to derived values only after the literal semantic value of the whole
(the proposition literally expressed) has been computed. It is this picture
which I think is unwarranted.
If I am right, the asymmetric dependence of derived meaning upon
literal meaning does not rule out an account according to which literal
meaning and derived meaning are on equal footing in terms of semantic
composition. Thus we can imagine that the literal meaning and the derived
meaning of a given expression are processed in parallel, in constructing an
interpretation for the whole utterance. In the model I have in mind, the
literal meaning of the expression is accessed first and triggers the
activation of associatively related representations. That literal meaning is
a natural candidate for the status of semantic value, but there are others:
some of the representations activated by association contribute further
candidates for the status of semantic value. All candidates, whether literal
or derived, are processed in parallel and compete. When an interpretation
which fits the broader context of discourse is found, it is selected and the
other candidates are suppressed.
On this view, derived meanings still proceed (associatively) from
literal meanings, which they indeed presuppose; but, although generated
serially, they are processed in parallel. The literal meaning has no
compositional privilege over derived meanings; they compete and it is
possible for some derived meaning to be retained (if it fits the broader

39
context of discourse) while the literal interpretation is suppressed. In other
words, the derived interpretation is associatively derived from the literal
interpretation, but it is not inferentially derived. Inferential derivation
entails computation of the literal value of the global sentence, while
associative derivation is a 'local' process which does not require prior
computation of the proposition literally expressed (Recanati 1993:263-6).
Consider, as an example, Geoff Nunberg's famous ham sandwich
(Nunberg 1979). The waiter says 'The ham sandwich has left without
paying'. On the Gricean picture the interpreter computes the proposition
literally expressed by the sentence — namely the absurd proposition that
the sandwich itself has left without paying — and from its absurdity infers
that the speaker means something different from what she says. On the
parallel model I have outlined the description 'the ham sandwich' first
receives its literal interpretation, in such a way that a representation of a
ham sandwich is activated; activation then spreads to related
representations, including a representation of the man who ordered a ham
sandwich.17 All these representations activated by the description 'the ham
sandwich' contribute potential candidates for the status of semantic value
of the expression; all of which are equally susceptible of going into the
interpretation of the global utterance. Now the ham sandwich orderer is a
better candidate than the ham sandwich itself for the status of argument
for '... has left without paying'. It is therefore the derived, non-literal
candidate which is retained, while the literal interpretation is discarded.
An important difference between the Gricean model (according to
which the literal interpretation is processed first) and the parallel model

17 As Sag (1981) and Nunberg (1995) pointed out, it is not the description

as a whole, but the predicate 'ham sandwich' in the description which has

a derived, non-literal value; the ham sandwich example is a case of

property transfer, not a case of deferred reference (see footnote 17). Since

nothing hinges on this point, I will simplify matters in this and the next two

sections by treating the description itself as a unit susceptible to both a

literal and a derived interpretation.

40
just outlined is this: on the parallel model is it possible for an utterance to
receive a non-literal interpretation without the literal interpretation of that
utterance being ever computed. The non-literal interpretation of the global
sentence does not presuppose its literal interpretation, contrary to what
happens at the constituent level. If the non-literal interpretation of some
constituent fits the context especially well it may be retained (and the
other interpretations suppressed) before the literal interpretation of the
sentence has been computed. Whether or not this sort of thing actually
happens, this is at least conceivable, on the parallel model.

2.3. Accessibility

When someone talks of 'wearing rabbit', the literal meaning of the mass
term 'rabbit' (namely rabbit stuff) is accessed, but it has to compete with
other candidates for semantic value. The more specific representation
rabbit fur is also activated since it is associatively connected to the
representations encoded by both 'rabbit' and 'wear'. As a result of this
multiple activation, it is possible for the representation rabbit fur to be
more active, in this context, than the less specific representation rabbit
stuff which is linguistically encoded. Whatever we think of this particular
example, it seems to me that the following situation can obtain: An
expression linguistically encodes a certain representation; that
representation becomes active when the expression is uttered, but
another, associatively related representation is also activated (in part - but
in part only - through the encoded representation) and turns out to be
more active in that context than the original representation from which (in
part) it derives.18 If we assume that the candidate for semantic value which

18 I assume that representations can be activated because they are being

processed (or have been recently), or because they receive activation from

associatively related representations which are themselves activated.

Frequency of processing is another important factor (among several

others); it can be conceived of as lowering the activation treshold of a

41
is retained and undergoes semantic composition is that which is most
accessible, i.e. that which corresponds to the most active representation
when the interpretation process stabilizes, we explain how a derived
meaning resulting from enrichment (or loosening, or transfer) can be
selected as semantic value, in lieu of the literal meaning from which it is
derived.
The same phenomenon can happen dynamically. Some
representation may be activated through its associative links to the
representation linguistically encoded and become more active than the
latter as a result of the coming into play of further linguistic material which
raises its activation level. Before considering such a case, let us look at a
simpler example of accessibility shift along the temporal dimension.
Consider sentence (2) from Recanati 1993: 265:

(2) John was arrested by a policeman yesterday; he had just stolen a

wallet.

In order to interpret the utterance one must assign a reference to the

pronoun 'he' in the second sentence. The two persons who were
mentioned in the first sentence, namely John and the policeman, are
obvious candidates, and we may suppose that there is no other
(sufficiently accessible) candidate. (There would be one if, for example,
the speaker pointed to someone while uttering the pronoun.) We may
suppose that when the pronoun 'he' is uttered one of the two candidates,
John or the policeman, is more accessible than the other; John because he
is foregrounded (qua subject of the sentence), or the policeman because
he was mentioned last. For the sake of the argument, let us suppose that
the second factor is more important than the first one, so that the
policeman is slightly more accessible than John when the pronoun is
uttered. This may well change when the other constituents of the sentence
are processed. When the predicate 'had just stolen a wallet' is uttered,
John becomes more accessible than the policeman as a candidate for the

representation. See e.g. Barsalou and Billman (1989).

42
status of referent of 'he', even if the policeman was more accessible at an
earlier stage in the processing of the utterance. For John is the subject of
'was arrested' and therefore occupies the role of the person being
arrested; now that role is linked to the role of the person doing the
stealing, in some relevant frame (on frames, see e.g. Fillmore and Atkins
1992). Because of this link, the representation of the referent of 'he' as the
person doing the stealing contributes some activation to the
representation of the person being arrested and therefore raises the
accessibility of John qua occupier of this role. John thus becomes the most
accessible candidate.
If we turn to a case in which some candidate is distinguished from
the others by its being literal, we see that the same sort of temporal shift
in accessibility can occur. When the words 'the ham sandwich' are uttered,
we may consider that the representation of the ham sandwich is more
active than other, related representations which are activated through
their links to that representation. Thus we may suppose that the
representation of the ham sandwich is more active than the 'derived'
representation of the ham sandwich orderer. This is similar to the fact that
the policeman is more accessible than John when the pronoun 'he' in (2) is
uttered, except that the policeman is (perhaps) more accessible because
he was mentioned last, while the representation of the ham sandwich is
more active because it is linguistically encoded and has some form of
priority over the ham sandwich orderer (derived value). In both cases, the
initial ranking is reversed when further linguistic material comes into play.
After the predicate in the sentence 'The ham sandwich has left without
paying' has been processed, the ham sandwich is no longer a more
accessible candidate than the ham sandwich orderer — the order of
accessibility is reversed. The explanation, again, is very simple and does
not appeal to inference on the hearer's part. The predicate 'has left
without paying' demands a person as argument; this raises the
accessibility of all candidates who are (represented as) persons. In this
way the representation of the ham sandwich orderer gains some extra
activation which makes him more accessible than the ham sandwich, after

43
the predicate has been processed.

2.4. Objections and responses

In the framework I have sketched the interpretation which eventually

emerges and incorporates the output of various pragmatic processes
results from a blind, mechanical process, involving no reflection on the
interpreter's part. The dynamics of accessibility does everything, and no
'inference' is required. In particular, there is no need to consider the
speaker's beliefs and intentions.
Dan Sperber objected to me in conversation that accessibility can
lead us astray: Sometimes the first interpretation that comes to mind (the
most accessible one) turns out not to be satisfactory and forces the hearer
to backtrack. According to Sperber, the possibility of such garden-path
effects shows that success, for a candidate semantic value, cannot be
equated with sheer accessibility. This objection is misguided, I think. The
most accessible interpretation at some stage s in the interpretation
process may well turn out to be unsatisfactory at some later stage s',
thereby resulting in a garden path effect and the need to backtrack. This
does not show that interpretational success cannot be cashed out in terms
of accessibility. At any given stage, the most accessible interpretation will
be the winning one (at that stage). In garden path utterances we have two
successive stages to consider. Some interpretation is the most accessible
one, hence wins, at s, but that interpretation fails to fit some schema,
hence loses, at a later stage s'. In an accessibility-based framework, this
means that this interpretation's accessibility at s' is no longer sufficient for
it to be the winning candidate (at s'). Another candidate (which was less
accessible at s, but turns out to be more accessible at s') takes over,
hence the garden path effect. The distinction between successive stages
of interpretation, together with the notion of an accessibility shift, is
sufficient to account for garden path effects within the accessibility-based
framework.
The notion of an accessibility shift also provides immediate answers

44
to the questions which, according to Kent Bach (1994b), my account of the
ham sandwich metonymy and similar phenomena in terms of local
pragmatic processes raises. Bach writes:

Recanati supposes that the process of metonymical transfer takes

place without the intrusion of a thought of the absurd proposition
associated with the literal meaning of (i) ['The ham sandwich is
getting restless']. That is, the hearer does not have to compute that
the speaker does not mean that a certain culinary item is getting
restless in order to determine what the speaker does mean. But, I
ask, how can the hearer go from the concept of ham sandwich to that
of ham-sandwich-orderer without first entertaining the absurd
minimal proposition? What triggers the 'local process' and, for that
matter, keeps it from being triggered in a cases like (ii) ['The ham
sandwich is getting eaten'], uttered in similar circumstances?
Recanati's account predicts that the hearer would entertain the
proposition that the ham-sandwich-orderer is getting eaten, since the
local process it posits would get triggered before a full proposition
were reached. And yet (ii) could be understood perfectly well. (Bach
1994b: 158)

But my account does not predict that the hearer would understand (ii)
('The ham sandwich is being eaten') as being about the ham sandwich
orderer. There is something, in my account, which keeps the local process
of metonymy from being triggered in a case like (ii). What keeps the local
process of metonymy from being triggered in a case like (ii) is precisely
what answers Bach's legitimate question: 'How can the hearer go from the
concept of ham sandwich to that of ham-sandwich-orderer without first
entertaining the absurd minimal [i.e. literal] proposition?' In my
framework, the process of metonymical transfer is 'local', not global, yet it
is sensitive to the linguistic (and extralinguistic) context in which the
expression which receives the metonymical interpretation occurs. In
particular, it is sensitive to the meaning of the predicate expression; this is

45
why it is triggered in (i) but not in (ii).
For the literal proposition to be entertained, the literal semantic
value of 'the ham sandwich' must undergo semantic composition with the
semantic value of the predicate; but the semantic value of the predicate
cannot be accessed without making the metonymical interpretation of the
subject-term more accessible than its literal interpretation, hence it
prevents the latter from going into the proposition expressed. The
activation of the concept which corresponds to the literal interpretation of
the predicate is both what triggers the process of metonymical transfer in
(i) and what prevents it from taking place in (ii). Yet the transfer's
sensitivity to the interpretation of the other constituents in the sentence
does not mean that the transfer is a global process involving the
computation of the absurd, literal proposition. The interpreter does not go
from the concept of ham sandwich to that of ham-sandwich-orderer after
having entertained the absurd literal proposition; rather, it is because the
interpreter goes from the concept of ham sandwich to that of ham-
sandwich-orderer (as a result of an accessibility shift resulting from the
interpretation of the predicate) that he or she does not entertain the
absurd literal proposition.
In this account, the metonymical transfer from ham sandwich to
ham-sandwich-orderer results from a shift in accessibility triggered by the
interpretation of another constituent in the sentence. The literal
interpretation of the subject-term was more accessible than the
metonymical interpretation before the predicate came into the picture, but
the metonymical interpretation becomes more accessible as a result of
interpreting the predicate. This accessibility shift has two consequences.
First, the literal value of the subject term in (i) does not undergo semantic
composition with that of the predicate, precisely because the semantic
value of the predicate makes the literal value of the subject term less
accessible than some other, non-literal interpretation. Second, if the
sentence contained another predicate which did not require a person, but
rather a culinary item, as argument, then the accessibility shift would not
occur and the description 'the ham sandwich' would be given the literal

46
interpretation; this account for the difference between (i) and (ii).

2.5. Interactive processing

So far I have considered the interpretation of a given constituent as if the

interpretations of the other constituents were fixed; but nothing is fixed.
All constituents can be given derived as well as literal interpretations, and
how we interpret a constituent cannot but affect how we interpret the
others. Consider (3), for instance.

(3) The city is asleep.

A city is not the sort of thing that sleeps (in the normal sense of the term),
hence the overall interpretation of (3) is likely to involve some process of
non-literal interpretation. There are various possibilities, though. 'The city'
can be interpreted metonymically as standing for 'the inhabitants of the
city' (transfer); or 'asleep' can be interpreted metaphorically in the sense
of 'quiet and showing little activity' (loosening). Given the likelihood of a
non-literal interpretation for some constituent, if 'the city' literally applies
to a city, 'asleep' will have to be taken non-literally; conversely, if 'asleep'
is literal, 'the city' will not be.
The same sort of trade off applies to seemingly unproblematic
examples such as the following:

(4) I finished the book.

(5) John heard the piano.

If 'finish' has its standard sense in (4), its object (what is said to be
finished) must be a process, for only processes can start or finish in the
standard sense (Pustejovsky 1991). It follows that 'the book' must be
interpreted non-literally — it must be interpreted as standing, not for a
certain book, but for the process of, say, reading a certain book.
Conversely, if 'the book' is interpreted literally and stands for a certain

47
book, 'finish' must have a derived sense and mean something like finished
reading. Pustejovsky (1991) defends the former analysis, Langacker (1984)
the latter.19
(5) is another example (discussed by Langacker) which can be
interpreted in two ways. We may construe 'the piano' as standing for the
sounds emitted by the piano (metonymy), in accordance with the
'postulate' that only sounds can be heard; or we may consider 'hear' as
polysemous. When it is said that 'only sounds can be heard', 'hear' is taken
in its basic sense ('hear'1), but there is another sense ('hear'2), which can
be defined as follows: An object is heard2 whenever the sound it emits is
heard1. Langacker analyses (5) as saying that John heard 2 the piano,
rather than as saying that he heard1 the sounds emitted by the piano.
Though (3), (4) and (5) exhibit the same sort of trade off, there is an
important difference between (3), on the one hand, and (4) and (5) on the
other hand. Depending on which constituent receives a non-literal
interpretation, the global interpretation of (3) itself will vary. In one case
the speaker means that the inhabitants of the city are sleeping, in the
other she means that the city itself is quiet and shows little activity; the
truth-conditions are distinct. It follows that it is easy to say which
constituent is responsible for the non-literal interpretation of the sentence:
we have only to consult our intuitions concerning the truth-conditions of
the utterance. (4) and (5) are very different, however. Whether one
constituent or the other bears the onus of non-literal interpretation, the
overall interpretation of the utterance does not seem to change. Hence it
is not easy to establish which analysis is right, and one must appeal to
indirect evidence, such as, for example, the fact that we can say things
like: 'I can both hear and touch the piano'. (This fact shows that Langacker
is right: when we say that we hear the piano, we are referring to the piano.

19 The sentence may also be seen as elliptical and contextually 'expanded'

into 'I finished reading the book' (Bach 1987, 1994a and b). I find this sort

of analysis too much unconstrained (Recanati 1993:261ff), so I don't take

it into consideration here.

48
It is the verb, not the noun-phrase, which has a derived sense in this
context.)20
The trade off talked about in this section is a particular case of a
more general and fairly complex phenomenon: the search for coherence in
interpretation. We have two constituents, A and B, each with (at least) two
possible interpretations (Ai and Aj, and Bi and Bj). Now there is a better 'fit'
between Ai and Bj or between Aj and Bi than there is between either Ai
and Bi or Aj and Bj, hence the 'trade-off' effect: the two global

20A number of similar tests are offered in Nunberg 1993 and 1995. For

example, Nunberg argues that (a) and (b) must be analysed in fairly

different ways, despite their superficial similarity.

(a) I am parked out back.

(b) This one <pointing to a car key> is parked out back.

(b) is a case of 'deferred ostension': an object is 'demonstrated' (the key),

but the ultimate 'referent', as opposed to the 'demonstratum', is another

object suitably related to the demonstratum, namely the car itself. In

contrast, 'I' does not refer to the car in (a); it normally refers to the

speaker. It's the predicate 'parked out back' that bears the onus of non-

literality. That the referent is the car rather than the (demonstrated) key in
(b) is established by the fact that semantic features of the referring

expression such as number or gender do NOT correspond to the key, but

to the car; thus we would say 'this one' (in the singular) even if we were

exhibiting a whole set of keys for the car in question. Also, we can say

things like:

(c) This <pointing to a car key>, which is an old Chevrolet, is parked out

back.

The relative clause clearly qualifies the car, not the key. These tests

provide evidence that the referent is the car, even if it is the key which is

'demonstrated'. When we turn to (a), the situation is totally different.

Gender and number correspond to the speaker, not to the car. We would

49
interpretations <Ai, Bj> and <Aj, Bi > are favored, in such a way that the
choice of either Aj or Ai as the correct interpretation for A 'coerces' a
particular interpretation (Bi or Bj) for B. The tendency to prefer coherent
interpretations (with a high degree of fit between the various semantic
values) is what we must now try to account for. But first, we must
introduce the notion of a schema (Rumelhart 1979, 1980; Langacker
1987).

2.6. The role of schemata

To say that two semantic values  and  'fit together' is to say that there is
an abstract schema which <, > instantiates. The particular case
discussed above, 'coercion', can be analysed as follows. An expression E
activates an abstract schema in which there is a slot for a value of a
certain type; as a result, the semantic value of E will preferably enter into
composition with a semantic value of the relevant type. Thus in a sentence
like 'the city is asleep', if we give to 'asleep' its literal value (thereby
activating the SLEEP schema), the value of 'the city' will have to be of the
relevant type (e.g. human or animal), hence non-literal.
The role of schemata in interpretation is best seen in connection with
more complex examples. Consider example (2) again:

(2) John was arrested by a policeman yesterday; he had just stolen a

wallet.

As I said earlier, John is selected as the referent for 'he' in the second
clause because (i) the referent of 'he' is said to have stolen and John is
known to have been arrested, and (ii) there is a frame or schema in which

not say 'We are parked out back' if there was a single car owner and

several cars (whereas we can say ' those <pointing to a key> are parked

out back' if there is a single key and several cars). And we can't say 'I,

who am an old Chevrolet, am parked out back'.

50
the two roles (stealing, and being arrested) are linked. This schema is
jointly activated by the predicates 'was arrested' and 'had stolen'. An
interpretation in which the same person steals and is arrested (and in
which he is arrested because he has stolen - see below) satisfies the
schema, and is more likely to be selected than one which violates it.
Given the link we have established between coherence (or fit) and
schemata, the question we must answer becomes: Why are schema-
instantiating interpretations more successful than others? In line with
what has been said so far, the following answer suggests itself.
Interpretational success - what brings a 'candidate' or potential semantic
value into the actual interpretation of the utterance - to a large extent
depends on the candidate's accessibility or degree of activation. Now a
schema is activated by, or accessed through, an expression whose
semantic value corresponds to an aspect of the schema. The schema thus
activated in turn raises the accessibility of whatever possible semantic
values for other constituents of the sentence happen to fit the schema.
The schema itself gains extra activation from the fact that some other
constituent of the sentence has a possible interpretation which fits the
schema. In such a case all 'candidates' or potential semantic values which
fit the schema evoked by some of them mutually reinforce their
accessibility and therefore increase the likelihood that they will be globally
selected as part of the interpretation of the utterance. Coherent, schema-
instantiating interpretations therefore tend to be selected and preferred
over non-integrated or 'loose' interpretations. As a result, schemata drive
the interpretation process.
The role played by schemata explains why the process of utterance
interpretation is to such a large extent top down and driven by world
knowledge. The interpreter unconsciously enriches the situation described
by the utterance with many details which do not correspond to any aspect
of the uttered sentence but are contributed in order to fit an evoked
schema.21 Thus, in the policeman example above, not only is the reference

21 Think of an example like (6):

(6) John hates the piano.

51
of 'he' determined by the Steal (x) - Is arrested (x) schema, but (among
other things) the causal interpretation of the relation between the two
events mentioned - the fact that the referent was arrested because he had
stolen - is also determined by the schema.

Contrary to (4)-(5), this example cannot be described in terms of

'coercion'. I mentioned earlier various contentions to the effect that only

sounds can be heard and only processes finished. By virtue of these

principles, some process of non-literal interpretation must take place in

order to make sense of phrases such as 'finish the book' or 'hear the

piano'. But a piano is certainly an object that can be hated, however

strictly one construes the predicate 'hate'. Still, some contextual

enrichment is in order, because to hate the piano is to hate it under some

aspect or dimension. One may hate the sounds emitted by the piano, or

one can hate playing the piano, or one can hate the piano as a piece of

furniture, etc. The relevant dimension is contextually provided through the

process of enrichment.

52
 Chapter 3

Relevance-theoretic objections

3.1. One or two systems?

There is much that is common between the view so far presented and
‘relevance theory’ (Sperber and Wilson 1986, Carston 2002b). The
emphasis upon the psychological dimension of utterance interpretation,
and the rejection of Minimalism, are among the important features that
the two frameworks share. But there are also some differences.
One difference that has been recently the focus of attention on the
part of relevance theorists is this. According to me, the primary pragmatic
processes involved in comprehension are not 'inferential' (Recanati
2002b). Only when the unreflective, normal process of interpretation yields
weird results does a genuine inference process take place whereby we use
evidence concerning the speaker's beliefs and intentions to work out what
he means. There is no doubt that our ability to do so is an important part
of our conversational competence, but the question at issue is: How
essential is that inferential ability? Can linguistic communication proceed
without it, at some basic level, or is it from the very start constituted by it?
Following Millikan and Burge, I reply that communication is not
constitutively inferential. As Burge writes,

We seem normally to understand content in a way whose

unconscious details (…) are not accessible via ordinary reflection. To
be entitled to believe what one is told, one need not understand or be
able to justify any transition from perceptual beliefs about words to
understanding of and belief in the words' content. One can, of course,
come to understand certain inferences from words to contents. Such
empirical meta-skills do enrich communication. But they are not
indispensable to it. (Burge 1993: 477)

53
 In my framework, however, conversational implicatures are
inferentially derived from premisses concerning the speaker's intentions in
saying what he says. They are arrived at by answering questions such as,
'Why is the speaker saying what he says?' Instead of merely retrieving
what is said through the operation of unconscious, primary pragmatic
processes, we reflect on the fact that the speaker says what he says and
use that fact, together with background knowledge, to infer what the
speaker means without saying it. The retrieving of conversational
implicatures therefore involves reflective capacities that are not exercised
in what Millikan calls 'normal language flow'.
As I emphasized in chapter 2, the contrast between primary and
secondary pragmatic processes corresponds to different levels of
processing. The determination of what is said takes place at a sub-
personal level, like the determination of what we see. But the
determination of what the speaker implies takes place at the personal
level, like the determination of the consequences of what we see.
Relevance theorists do not accept my distinction between the two
sorts of process. They advocate a unified view, according to which « the
various different pragmatic tasks are performed by processes that
comprise a single system, which takes decoded linguistic meaning as its
input and delivers the propositions communicated » (Carston 2002a: 141).
And they insist that the processes in question are uniformly inferential,
though in a sense that does not entail reflectiveness.
According to relevance theorists, there is nothing 'special' about
implicatures. They accept that in some cases a conscious, explicit process
of inference takes place, which disrupts normal language flow; 22 but it

22 "There clearly are times at which the normal communicative flow is

disrupted: certain instances of garden-path utterances, especially when

exploited by speakers for particular, often humorous, effects; some cases

of complex figurative use which require an effortful conscious search for

interpretation; other cases where there is some apprehended difficulty in

54
would be a mistake, they argue, to think of implicature derivation on this
model. Implicature derivation is as automatic and unconscious as the
other processes I treat as 'primary', viz. enrichment or loosening. In other
words, there are two forms of inference: a spontaneous, automatic,
unconscious form of inference which is involved both in explicature and
implicature derivation, and an effortful, conscious form which is involved
only in cases in which something goes wrong and there is disruption of
normal language flow. As Carston puts it,

The appropriate distinction within modes of processing and levels of

explanation would seem to be between, on the one hand, a modular
(sub-personal) pragmatic processor which, when all goes well,
quickly and automatically delivers speaker meaning (explicatures
and implicatures), and, on the other hand, processes of a conscious
reflective (personal-level) sort which occur only when the results of
the fomer system are found wanting' (Carston 2002a: 146).

In what follows I will discuss the arguments offered by relevance theorists

in support of their position. This will enable me to clarify my view and to
clear up some misunderstandings.

3.2. Personal and subpersonal inferences

Communication, in its most basic form, is non-inferential in the sense that

it does not rest on reasoning. Reid defines 'reasoning' as

the process by which we pass from one judgment to another which is

the consequence of it. Accordingly, our judgments are distinguished
into intuitive, which are not grounded upon any preceding judgment,
and discursive, which are deduced from some preceding judgment
by reasoning (Reid, 1969: 710).

satisfying oneself that the intended interpretation has been reached (it

doesn't seem sufficiently relevant, for instance)" (Carston 2002a: 145).

55
Evidently, comprehension is intuitive rather than discursive, except in the
special situations in which one adopts a reflective stance towards the
ongoing discourse. Normally we do not have to reason to understand what
the others are saying: the judgment that the speaker has said that p is
made directly upon hearing the utterance, without being inferentially
grounded in some prior judgment to the effect that the speaker has
uttered sentence S.
Relevance theorists do not deny that comprehension is intuitive
rather than discursive. When they say that communication is
fundamentally inferential, they mean it in a sense which is compatible with
its being intuitive. According to them, there are two sorts of inference. One
is conscious, explicit inference — what Reid calls 'reasoning'. Let us call
that 'inference in the narrow sense'. But there is also inference in the
broad sense: a type of inference which may well occur unconsciously, in
such a way that the judging subject is aware only of the conclusion of the
inference (which is, therefore, available not as the conclusion of an
inference but as an immediate, intuitive judgment).
Relevance theorists admit, indeed emphasize, that the inferential
procedure underlying ordinary understanding is unconscious. As Sperber
says, « when most of us talk of reasoning, we think of an occasional,
conscious, difficult, and rather slow mental activity. What modern
psychology has shown is that something like reasoning goes on all the
time — unconsciously, painlessly, and fast » (Sperber 1995: 195). It is in
that modern psychological sense that communication is said to be
'inferential'.
That there is such a broad notion of inference in contemporary
cognitive science is beyond question. For example, Marr writes that « the
true heart of visual perception is the inference from the structure of an
image about the structure of the real world outside » (Marr 1982: 68). That
is a very broad sense of 'inference' indeed. Nothing prevents the input to

56
an inference in that sense from being nonconceptual. 23 Sperber and Wilson
use ‘inference’ more restrictively. For there to be inference, they claim, the
transition must be from a conceptual representation to a conceptual
representation. Moroever, the transition must be truth-preserving. Only if
these conditions are satisfied will a cognitive transition count as a genuine
'inference'.24 According to Sperber and Wilson, pragmatic inferences are
genuine inferences by this criterion.25 Still, they are not inferences in the
narrow sense – explicit, conscious inferences conducted at the personal
level. They are inferences only in the broad, psychological sense. That is
true whether we talk of the inferences which underlie my 'primary
pragmatic processes' (enrichment, saturation etc.) or of those through
which conversational implicatures are generated.
In response, I wish to maintain that implicature generation rests on a
process of inference in the narrow sense: a process of conscious inference,
corresponding to Reid's description of 'reasoning'. But the notion of
conscious inference that is relevant here does not quite correspond to the
description given by relevance theorists. Sperber contrasts explicit,
conscious reasoning, which is a voluntary, effortful and slow activity, with
spontanous inference, which is effortless, fast, and takes place
unconsciously. That distinction I find misleading in the present context, for

23 Representations of intensity changes in terms of zero-crossings are

clearly nonconceptual (Crane 1990), yet Marr describes the transition from

such proximal representations to distal representations of edges as an

inference. Fodor (1983) uses ‘inference’ in the same broad way.

24 "[Whenever] the input and the output of a [computation] are not

related as premise and conclusion in an argument, the computation is not

interpretable as inferential"(Sperber and Wilson 1987: 137).

25 This is debatable. In 'Does linguistic communication rest on inference?'

(Mind and Language 17, 2002), pp. 122-3, I have questioned the claim that

primary pragmatic processes can be described as truth-preserving

operations on conceptual representations.

57
it mixes distinct issues. Like Sperber, I think that inference in the narrow
sense is essentially conscious: it takes place only if one judgment (the
conclusion) is grounded in another judgment (the premise), and if both
judgments, as well as the fact that one is grounded in and justified by the
other, are available (consciously accessible) to the judging subject.
Secondary pragmatic processes are inferential in the narrow sense
because they satisfy that essential condition — the availability condition.
But it would be wrong to claim that narrow inferences are necessarily
effortful, slow and under voluntary control. Among inferences in the
narrow sense, some — including those that underly the retrieval of
conversational implicatures — are typically spontaneous: the inference is
drawn more or less automatically.
Consider Sperber's own example of spontanous inference: you hear
the doorbell ringing, and you form the belief 'There is someone at the
door' (Sperber, 1997: 77-78). That belief is inferentially derived from a
prior belief (to the effect that the doorbell is ringing) directly based on
perception, but it does not issue from a process of explicit reasoning:

If challenged, you might be able to produce, ex post facto, a missing

premise that, together with the perceptual belief, warrants the
inferential belief. However, the fact is that you arrived at the
inferential belief without engaging in deliberate or conscious
inference (Sperber 1997: 78).

I grant that there is a difference between spontaneous inference and

explicit reasoning, but I deny that this corresponds to the distinction
between inferences in the broad and the narrow sense. On my view,
spontaneous inferences of the sort mentioned by Sperber are 'inferences
in the narrow sense'.
In all the examples given by Sperber in the passage from which I've
just quoted, including the doorbell example, the availability condition is
satisfied: the subject makes two judgments, one based on perception, the
other based on inference from the first belief. The two judgments are

58
conscious and available to the subject. Moreover, the subject is aware that
the second judgment is grounded in, and justified by, the first. If she says,
'There is someone at the door', and is asked 'How do you know?', she will
reply something like: 'The doorbell is ringing'. The perceptual judgment to
the effect that the doorbell is ringing justifies, and is known to justify, the
nonperceptual judgment that there is someone at the door. Since there are
two judgments standing in the appropriate relation to each other, Reid's
definition of 'reasoning' applies, even though this piece of reasoning is
spontaneous, effortless and fast. The only thing that is not conscious here
are the inferential steps needed to bridge the gap between the first,
perceptual belief and the second, inferential belief. Still, the inference,
though not explicit, is conscious in the sense that the availability condition
is satisfied.26 This is in contrast to cases in which the availability condition

26 The distinction between the two notions is crucial. Overlooking it leads

to fallacies, as in this passage from Michael Thau’s Consciousness and

Cognition (Oxford University Press, 2002), p. 149 :

Once we abandon the idea that conversational participants

consciously calculate what’s implicated from what is said, there’s no

reason to think that they should always have the ability to

consciously distinguish the former from the latter.

This is fallacious because the main reason for denying that conversational

participants consciously calculate what’s implicated from what is said is

that they don’t do so explicitly. (As Thau rightly points out, p. 145, « when I

understand that your utterance [‘Fergie was sober today’] is intended to

convey that Fergie is often drunk, I don’t have any explicit thoughts about

conversational maxims, nor do I perform any explicit inferential

reasoning. »). But this has no bearing at all on the issue, whether or not

the availability condition is satisfied. In this particular case I take it as

obvious that the availability condition is satisfied : the interpreter is aware

that the speaker has said that Fergie is sober today (thereby implying that

59
is not satisfied: cases, for example, where the subject is aware only of one
judgment, the alleged inferential source of that judgment being
unavailable to consciousness; or cases in which both judgments are
available, but the subject is unaware of one being inferentially derived
from the other.
I conclude that there are two sorts of inference in the narrow sense :
explicit reasoning, and spontaneous inference. In both cases the
availability condition is satisfied. Inferences in the broad sense are
characterized by the fact that the availability condition is not satisfied.
Now it is my contention that, if primary pragmatic processes are indeed
inferential, they are inferential only in the broad sense. The interpreter is
not aware that his judgment, to the effect that the speaker has said that p,
is inferentially derived from a prior judgment (e.g. a judgment to the effect
that the speaker has uttered sentence S). Similarly, some theorists claim
that perceptual judgments themselves are inferential, but the inferences
at issue do not satisfy the availability condition: the perceiver is not aware
that his perceptual judgment that p is based on a prior judgment (to the
effect that things seem thus and so, or whatever). In contrast, when I
judge that John is still here because I can see his car, that judgement is
inferential in the narrow sense. Similarly, secondary pragmatic processes
are inferential in the narrow sense, because the availability condition is
satisfied: the subject is aware of the relation between the implicature and
what is said. That is true even though (normally) conversational
implicatures are not generated through a process of explicit reasoning. As
relevance theorists rightly emphasize, implicature generation is often as
automatic, as effortless as primary pragmatic processes. But this is
irrelevant to the distinction I want to make. Despite their spontanous and
automatic character, secondary pragmatic processes are conscious in the
minimal sense that the availability condition is satisfied. In primary
pragmatic processes, the availability condition is not satisfied.

3.3. Implicature or enrichment?

he is often drunk).

60
Relevance theorists have provided a couple of arguments supporting their
view that implicatures and explicatures are not processed by separate
systems, a 'primary' pragmatic system that works blindly and in parallel,
and a 'secondary' system that takes as input the output of the primary
system. Those arguments attempt to show that the computation of what is
said does not have to precede the derivation of implicatures, contrary to
what my account suggests.
Following Herb Clark, Carston points out that "there is a class of
widely recognized implicatures, known as 'bridging' implicatures, which
have to precede the full derivation of an explicature » (Carston 2002a :
143). One of Clark's original examples was (1), which is said to implicate
(2).

(1) The picnic was awful. The beer was warm.

(2) The beer was part of the picnic.

Carston writes: "in the absence of any argument that denies the status of
implicature to assumptions like those in (2), they seem to present strong
evidence in favour of a system of pragmatic interpretation which derives
explicatures and implicatures in parallel" (Carston 2002a: 143).
Now I am puzzled by this particular example because it is so
strikingly similar to other examples of 'bridging' which Carston herself
(rightly) treats as instances of enrichment. If 'She took out her key and
opened the door <with the key>' is an instance of bridging enrichment,
the same thing should be said of the Clark example, since the only
significant difference between the two cases is the fact that there is
conjunction in one case, parataxis in the other. Conversely, if (2) is treated
as a conversational implicature of (1), then 'the door was opened with the
key' should be treated as a conversational implicature of 'She took out her
key and opened the door', contrary to Carston's own analysis in terms of
enrichment. As far as I am concerned, I opt for the enrichment account,
and I therefore deny that (2) is a conversational implicature of (1). (The

61
enrichment of incomplete descriptions is a standard phenomenon, which I
have already mentioned. The enrichment of 'the beer' into something like
'the beer that was part of the picnic' is like the enrichment of 'the table'
into 'the table of the living room'.)27
Another example of implicature invoked by Carston in arguing
against my view turns out to be best treated as a case of enrichment. She
writes:

Let's consider some examples of utterances which clearly

communicate an implicature, starting with the first example Grice
gave when illustrating the role of his conversational maxims:
(3) A: I am out of petrol.
B: There is a garage around the corner.
(Gloss: B would be infringing the maxim 'Be
relevant' unless he thinks, or thinks it possible, that
the garage is open, and has petrol to sell; so he
implicates that the garage is, or at least may be,
open, etc.)
Now, B could have given a more explicit response to A, one in which
the information that petrol is currently being sold at a garage round
the corner is part of what is said by the utterance. For instance she
could have uttered the sentence in (4):
(4) B': There is a garage round the corner
which sells petrol and is open now.
According to [Recanati's view], this utterance would have maintained

27 Carston would probably argue that examples like (1) involve both

enrichment and implicature generation. In (1), for example, the speaker

implicates that there was beer at the picnic, and the explicature of (1) is

enriched on the basis of that 'implicated premiss': 'the beer' is understood

as 'the beer that was part of the picnic'. As far as I am concerned,

however, I am unwilling to treat what relevance theorists call 'implicated

premisses' as genuine implicatures (see §3.5).

62
 the normal language flow while the one B actually gave, in (3),
disrupts that normal language flow. It is perhaps difficult to have a
sure sense of what is meant by the notion of 'normal' language flow,
but the exchange in (3) seems to be about as natural, normal and
flowing a conversation as there is, while, arguably, the implicature-
less one in (4) is somewhat awkward, being quite unnecessarily
explicit (in the absence of any doubt about the functioning of the
garage. (Carston 2002a: 144)

I have already said in what sense I mean that implicatures are special:
their distinctness from what is said is grasped, while the aspects of
meaning that are contextually provided through primary pragmatic
processes are undistinguished. Be that as it may, I deny that a
conversational implicature is involved in the Gricean example mentioned
by Carston. The content of B's utterance in (3) is enriched into the
proposition which (4) explicitly articulates. This is very standard. In the
relevance-theoretic framework favoured by Carston herself the word
'garage' can perhaps be taken to contextually express the ad hoc concept
GARAGE*, paraphrasable as 'a garage that is presumably open and has
petrol to sell'.
The inappropriateness of the alleged counterexamples should not
blind us to the important issue raised by relevance theorists: if there are
cases in which the derivation of an implicature must come before the
computation of what is said, does this not argue against my account? To
this issue I now turn.

3.4. Mutual adjustment of explicature and implicature

According to relevance theorists, I am committed to a sequential view of

the derivation of implicatures and am therefore unable to account for the
phenomenon of 'mutual adjustment' between implicature and explicature.
My view, it is said,

63
 assumes that explicature derivation precedes implicature derivation,
and, thereby, commits [me] to a sequential processing model. By
contrast, a relevance-based comprehension model is not a
sequential model: the implicature derivation may coincide with,
precede, or follow the implicature derivation, and the content of
derived implicatures may affect the way explicatures are developed.
(Matsui 2001: 250)

Relevance theorists have documented the phenomenon of 'mutual

adjustment' between implicature and explicature. As Sperber and Wilson
say, "expectations of relevance warrant the derivation of specific
implicatures, for which the explicit content must be adequately enriched"
(Sperber and Wilson 1998: 194). One of their examples is the following
exchange:

A: Do you want to go to the cinema?

B: I am tired.

A primary pragmatic process of enrichment maps the lexical concept

TIRED to the ad hoc concept TIRED*, which can be paraphrased as
something like 'tired to an extent that makes going to the cinema
undesirable to B'. Enrichment here is driven by the need to derive from B's
utterance an implicated response to A's question. Similarly, at the end of
Chapter 1, I mentioned the connection between enrichment and
implicature derivation:

By saying that she's had breafkast, the speaker implies that she is
not hungry and does not want to be fed. By saying that the child is
not going to die, the mother implies that the cut is not serious; and
so forth. Now those implicatures can be worked out only if the
speaker is recognized as expressing the (nonminimal) proposition
that she's had breakfast that morning, or that the child won't die
from that cut. Clearly, if the speaker had had breafkast twenty years

64
 ago (rather than that very morning), nothing would follow
concerning the speaker's present state of hunger and her willingness
or unwillingness to eat something. The implicature could not be
derived, if what the speaker says was not given the richer,
temporally restricted interpretation.

So I agree with Sperber and Wilson about the phenomenon they call
'mutual adjustment'. Carston concludes that " there is no generalization to
be made about which of the two kinds of communicated assumption
[explicature and implicature] is recovered first and functions as input to
the recovery of the other; the parallel adjustment process entails that
neither is wholly temporally prior to the other." (Carston 2002a: 143).
Rather, the proper generalization to make is this:

Any interpretation... results from mutual adjustment of the explicit

and implicit content of the utterance. This adjustment process
stabilises when the hypothesised explicit content is warranted by the
hypothesised explicit content together with the context. (Sperber
and Wilson 1998: 197)

But there is nothing with which I find that I have to disagree here. So I
deny that my view commits me to a 'sequential model' in the sense under
discussion. What matters from my point of view is only this: there is an
asymmetry between implicature and explicature, in that the implicature
must be grounded in, or warranted by, the explicature. Both Carson and
Sperber and Wilson grant this point, at least as far as 'implicated
conclusions' are concerned. So there is no disagreement on this score.

3.5. Implicated premisses

Deirdre Wilson once objected to me that implicatures may be premisses as

well as conclusions, hence they need not be 'grounded in' or 'warranted
by' the explicature. Implicated premisses are background assumptions

65
that either serve as auxiliary premisses in deriving the implicatures from
the explicature, or, possibly, are appealed to in enriching the content of
the utterance so as to yield an appropriate explicature. Wilson gave as
example the following exchange:

A: Why didn't you invite me to your party?

B: I invited only nice people.

The implicature here is the premiss that A is not nice. (B's utterance
provides a relevant answer to A's question only if this is assumed.) The
alleged 'bridging implicature' invoked by Carston as counterexample to my
view is also an implicated premiss. That there was beer among the
beverages served at the picnic is an implicated premiss by appeal to which
one can enrich 'the beer' into 'the beer served at the picnic'.
In general I don't think so-called 'implicated premisses' are genuine
implicatures in the sense in which I am using that term. They are not part
of what the speaker means, i.e. intentionally communicates, but part of
what he takes for granted (presupposes) and expects the hearer to take
for granted. Thus I draw a distinction between the common ground to
which such premisses belong, and the explicit or implicit content of the
utterance — what the speaker specifically intends to communicate.
Implicatures are, by definition, an aspect of content. Since that is not true
of implicated premisses, they do not count as genuine implicatures (in my
framework).
To be sure, there are cases in which an implicated premiss is part of
what is meant. Thus in the above exchange we may imagine that B's point
precisely is to convey to A that he (A) is not nice. Even in that case I would
not treat that as an instance of implicature, but, rather, as an instance of
'staging' (§5.5). The speaker ostensively speaks as if it was common
ground that A is not a nice person. The mechanism at work here is more
similar to that which underlies irony than to the inferential mechanism
characteristic of implicature. However, let us assume that there is a
genuine implicature here. What, exactly, is implicated? Wilson says it is

66
the premiss that A is not nice. But there is an alternative construal. We
may consider that what the speaker implicates is that he considers A as
not a nice person. Is this a premiss or a conclusion? Arguably, like other
implicatures it is the conclusion of an inference taking as premiss the
speaker's saying what he says. The speaker's saying that only nice
persons were invited, in response to the question 'Why didn't you invite
me?', implies that he considers the hearer as not a nice person. If this is an
implicature, one may argue that it is, indeed, grounded in or warranted by
the speaker's locutionary act (i.e. his saying of what is said), just as
ordinary implicatures are.

67
 Chapter 4

The Syncretic View

4.1. Four levels?

In Chapter 1 I described two approaches to what is said; the standard

approach, based on the Minimalist principle, and the availability-based
approach which entails rejecting Minimalism. The two pictures are
summarized in figure 1.2 (p. 00). Now an intermediate position is
available, according to which there are two equally legitimate notions of
what is said: a purely semantic, minimalist notion, and a pragmatic notion
('what is stated' as opposed to what is implied). If we accept this
suggestion, we end up with four levels instead of three:

literal sentence meaning

meaning what is saidmin
vs
speaker's what is saidprag
meaning what is implied

Such a compromise would seem to be acceptable to both parties.

The minimalist wants to isolate a purely semantic notion of content, that
is, a notion of the content of a sentence (with respect to a context) which
is compositionally determined and takes pragmatic elements on board
only when this is necessary. His opponent wants to capture the intuitive
notion of 'what is said' (as opposed to what is implied) and stresses that
what is said in that sense is, to a large extent, determined in a top down
manner by the context. The two notions can be integrated within a unified
framework if one accepts to replace the traditional triad by a four-level
picture (fig. 4.1).

[Figure 4.1 about here]

68
 Several authors have made proposals to that effect. In 'The
Pragmatic Fallacy', Nathan Salmon distinguishes two senses of the phrase
'what is said': what is said in the strict and philosophical sense (the
semantic content of the sentence, with respect to the context at hand) and
what is said in the loose and popular sense (the content of the speaker's
speech act). What is said in the loose and popular sense is typically richer
than the sentence's semantic content, yet it does not encompass what the
speaker merely conveys or implicates in Grice's sense. Salmon, in effect,
draws a threefold distinction between (i) what is said in the minimalist
sense (the semantic content of the sentence, in the speaker's context), (ii)
what the speaker asserts, and (iii) what he or she implies. When the
conventional meaning of the sentence-type is added, this yields something
very much like the four-level picture above.
Kent Bach defends a similar view, with one more level (Bach 1994a,
1994b). 'What is said' is so minimal, in Bach's conception, that it need not
even be propositional. It may be a 'propositional radical'. Thus if, pointing
to Tipper, I say 'He is ready' what I say is that Tipper is ready. This
becomes a full-fledged proposition only if an answer is provided to the
question: ready for what? A pragmatic process of 'completion' must
therefore take place to make the Bachian 'what is said' into a complete
proposition. To go from that proposition (resulting from completion) to
what is actually asserted a further pragmatic process of expansion often
comes into play. That process is clearly non-minimalist: it is neither
triggered by a linguistic constituent, nor necessary in order to achieve
propositionality. ('Expansion' is Bach's word for free enrichment.) So in
Bach's framework there are five distinct notions: the sentence's linguistic
meaning; what is said in the less-than-minimal sense; the minimal
proposition resulting from completion (= what is said in the minimalist
sense); the non-minimal proposition resulting from expansion (= what is
said in the pragmatic sense, i.e. the 'explicature', in relevance-theoretic
terminology); and, finally, the conversational implicatures of the utterance
(not to mention what the speaker nonliteraly communicates). The

69
contextual components of meaning generated through completion and
expansion are called by Bach 'conversational implicitures' because they
are 'implicit in' what is said, in contrast to the implicatures which are
'implied by' what is said (or the saying of it).
A third author worth mentioning in this connection is Scott Soames.
In Beyond Rigidity he writes:

When speaking of the information carried by an assertive utterance

of a sentence in a context, one must distinguish (i) the semantic
content of the sentence uttered in the context; (ii) what the speaker
says (asserts) by uttering the sentence; (iii) what the speaker
implies, implicates, or suggests... (i) is standardly included in (ii),
but... in the case of many utterances, (ii) is not exhausted by (i).
(Soames 2002: 86)

Like Salmon and Bach, Soames contrasts what is said in the semantic
sense with what is said in the pragmatic sense (i.e. what is stated or
asserted). It is what is said in the pragmatic sense which arguably serves
as input to secondary pragmatic processes and satisfies the availability
condition. As we have seen, that entails that what is said (in the pragmatic
sense) cannot satisfy Minimalism. But nothing follows regarding what is
said in the semantic sense; hence there is no reason to give up Minimalism
when dealing with semantic content.
Whatever we think of the details of the views held by Bach, Salmon
or Soames, it is clear that, by freeing us of the limitations of the traditional
three-level picture, they make a reconciliation of Minimalism and
Availability possible. One can be a minimalist with respect to what is said
in the strict and literal sense, and at the same time give up Minimalism
when it comes to what the speaker asserts, which is subject to the
Availability constraint.
The syncretic view promises to give us the best of both worlds. It
sounds appealing and reasonable. As Jonathan Berg pointed out, I myself
have proposed something along similar lines (Berg 1998: 466-7). Yet I

70
think there is a serious problem with that view — a problem which makes
it rather unattractive, despite its initial appeal.
The weak point in the syncretic view is the very notion of 'what is
said in the strict and literal sense', i.e. the minimalist notion of what is
said. How are we to understand that notion? A natural and widespread
interpretation runs as follows: 'What is said' in the minimalist sense is
what the sentence says (with respect to the context at hand), as opposed
to what the speaker says by uttering the sentence. Thus interpreted,
however, the syncretic view is closely related to a traditional way of
construing the semantics/pragmatics distinction which I think must be
rejected. In the next two sections of this chapter, I will spell out that
conception and show what is wrong with it.

4.2. Semantics and pragmatics: the standard picture

On the standard picture, knowing a language is like knowing a theory by

means of which one can deductively establish the truth-conditions of (or
the proposition expressed by) sentences which one has never encountered
before. Semantic interpretation is the process whereby an interpreter
exploits his or her knowledge of a language, say L, to assign to an
arbitrary sentence of L its truth-conditions.
Pragmatic interpretation is a totally different process. It is not
concerned with language per se, but with human action. When someone
acts, there is a reason why he does what he does. To provide an
interpretation for the action is to find that reason, that is, to ascribe the
agent a particular intention in terms of which we can make sense of the
action.
A distinguishing characteristic of pragmatic interpretation is its
defeasability. The best explanation we can offer for an action given the
available evidence may be revised in the light of new evidence. Even if an
excellent explanation is available, it can always be overriden if enough
new evidence is adduced to account for the subject's behaviour. It follows
that any piece of evidence may turn out to be relevant for the

71
interpretation of an action. In other words, there is no limit to the amount
of contextual information that can affect pragmatic interpretation.
A particular class of human actions is that of communicative actions.
That class is defined by the fact that the intention underlying the action is
a communicative intention — an intention such that (arguably) its
recognition by the addressee is a necessary and sufficient condition for its
fulfilment (Searle 1969, Bach and Harnish 1979). To communicate that p is
therefore to act in such a way that the addressee will explain one's action
by ascribing to the agent the intention to communicate that p. For
communication to succeed, the addressee must not only understand that
the agent does what he does in order to communicate something to her;
she must also understand what the agent tries to communicate. To secure
that effect the communicator will do something which will evoke in the
adressee's mind that which he wants to communicate. To that purpose the
communicator may use icons, or indices, or symbols, that is, conventional
signs. It is, of course, symbols that are used when the communicator and
the addressee share a common language.
At this point semantic interpretation and pragmatic interpretation
make contact with each other. A speech act is an action performed by
uttering a sentence in some language, say L. Let us assume that the
sentence has a certain semantic interpretation in L: it means that p. Since
the speaker utters a sentence which means that p and manifests the
intention to communicate something to the hearer, one reasonable
hypothesis is that he intends to communicate that p. If that is the best
explanation for the action given the available evidence, the hearer will
settle for it and (if that was indeed the speaker's intention) the
communicative intention will be fulfilled: the speaker will have succeedeed
in communicating that p to the hearer. In this case the speech act will be
assigned a particular content as a result of pragmatic interpretation; and
that content will coincide with the content which comes to be assigned to
the sentence as a result of semantic interpretation. That is not really a
coincidence, of course; for the semantic interpretation of the sentence was
part of the evidence used in pragmatically determining the content of the

72
speech act. But there are cases in which the two contents do not coincide:
the sentence means that p, but that is not what the speaker means —
what he manifestly intends to communicate.
So far I have expounded the standard picture of the relation between
semantic and pragmatic interpretation. I will criticize it shortly (§4.3).
Before doing so, let me spell out the connection between that picture and
the Syncretic View talked about in the first section.
On the standard picture, as we have just seen, there is a basic
distinction between what the sentence says and what the speaker means,
even when they coincide. What the sentence says is determined by
semantic interpretation, that is, deductively and without paying any regard
to the speaker's beliefs and desires. Of course, one needs to make sure
that the speaker utters what he does as a sentence of L; and that may
require a good deal of pragmatic interpretation. But once it is determined
that the utterance at issue counts as an utterance of a particular sentence
of L, semantic interpretation takes over, and the content of that sentence
is mechanically determined.— On the other hand what the speaker means
is determined by pragmatic interpretation. Pragmatic interpretation
invoves an assessment of the speaker's communicative intentions, given
an overall assumption of rationality. As I pointed out any piece of
contextual information may turn out to be relevant to establishing the
correct interpretation for the speech act.
In this framework there is room for the notion of what the speaker
says — what is said in the pragmatic sense, or, as Bach puts it, 'what is
stated'. Sometimes we can distinguish two components within what the
speaker means: what he states, and what he implies in virtue of stating it.
What is stated may, but need not, be identical to what the sentence says.
Often what is stated is richer than what the sentence says, as we have
seen. Be that as it may, there are two notions of what is said: one is the
output of semantic interpretation. It is what the sentence says. The other
is a particular aspect of speaker's meaning. It is what the speaker states.
That distinction is the core of the Syncretic View.

73
 4.3. Semantic underdeterminacy

I think there is something deeply wrong with the standard picture (and the
Syncretic View insofar as it is based on it). What is wrong — or one of the
things that are wrong — is the assumption that semantic interpretation
can deliver something as determinate as a complete proposition. On my
view semantic interpretation, characterized by its deductive character,
does not deliver complete propositions: it delivers only semantic schemata
— propositional functions, to use Russell's phrase.
By saying that semantic interpretation properly conceived delivers
only schemata, not complete propositions, I do not mean that it delivers
only characters in Kaplan's sense (Kaplan 1989). When a sentence
contains an indexical, like 'I' or 'tomorrow', the meaning of the indexical
(its character) contextually determines its content in a very
straightforward manner. There is no reason not to consider that aspect of
content-determination as part of semantic interpretation. For the type of
context-dependence exhibited by (pure) indexicals has nothing to do with
the radical form of context-dependence which affects speaker's meaning.
The hallmark of the more radical form of context-dependence is the fact
that any piece of contextual information may be relevant. But the context
that comes into play in the semantic interpretation of indexicals is not the
total pragmatic context; it is a very limited context which contains only a
few aspects of the pragmatic context: who speaks, when, where, and so
forth. As Bach puts it,

There are two quite different sorts of context, and each plays quite a
different role. Wide context concerns any contextual information
relevant to determining the speaker's intention and to the successful
and felicitous performance of the speech act... Narrow context
concerns information specifically relevant to determining the
semantic values of [indexicals]... Narrow context is semantic, wide

74
 context pragmatic. 28

But most context-sensitive expressions are semantically

indeterminate rather than indexical in the strict sense. A possessive
phrase such as 'John's car' means something like the car that bears
relation R to John, where 'R' is a free variable. The free variable must be
contextually assigned a particular value; but that value is not determined
by a rule and it is not a function of a particular aspect of the narrow
context. What a given occurrence of the phrase 'John's car' means
ultimately depends upon what the speaker who utters it means. It
therefore depends upon the wide context. That is true of all semantically
underdeterminate expressions: their semantic value varies from
occurrence to occurrence, just as the semantic value of indexicals does,
yet it varies not as a function of some objective feature of the narrow
context but as a function of what the speaker means. It follows that
semantic interpretation by itself cannot determine what is said by a
sentence containing such an expression: for the semantic value of the
expression — its own contribution to what is said — is a matter of
speaker's meaning, and can only be determined by pragmatic
interpretation.
Even if we restrict our attention to expressions traditionally
classified as indexicals, we see that they involve a good deal of semantic
indeterminacy. This is true, in particular, of demonstratives. The reference
of a demonstrative cannot be determined by a rule, like the rule that 'I'
refers to the speaker. It is generally assumed that there is such a rule,
namely the rule that the demonstrative refers to the object which happens
to be demonstrated or which happens to be the most salient, in the
context at hand. But the notions of 'demonstration' and 'salience' are
pragmatic notions in disguise. They cannot be cashed out in terms merely
of the narrow context. Ultimately, a demonstrative refers to what the
speaker who uses it refers to by using it.

28 From the handout of a talk on 'Semantics vs Pragmatics', delivered in

1996.

75
 To be sure, one can make that into a semantic rule. One can say that
the character of a demonstrative is the rule that it refers to what the
speaker intends to refer to. As a result, one will add to the narrow context

a sequence of 'speaker's intended referents', in such a way that the nth

demonstrative in the sentence will refer to the nth member of the

sequence. Formally that is fine, but philosophically it is clear that one is
cheating. We pretend that we can manage with a limited, narrow notion of
context of the sort we need for handling indexicals, while in fact we can
only determine the speaker's intended referent (hence the semantic
referent, which depends upon the speaker's intended referent) by
resorting to pragmatic interpretation and relying on the wide context. 29

We encounter the same sort of problem even with expressions like

'here' and 'now' which are traditionally considered as pure indexicals
(rather than demonstratives). Their semantic value is the time or place of
the context respectively. But what counts as the time and place of the
context? How inclusive must the time or place in question be? It depends

29 One may acknowledge the need to appeal to the wide context in

interpreting demonstratives while arguing that the appeal thus made is

'pre-semantic'. The speaker's directing intention is necessary to endow a

demonstrative with a character in the first place, one might say. We need

the wide context to fix the expression's character, but once the character

is thus determined, through a pragmatic procedure akin to

disambiguation, it maps the narrow context onto the proposition

expressed by the utterance in that context. Similarly, we might say that

whenever there is semantic underdeterminacy, some form of pragmatic

disambiguation must take place before the process of semantic

interpretation can start. — To argue in this way, it seems to me, is to

acknowledge that semantic interpretation by itself cannot determine the

content of a sentence containing a semantically underdeterminate

expression.

76
on what the speaker means, hence, again, on the wide context. We can
maintain that the character of 'here' and 'now' is the rule that the
expression refers to 'the' time or 'the' place of the context — a rule which
automatically determines a content, given a (narrow) context in which the
time and place parameters are given specific values; but then we have to
let a pragmatic process take place to fix the values in question, that is, to
determine which narrow context, among indefinitely many candidates
compatible with the facts of the utterance, serves as argument to the
character function. On the resulting view the (narrow) context with respect
to which an utterance is interpreted is not given, it is not determined
automatically by objective facts like where and when the utterance takes
place, but it is determined by the speaker's intention and the wide context.
Again, we reach the conclusion that pragmatic interpretation has a role to
play in determining the content of the utterance, in such a case.
To sum up, either semantic interpretation delivers something gappy,
and pragmatic interpretation must fill the gaps until we reach a complete
proposition. Or we run semantic interpretation only after we have used
pragmatic interpretation to pre-determine the values of semantically
indeterminate expressions, which values we artificially feed into the
narrow context. Either way, semantic interpretation by itself is powerless
to determine what is said, when the sentence contains a semantically
indeterminate expression.
Now I take it that such expressions can be found all over the place.
Moreover, semantic indeterminacy is not limited to particular lexical items.
One can follow Waismann and argue that the satisfaction conditions of any
empirical predicate are semantically indeterminate and subject to
pragmatic interpretation (§9.4). There is also constructional indeterminacy.
For example consider something as simple as the Adjective+Noun
construction, as in 'red pen'. What counts as a red pen? A pen that is red.
But when does a pen count as red? That depends upon the wide context.
The satisfaction conditions of 'red pen' can only be determined by
pragmatic interpretation (§9.3).
Suppose I am right and most sentences, perhaps all, are

77
semantically underdeterminate. What follows? That there is no such thing
as 'what the sentence says' (in the standard sense in which that phrase is
generally used). There is a single notion of what is said, and that is a
pragmatic notion: saying, as Grice claimed, is a variety of non-natural
meaning, characterized by the role which the conventional meaning of the
sentence plays in the hearer's intended recognition of the speaker's
communicative intention.

4.4. The minimal proposition as 'common denominator'

My conclusion at this point should not be overstated. I am not (yet) saying

that the Syncretic View cannot be made sense of, only that a particular
interpretation (the standard interpretation) must be rejected. There is, I
claimed, no such thing as 'what the sentence says' in the standard sense,
i.e. no such thing as a complete proposition autonomously determined by
the rules of the language with respect to the context but independent of
speaker's meaning. As Bach points out, what the sentence says, in the
purely semantic sense, "excludes anything that is determined by [the
speaker's] communicative intention (if it included that, then what is said
would be partly a pragmatic matter)" (Bach 2001: 21). It is for that reason
that I say there is no such thing: in order to reach a complete proposition
through saturation, we must appeal to speaker's meaning. That is the
lesson of semantic underdeterminacy.
There is another possible interpretation for the semantic notion of
'what is said', however. Instead of construing what is said as a
nonpragmatic property of the sentence, independent of speaker's
meaning, we can start with the pragmatic notion of what is said, and
define the semantic notion in terms of it. What is said in the minimal sense
can thus be defined as what is said in the full-fledged, pragmatic sense
minus the contextual ingredients that are optional and whose provision is
context-driven. To filter out the optional ingredients, while retaining the
contextual ingredients that are necessary for propositionality (reference of
indexicals, etc.), one may follow Soames and define the semantic content

78
of a sentence s relative to a context C as

that which would be asserted and conveyed by an assertive

utterance of s in any normal context in which the reference of all
indexicals in s is the same as their reference in C. (Soames 2002:
106).

Soames's strategy therefore consists in abstracting what is said in the

semantic sense from what is said in the pragmatic sense. That is possible
because, according to Soames, the semantic content of the sentence is
included in the content of the assertion. To get to the semantic content,
one only has to filter out those aspects of assertion content that go
beyond semantic content and are tied to specific contexts of utterance.
What remains, i.e. the 'common denominator', is the minimal proposition
expressed by the sentence itself:

The semantic content of a sentence relative to a context is

information that a competent speaker/hearer can confidently take to
be asserted and conveyed by an utterance of the sentence in the
context, no matter what else may be asserted, conveyed, or
imparted. It is a sort of minimal common denominator determined
by the linguistic knowledge shared by all competent speakers,
together with contextually relevant facts such as the time, place,
and agent of the context; the identity of individuals demonstrated by
the speaker; and the referents of the names, as used in the context.
As such, the semantic content of a sentence functions as a sort of
minimal core around which speaker/hearers can structure the totality
of information the sentence is used to communicate in a given
context. (Soames 2002: 109)

This alternative strategy is also a failure, however. What is said in

the minimal, semantic sense cannot be abstracted from what is said in the
pragmatic sense simply because it need not be part of it. Soames's claim

79
that semantic content is included in assertion content seems plausible
because his examples are all cases in which the asserted content is richer
than the alleged semantic content. He gives the following sort of example:

A man goes into a coffee shop and sits at the counter. The waitress
asks him what he wants. He says, "I would like coffee, please." The
sentence uttered is unspecific in several respects — its semantic
content does not indicate whether the coffee is to be in form of
beans, grounds, or liquid, nor does it indicate whether the amount in
question is a drop, a cup, a gallon, a sack, or a barrel. Nevertheless,
it is obvious from the situation what the man has in mind, and the
waitress is in no doubt about what to do. She brings him a cup of
frehsly brewed coffee. If asked to describe the transaction, she might
well say, "He ordered a cup of coffee" or "He said he wanted a cup of
coffee", meaning, of course, the brewed, drinkable kind. In so doing,
she would, quite correctly, be reporting the content of the man's
order, or assertion, as going beyond the semantic content of the
sentence he uttered. (Soames 2002: 78)

Now Soames thinks that in such cases several propositions are asserted,
including (i) the unspecific proposition literally expressed (to the effect
that the man wants coffee in some form or other) and (ii) more specific
propositions recoverable from the literal proposition and the context.
Those more specific propositions resulting from enrichment are tied to the
particular context in which they are generated, hence they can be filtered
out by considering other contexts in which that sentence would be uttered
and the indexicals would be given the same semantic values. Soames
equates the minimal proposition expressed by the sentence with the
proposition which would be asserted in all such contexts.
The problem is that enrichment is only one optional process among
others. Only in the case of enrichment, however, is it plausible to suggest
that the minimal proposition itself is part of what is asserted. Soames
glosses assertion in terms of commitment: "assertively uttering a sentence

80
with the intention to assert or convey p involves doing so with the
intention of committing oneself to p" (73). Since one cannot commit
oneself to the truth of a specific proposition p without committing oneself
to the truth of a less specific proposition q which it entails, it makes sense
to say that the minimal proposition q is asserted when one (intuitively)
asserts an enriched proposition p. But the principle that the minimal
proposition is part of what is asserted (hence can be abstracted from it)
does not hold when the primary processes at issue are instances of
loosening or transfer. The speaker who assertively utters 'The ham
sandwich left without paying', thereby referring to the ham sandwich
orderer, does not assert the minimal proposition that the sandwich itself
left without paying! Hence the minimal proposition cannot be defined as
the common denominator — what is asserted in all contexts in which the
sentence is uttered and the indexicals are given the same semantic values
as in the current context.
To be sure, the counterexamples involving loosening or transfer are
taken care of, in Soame's framework, by the notion of a 'normal' context,
that is, of a context in which the sentence "is used with its literal
meaning", i.e. "is not used metaphorically, ironically, or sarcastically".
However this qualification cannot be invoked in the context of the present
debate without begging the question. We are supposed to start with the
intuitive (nonminimalist) notion of what is said, which sometimes is
affected by loosening or transfer. If, following Soames, we want to build
the notion of the minimal proposition out of what is said in that sense, we
cannot arbitrarily set aside the cases that potentially threaten the
enterprise.

4.5. Interaction between saturation and optional pragmatic

processes

The problem with the 'minimal proposition' is that it results from, and
presupposes, the process of saturation (i.e. the contextual assignment of
semantic values to indexicals and other context-sensitive expressions) — a

81
process that, in most cases, is impossible without appealing to speaker's
meaning. It follows that the minimal proposition is not independent from
speaker's meaning. On the other hand, the alleged minimal proposition
does not necessarily correspond to an aspect of what the speaker means
or asserts.
Another, related difficulty is this. To get the minimal proposition, we
must separate those aspects of speaker's meaning which are provided in
the course of saturation, i.e., which fill gaps in the representation
associated with the sentence as a result of purely semantic interpretation,
and those aspects of speaker's meaning which are optional and enrich or
otherwise modify the representation in question. The former, but not the
latter, contribute to the minimal proposition. Those aspects of speaker's
meaning cannot easily be separated, however. They are indissociable,
often mutually dependent aspects of a single process of pragmatic
interpretation. This can be seen by considering yet another example in
which the interpretation of various constituents in a sentence proceeds in
an interactive manner.
As we saw in §2.4, how we interpret a particular constituent often
depends upon how we interpret the other constituents in the sentence.
The cases described in §2.4 were cases in which one constituent had to be
interpreted non-literally if the others were interpreted literally. There are
also cases in which a non-literal interpretation for a constituent prompts a
non-literal interpretation for another constituent, and cases of interaction
involving e.g. saturation. Thus in interpreting an utterance of: 'Mike
finished John's book', we have to assign a contextual value to 'finish'
(finish reading, finish writing, finish binding, etc.) and a contextual value to
'John's book' (book owned by John, written by John, read by John, etc.).
There is an obvious interaction between the two values respectively
assigned to the two constituents; the overall assignment must be
coherent. If 'John's book' is interpreted as meaning the book written by
John, it's unlikely that 'finished' will be interpreted as meaning finished
writing.
The cases of interaction that are relevant to the present discussion

82
are cases in which the primary pragmatic processes which interact
respectively belong to the two broad classes I have distinguished: the
primary pragmatic processes which are semantically mandatory because
no proposition would be expressed without them (saturation), and the
processes which are optional, i.e. non-mandatory (enrichment, loosening,
and transfer). We naturally interpret 'He took out his key and opened the
door' as meaning that he opened the door with the key, but even if we did
not 'enrich' the interpretation of the utterance in this way the latter would
still express a complete proposition (namely the proposition that he did
two things: take out his key, and open the door in some way or other). In
this regard enrichment is optional and differs from saturation; for 'I read
John's book' does not express a complete proposition if a certain relation
between John and the book is not contextually provided.
Now it is possible for two pragmatic processes respectively
belonging to the two classes distinguished above to interact in such a way
that the mandatory process depends upon the optional one, and cannot
operate until the latter has. For example, take the construction: the x of y,
or even better the possessive construction: y's x. For the sentence where
such a construction occurs to express a definite proposition, a certain
relation between x and y must be contextually provided. Which relation is
actually provided as part of the interpretation of the utterance obviously
depends upon the identity of the relata and what we know about them.
The relata, x and y, will themselves be referred to by means of linguistic
expressions, say NP1 and NP2. Now it's easy to imagine a case in which,
say, NP2 can undergo a process of non-literal interpretation. For example a
definite description such as 'the lion' can be either interpreted literally (as
referring to a lion) or non-literally (as referring metonymically to, say, the
warrior whose shield carries a picture of a lion). How we interpret the
description determines the identity of one of the relata, and we have seen
that (in some cases at least), how we interpret the relation expressed by
the possessive construction depends upon the identity of the relata. I want
to argue that, in such a case, if the description is used non-literally (if it
refers to the warrior), then a pragmatic process which must take place for

83
the utterance to express a complete proposition (namely the assignment
to the possessive construction of a particular relation as contextual value)
cannot take place unless a non-literal interpretation has been antecedently
provided for some other constituent in the sentence. It follows that the
utterance is semantically indeterminate at the purely literal level, for the
mandatory process of saturation - without which no complete proposition
would be expressed - itself depends on some feature of the non-literal
interpretation of the utterance and therefore cannot be achieved at the
literal level. This is a case in which it is not possible (even in principle) to
determine the proposition literally expressed before one has determined
what the speaker non-literally means.
Let us now actually construct the example by filling in the context.
Suppose that there has been a ritual fight between respectively five
warriors and five beasts. The beasts are a wolf, a lion, a hyena, a bear, and
an alligator; the warriors are armed with swords, and carry shields with
distinctive decorations (the first warrior has the moon on his shield, the
second one has the Eiffel Tower, the third one has the Metro-Godwin-Mayer
roaring lion, etc.). Each warrior is assigned a particular beast which he or
she must stab to death. After the ritual fight, the five beasts lay on the
ground with a sword through their body. This is the context. Now suppose
that in this context I utter

Bring me the lion's sword - I want to have a look at it.

In this context I think there are two accessible interpretations for 'the lion';
the lion can be interpreted literally as referring to the lion (one of the
beasts), or non-literally as referring to the warrior who has (a picture of) a
lion on his shield. If we choose the first interpretation, the relation which
will be contextually assigned to the possessive construction will be one of
the salient relations which hold between the lion (the animal) and the
sword which can be seen emerging from its pierced body. If we choose the
second interpretation, the relation will be totally different; it will be one of
the salient relations between the warrior and the sword which he used in

84
his fight against, say, the bear.
Now suppose the correct interpretation is the second one: the
speaker actually refers to the warrior with a lion on his shield and wants to
see his sword (i.e. the sword which emerges from the bear's body). What
is the minimal proposition expressed by the utterance, in this context? I
doubt that there is one, but if the minimalist insists, here is the only
available procedure for determining it: To get the minimal proposition, we
must give the word 'lion' its literal interpretation, because the nonliteral
interpretation results from an optional, non-minimalist process; and we
must assign a particular value to the variable ‘R’ carried by the possessive
construction. Which value? Well, the value which corresponds to what the
speaker actually means. (Remember: There is no other way to contextually
determine a value for the genitive, than by appeal to speaker's meaning.)
But that value is that which goes together with the intended nonliteral
interpretation of 'lion'! The result is a monster: what the phrase 'the lion's
sword' contributes to what is said in the minimalist sense is something like
the sword which the lion (the animal) used during the fight. The minimal
proposition thus determined is absurd and evidently corresponds to no
stage in the actual process of understanding the utterance.

4.6. Taking stock

We have seen that the minimal proposition which the Syncretic View
posits as the semantic content of the utterance, and which results from
saturation alone, is not 'what the sentence says' in the objectionable sense
glossed in §4.2 and criticized in §4.3. It is not autonomously determined by
the rules of the language independent of speaker's meaning. At the same
time, the minimal proposition does not necessarily correspond to an
aspect of what the speaker asserts (§4.4). The minimal proposition is a
hybrid which goes beyond what is determined by the rules of the language
yet has no psychological reality and need not be entertained or
represented at any point in the process of understanding the utterance
(§4.5).

85
 Do we need such a notion in theorizing about language and
communication? Many people claim that we do, but I can hardly
understand why. In a forthcoming paper ('Semantics, Pragmatics, and the
Role of Semantic Content'), King and Stanley argue against semantic
theories which ascribe (functional) characters to sentences on the basis of
the characters of their parts, on the grounds that "the job of character is to
give us content, and we can assign contents to complex expressions in
contexts using only the characters of the parts, and combining the
contents they determine in those contexts." They conclude that

Both a semantics that assigns characters to simple expressions and

recursively assigns characters to complex expressions and a
semantics that assigns characters to only simple expressions allow
for an assignment of the same contents in contexts to simple and
complex expressions. So unless the functional characters of complex
expressions have some additional job to do, they are unnecessary.
But there seems to be no such additional job.

The same sort of argument seems to me to rule out the minimal

proposition as unnecessary. What must ultimately be accounted for is what
speakers say in the pragmatic sense. The job of characters, contents etc.
is to contribute to the overall explanation. But it is sufficient to assign
semantic contents (in context) to simple expressions. Pragmatic processes
will operate on those contents, and the composition rules will compose the
resulting pragmatic values, thereby yielding the content of the speaker's
assertion. Of course it is possible to let the composition rules compose the
semantic contents of the constituent expressions, thereby yielding the
minimal proposition expressed by the sentence. (An absurd proposition, in
the lion's sword case.) However, the content of the speaker's assertion will
still be determined by composing the pragmatic values resulting from the
operation of pragmatic processes on the contents of the constituent
expressions; so it is unclear what additional job the minimal proposition is
supposed to be doing.

86
 4.7. Minimal or reflexive?

If one insists in using a purely semantic notion of 'what is said', i.e. a

notion of what is said which is propositional (truth-evaluable) yet "excludes
anything that is determined by the speaker's communicative intention",
there is a much better candidate than the alleged minimal proposition. For
every utterance, there arguably is a proposition which it expresses in
virtue solely of the rules of the language, independent of speaker's
meaning: that is the 'reflexive' proposition in the sense of John Perry (a
variant of Stalnaker's diagonal proposition). 30 The main difference between
the minimal proposition and the reflexive proposition is that the reflexive
proposition is determined before the process of saturation takes place. The
reflexive proposition can't be determined unless the sentence is tokened,
but no substantial knowledge of the context of utterance is required to
determine it. Thus an utterance u of the sentence 'I am French' expresses
the reflexive proposition that the utterer of u is French. That it does not
presuppose saturation is precisely what makes the reflexive proposition
useful, since in most cases saturation proceeds by appeal to speaker's
meaning. If one wants a proposition that's determined on purely semantic
grounds, one had better not have it depend upon the process of
saturation.
Soames considers the possibility of equating 'what the sentence
says' with the reflexive proposition or something close to it, but he rejects
that option with the following argument:

Consider... the first-person singular pronoun as it occurs in a

sentence I am F. There is no such thing as "what this sentence says"
independent of the context of utterance in which it is used. The
competence conditions associated with the first-person singular
pronoun guarantee that when I assertively utter the sentence, I use

30 See R. Stalnaker, Context and Content, Oxford University Press 1999; J.

Perry, Reference and Reflexivity, CSLI Publications 2001.

87
 it to say something about me, whereas when you assertively utter it,
you use it to say something about you. One might be tempted to
suppose that there is some more general thing that the sentence
"says" in every context — namely, the proposition expressed by the
speaker is F (or some such thing). But this will not do. Our notion of
"what a sentence says" is tied to what speakers who assertively
utter the sentence say. Typically, when I assertively utter I am F, I
don't assert that I am speaking or using language at all. Further, the
proposition that I assert when I assertively utter such a sentence
may be true in a possible circumstance in which no one is using
langage, and someone may believe this proposition without
believing anything about me being a speaker. (Soames 2002: 104)

Soames' main objection is that the alleged reflexive proposition is not (part
of) what the speaker asserts. As we have seen, however, the same thing
often holds for the minimal proposition posited by the syncretists. The
advantage of the reflexive proposition over the minimal proposition is that
it (the reflexive proposition) is determined solely by the rules of the
language, independent of speaker's meaning, in such a way that there is a
path to the reflexive proposition that does not go through the speaker's
meaning; hence it does not matter much if that proposition can't be
reached by abstraction from what the speaker asserts.
The reflexive proposition is admittedly distinct from that which the
speaker asserts – they have different possible worlds truth-conditions, as
Soames points out — but why is this an objection? Are we not supposed to
draw a distinction between the proposition expressed by the sentence and
the proposition asserted by the speaker who utters that sentence? Note
that we can, if we wish, incorporate into the reflexive proposition
something to the effect that the linguistic mode of presentation associated
with the first person pronoun will not be part of the proposition asserted,
while the reference it contextually determines will be. Thus we might take
the reflexive proposition expressed by an utterance u of 'I am French' to be
the proposition that there is an x such that x utters u and u is true iff x is

88
French. This comes as close as one can get to capturing, in propositional
format, the information provided by the utterance in virtue solely of the
linguistic meaning of the sentence 'I am French'. (See Recanati 1993 for an
analysis along those lines.) Such a reflexive proposition determines that
the proposition contextually asserted by 'I am French' will consist of the
reference of 'I' and the property of being French. The reflexive proposition
is therefore 'tied to what speakers who assertively utter the sentence say',
even if it is not (part of) what they say.
I conclude that there may be a way of preserving the notion of 'what
the sentence says', in the purely semantic sense, if one wants to; but it
does not support the Syncretic View with its four levels. What
characterizes the reflexive proposition is that, although fully propositional,
it does not incorporate those contextual ingredients whose provision is
linguistically mandated; it is much closer to the linguistic meaning of the
sentence — indeed it is directly and immediately determined by the
linguistic meaning of the sentence. Appealing to the reflexive proposition
instead of the minimal proposition takes us back to the nonminimalist
picture with its three basic levels: the linguistic meaning of the sentence
(and the reflexive proposition it directly and immediately determines);
what is said in the pragmatic sense; and what is implied or otherwise
conveyed by the utterance.

89
 Chapter 5

Nonliteral uses

It may be thought that the pragmatic approach to what is said advocated

in the first four chapters of this book blurs the commonsensical distinction
between what is (literally) said and what is nonliterally conveyed. What is
said, in the pragmatic sense, already incorporates the derived, nonliteral
values resulting from primary pragmatic processes of the optional variety.
Is there still room for a contrast between literal and nonliteral speech?
Sperber and Wilson think this distinction belongs to folklinguistics and
must be repudiated (Wilson and Sperber 2002: 622-4). In this chapter,
however, I will attempt to show that the ordinary, folk-theoretical notion of
nonliteral use can be rescued. It corresponds to a range of phenomena
which constitute a natural class, from a phenomenological point of view.

5.1. Nonliteral uses as nonminimal departures from literal

meaning

Let us start with a sense of the phrase 'literal meaning' which is

reasonably clear and raises no particular problem. In that sense, the literal
meaning of a linguistic expression is its conventional meaning: the
meaning it has in virtue of the conventions which are constitutive of the
language. Thus understood literal meaning is a property of the expression-
type; for it is the expression-type which the conventions of the language
endow with a particular meaning. Literal meaning, in that sense, I will dub
't-literal meaning' (with 't' standing for 'type'), in order to distinguish the
sense just introduced from other possible senses of the phrase 'literal
meaning'.
Particular occurrences of an expression also possess meaning. First,
every occurrence inherits the meaning of the expression-type of which it is
an occurrence. Every occurrence of a meaningful expression-type

90
therefore possesses a t-literal meaning. Second, an occurrence also
possesses a meaning which depends not merely upon the conventional
significance of the expression-type, but also on features of the context of
use. That meaning is, by definition, not t-literal. Insofar as it departs from
the meaning of the expression-type, it may even be said to be 't-
nonliteral'. Yet it need not be 'nonliteral' in the ordinary sense. When we
speak of 'nonliteralness', in the ordinary sense, we mean that what is
meant departs from t-literal meaning in a fairly specific way. Not any old
departure from t-literal meaning adds up to nonliteralness in the ordinary
sense.
Consider the following example. Suppose that Paul is thirsty, and I
tell you, while pointing to him: "He is thirsty". I thereby say of Paul that he
is thirsty. What is thus said is not t-literal because the reference to Paul is
something that is achieved by partly contextual means. The demonstrative
pronoun 'he', qua expression-type, does not refer to Paul. The semantic
conventions of the language do not assign that expression a reference, but
merely a rule of use in virtue of which it may, in context, acquire a
reference. Since the reference of the expression is not fixed by the
semantic conventions of the language, independent of context, it is not
part of the t-literal meaning of the sentence. The proposition that Paul is
thirsty (at the time of utterance) therefore counts as t-nonliteral, but of
course no one wants to say that there is anything 'nonliteral' in the
ordinary sense going on in that example (as described so far). The speaker
is speaking literally, in the ordinary sense of the word. It may be that he is
communicating something nonliterally by his utterance, but that can only
be something different from the proposition that Paul is thirsty. If the
speaker means that Paul should be offered a drink, that aspect of the
meaning of his utterance will indeed count as nonliteral in the ordinary
sense. The proposition that Paul should be offered a drink is conveyed
without being literally expressed.
What is the difference between the proposition that Paul is thirsty
and the proposition that he should be offered a drink, in the above
example? The former departs from the t-literal meaning of the sentence

91
since it includes something (the reference of 'he') which depends upon the
context and not merely upon the conventional significance of the uttered
words. Yet it is the words themselves which, in virtue of their conventional
significance, make it necessary to appeal to context in order to assign a
reference to the demonstrative. It is part of the t-literal meaning of
indexical expressions that they should be assigned a reference in context.
In interpreting indexical sentences, we go beyond what the conventions of
the language give us, but that step beyond is still governed by the
conventions of the language. The rule of use which constitutes the t-literal
meaning of indexical expressions is what triggers the search for a
contextual value. (This is the process I call 'saturation'.) The departure
from t-literal meaning is therefore pre-determined by t-literal meaning.
Whenever that is so, I say that the departure is 'minimal'. When the
meaning of an utterance only minimally departs from t-literal meaning,
that meaning does not count as nonliteral in the ordinary sense. Only
nonminimal departures make for nonliteralness in the ordinary sense. That
is precisely what we have in the other case — when the speaker says that
Paul is thirsty and implies that he should be offered a drink. For there is
nothing in the sentence-type 'He is thirsty' that triggers the contextual
generation of the implicature that Paul should be offered a drink.
Let me define: the meaning conveyed by an utterance is 'm-literal' iff
it involves only minimal departures from t-literal meaning. (The case in
which no departure is involved may count as a limiting case of m-
literalness. I doubt that there are such cases, however.) Standard cases of
nonliteralness, in the ordinary sense, are cases of m-nonliteralness: they
involve nonminimal departures from t-literal meaning. Yet, as we shall see,
not all cases of m-nonliteralness (i.e. not all cases involving such
departures) count as 'nonliteral' in the ordinary sense.

5.2. Nonliteral uses and secondary meaning

More often than not, nonliteral meaning is secondary meaning — meaning

derived from some more basic, primary meaning which it presupposes. In

92
the above example, the proposition that Paul is thirsty is primary. By
asserting that proposition, the speaker indirectly conveys something more:
that Paul should be offered a drink.
Conversational implicatures and indirect speech acts obviously have
a secondary character. Let us start with conversational implicatures. They
are a special case of pragmatic implication. Pragmatic implications are the
implications of actions. When I do something, my so doing may imply
various things. For example, my taking an unmbrella implies that I think it
will rain. Conversational implicatures are pragmatic implications of an act
of speech. They fall under the schema: the speaker's saying that p implies
that q. Not all implications of an act of speech are implicatures, however.
My saying that Frege died in 1940 implies that I am ignorant, but that is
not something I mean by my utterance. Arguably, a necessary condition
for something to count as an implicature is that it be part of what the
speaker means by his utterance. For that condition to be satisfied, the
speaker must overtly intend the hearer to recognize the pragmatic
implication of his utterance, and to recognize it as intended to be
recognized. (Other conditions may have to be satisfied for a pragmatic
implication of a speech act to count as a genuine conversational
implicature, but I will be content with what I have just said.)
Since what is conversationally implicated is implied by the speaker's
saying what she says, it immediately follows that conversational
implicatures have a secondary character. The speaker implies that q by
saying that p. In order to derive the implicature, we need the premiss that
the speaker has said what he has said; therefore we need to have
identified the primary content of his utterance.
The same thing holds, even more obviously, for indirect speech acts.
As their name indicates, indirect speech acts are performed 'indirectly', via
the performance of another speech act which falls within the illocutionary-
act potential of the uttered sentence and is said to be performed directly.
For example, I may make a request either directly ('Pass me the salt,
please') or indirectly, by asking a question ('Can you reach the salt?') or by
making a statement ('I can't reach the salt'). Twenty years ago I suggested

93
that indirect speech acts are nothing but a special case of conversational
implicature, where the speaker's intention to perform the indirect speech
act is conversationally implicated by his performance of the direct speech
act (Recanati 1979, Recanati 1987: 121-126). Be that as it may, everyone
agrees that there is much in common between conversational implicatures
and indirect speech acts. What they have in common is, in effect, their
secondary character: The interpretation of both conversational
implicatures and indirect speech acts involves an inference from the
utterance's primary meaning to its derived meaning.
In both conversational implicatures and indirect speech acts, the
meaning of the utterance involves something over and above its primary
meaning — something whose derivation presupposes the primary
meaning. In other cases, like irony, the situation is more complex. Suppose
the speaker says 'Paul really is a fine friend', in a situation in which just the
opposite is known to be the case. The speaker does not really say, or at
least she does not assert, what she 'makes as if to say' (Grice's phrase).
Something is lacking here, namely the force of a serious assertion. While
in conversational implicature, the speaker asserts something and conveys
something more as well, in irony the speaker does less than assert what
she would normally be asserting by uttering the sentence which she
actually utters. What the speaker does in the ironical case is merely to
pretend to assert the content of her utterance. Still, there is an element of
indirectness here, and we can maintain that irony also possesses a
secondary character. By pretending to assert something, the speaker
conveys something else, just as, in the other types of case, by asserting
something the speaker conveys something else. By pretending to say of
Paul that he is a fine friend in a situation in which just the opposite is
obviously true, the speaker manages to communicate that Paul is
everything but a fine friend. She shows, by her utterance, how
inappropriate it would be to ascribe to Paul the property of being a fine
friend. The utterance has a primary meaning — it expresses the
proposition that Paul is a fine friend; and it is by expressing that primary
meaning (under the 'pretense' mode, and without the force of a serious

94
assertion) that the speaker is able to convey what she conveys.
Whenever the meaning conveyed by an utterance is secondary and
derived from some antecedent meaning expressed by the utterance, it is
'nonliteral' in the ordinary sense. Let us therefore introduce a third notion
alongside t-literalness and m-literalness: that of p-literal meaning, where
'p' stands for 'primary'. An interpretation for an utterance is p-literal just in
case it directly results from interpreting the sentence (in context), without
being derived from some antecedently determined meaning by an
inferential process akin to that which is involved in conversational
implicatures, indirect speech acts, etc.
The question I want to raise concerns the relation between p-
literalness and m-literalness. We have seen that a meaning is m-literal if it
departs from the t-literal meaning of the sentence only minimally. Now we
have another notion, that of p-literal meaning. To what extent do they
coincide? In the example I used ('He is thirsty'), the proposition that Paul is
thirsty was both m-literal and p-literal, in contrast to the other proposition
(the implicature), which was neither. How general is this coincidence? More
specifically: Is a primary meaning necessarily a meaning that only
minimally departs from the t-literal meaning of the sentence? A positive
answer to that question tends to be assumed in the literature on this topic,
but it should be clear, from what I said in previous chapters, that I think
that is a mistake.

5.3. Nonminimal departures without secondariness

There are, I hold, meanings that are primary yet involve nonminimal
departures from the conventional meaning of the sentence. Examples
abound in the pragmatic literature. Let us start with two conjunctive
utterances which have been analysed and discussed many times:

(1) They got married and had many children

(2) The policeman raised his hand and stopped the car

95
 We naturally interpret (1) as depicting a situation in which marriage
took place before the coming of the children. Yet, according to Grice, this is
not encoded in the meaning of the sentence. Not only is the temporal
ordering of the events not part of t-literal meaning; it is not even part of
m-literal meaning.31 Remember how m-literal meaning was characterized:
m-literal meaning is not fully conventional since it involves, or may
involve, contextual ingredients (e.g. assignment of values to indexicals),
yet it departs from t-literal meaning only minimally, where a 'minimal'
departure from t-literal meaning is a departure that is itself governed by
the conventions of the language. But the temporal order between the
events described by the conjuncts in (1) is not something that the t-literal
meaning of the sentence forces an interpreter to specify. There are
conjunctive utterances similar in all formal respects to (1) which can be
given a quite satisfactory interpretation without specifying that order.
Grice concludes that the temporal ordering of the events, though strongly
suggested by the order of the conjuncts, is not part of 'what is said' by the
utterance. ('What is said' is Grice's term for the m-literal meaning.) The
contextual provision of that component of the meaning of the utterance
constitutes a 'nonminimal' departure from t-literal meaning.
Example (2) is similar. As Rumelhart pointed out, we naturally
interpret the sentence as depicting a scene in which the stopping of the
car was caused by the raising of the policeman's hand. Moreover, we
interpret the relevant form of causation as involving no direct physical
contact between hand and car but rather the mediation of intentional
states: the raising of the policeman's hand is understood to have caused
the driver (i) to recognize the policeman's intention that she should stop
the car and (ii) to stop the car so as to comply with the policeman's
request. In other words, we understand (2) as saying that the policeman
stopped the car by signalling to the driver that she was to stop. Yet neither
the form of causation, nor even the existence of a causal link between
hand-raising and car-stopping, is encoded in sentence (2). Nor is the
contextual provision of those elements of utterance meaning required in

31 This is controversial. See §6.2 for an alternative analysis.

96
virtue of some aspect of the t-literal meaning of the sentence. Once again
we find that the interpretation of the utterance involves nonminimal
departures from t-literal meaning.
Still, I maintain that the meanings thus conveyed by (1) and (2) are
p-literal. In conversational implicatures, indirect speech acts, irony etc.,
there is something which is said (or which the speaker 'makes as if to
say'), and something else that is implied by saying what is said. This
distinction between two separate components, one dependent upon the
other, is actually part of the meaning of the utterance: whoever
understands the utterance realizes that something is said and something
else is implied by saying it. For example, if I am asked whether I talked to
Freddy, and I reply

I don't talk to crooks

a normal hearer will understand me as saying that I don't talk to crooks

and thereby implying that I did not talk to Freddy. (Someone who does not
understand that does not understand the utterance.) Moreover, as Grice
insisted, the understander will grasp the inferential connection between
what is said (or the saying of it) and what is implied. (In this particular
case, the connection involves the premiss: 'Freddy is a crook'.) Faced with
(1) and (2), however, an interpreter does not construct an m-literal
interpretation — an interpretation that differs only minimally from t-literal
meaning — and use it to infer the m-nonliteral elements. The m-nonliteral
interpretations are arrived at directly, as a result of the interaction of the t-
literal meaning of the words (and constructions), salient features of the
speech situation, expectations created by the discourse, schemata stored
in memory and evoked by the words, etc. As I pointed out in Chapters 1 to
3, there is no inference here, or at least no inference of which the
interpreters themselves are aware.
Consider a third (and last) example which I already mentioned
several times:

97
(3) You are not going to die

In the context imagined by Kent Bach, a boy is crying because of a minor

cut and her mother utters (3). What she means is: 'You're not going to die
from that cut'. But m-literally, Bach points out, the utterance expresses the
proposition that the kid will not die tout court. The extra element
contextually provided (the implicit reference to the cut) is not a
component of m-literal meaning: it is not triggered by anything in the
sentence, nor is it necessary for the utterance to express a complete
proposition. On the other hand, that element is not 'inferred' on the basis
of m-literal meaning: the hearer does not construct the m-literal
interpretation, realize that the utterance, thus interpreted, would violate
Grice's maxim of quality by entailing something blatantly false (that the
child is immortal), and infer what the speaker's actually means; rather, she
directly understands the mother's utterance as characterizing the situation
created by the cut (Recanati 1999). Once again, the conveyed meaning
turns out to be m-nonliteral, but p-literal nonetheless.

5.4. The transparency condition

In the sort of case I have just discussed (the policeman example, etc.), the
interpretation of the utterance goes well beyond what the conventions of
the language dictate. The conventional meaning of the sentence is not
only 'completed' from the bottom up by assigning contextual values to
indexicals and other free variables, it is also enriched in a top-down
manner by appealing to background assumptions and world-knowledge.
That process of enrichment yields an output that is not m-literal, yet there
is nothing nonliteral in the ordinary sense in that output. Nonliteral
interpretations (in the ordinary sense) are 'special', by definition, but there
is nothing special about the interpretations of utterances like (2).
Utterance interpretation, in general, proceeds by matching the linguistic
meaning of the sentence to the particular situation or sort of situation the
speaker intends to talk about. The output of this matching process

98
typically is a richer meaning, as (2) clearly illustrates: a number of
features which are not linguistically encoded are nonetheless incorporated
into the described scene as a result of top down pressures. This is the
phenomenon of sense modulation, to be described in Chapter 9.
In contrast, there is something special about the interpretation of p-
nonliteral utterances; it is a two-step procedure instead of being a one-
step procedure. The interpreter first determines the utterance's primary
meaning, then infers some additional meaning. This two-step process
(which may be short-circuited or conventionalized — a complication which
I shall ignore) does not take place all the time: it takes place only when the
speaker conveys something indirectly. But the process of enrichment in
virtue of which what the speaker means by an utterance goes beyond the
conventional significance of his words is universal: there is no utterance,
however explicit, whose interpretation does not involve adjusting the
conventional meanings of words to the particulars of the situation talked
about. Communication would be impossible if things were not so.
The 'special' character of nonliteral communication is not the whole
story, however. Even if there were something special about m-
nonliteralness — if, say, most uses of words were m-literal — a use of
words would still not count as nonliteral in the ordinary sense merely in
virtue of the fact that it is m-nonliteral. A use of words counts as nonliteral
in the ordinary sense only if there is something special about that use that
is, or can be, perceived by the language users themselves. That is so
because nonliteralness is a feature of the interpretation of utterances, and
the interpretation of utterances is something that is bound to be available
to the language users who do the interpreting. Now m-nonliteralness per
se is not transparent to the language users. The speaker and hearer need
not be aware that in their understanding of the uttered words they are
going beyond what the conventions of the language dictate. It is not part
of their competence to reflect upon the complex cognitive processes
through which the meaning of an utterance gets built up from a number of
distinct sources. On the other hand, we have seen that whenever a
meaning is accessed via an inference from a primary meaning, as in

99
(genuine) conversational implicatures, the language users themselves are
aware of the distinction between the two layers of meaning as well as of
the connection between them. In other words, p-nonliteralness is
transparent to the language users; hence it counts as nonliteralness in the
ordinary sense. (This transparency is not a contingent property of p-
nonliteralness. It is definitive of p-nonliteralness that the sort of inference
at issue is conducted at the 'personal', rather than 'sub-personal', level
and is therefore available to the language users.)
To sum up, for something to count as nonliteral in the ordinary sense
it must not only go beyond the conventional significance of the uttered
words (m-nonliteralness), but it must be felt as such: the language users
must be aware that the conveyed meaning exceeds the conventional
significance of the words. That condition I dub the 'transparency
condition'. It is satisfied whenever the conveyed meaning has a secondary
character, as in conversational implicatures and indirect speech acts.

5.5. Varieties of nonliteral meaning

Is secondariness a necessary, or merely a sufficient condition of

transparency? If it is necessary, then only p-nonliteral instances of m-
nonliteral meaning (conversational implicatures, indirect speech acts, etc.)
will count as nonliteral in the ordinary sense. If it is merely sufficient, then
presumably there will be p-literal instances of m-nonliteral meaning which
will count as nonliteral in the ordinary sense, because the transparency
condition has been satisfied through something other than indirectness. I
see no reason to deny that there are such instances.
The paradigm case of nonliteral meaning is metaphor. Now
metaphor, in its most central varieties, I count as p-literal. To re-use an
example from Chapter 2, if I say that the ATM swallowed my credit card, I
speak metaphorically; there can be no real 'swallowing' on the part of an
ATM, but merely something that resembles swallowing. Still, an ordinary
hearer readily understands what is said by such an utterance, without
going through a two-step procedure involving the prior computation of the

100
'literal' meaning of the utterance (whatever that may be) and a secondary
inference to the actual meaning. Knowing the linguistic meaning of
'swallow', and knowing what sometimes happens with ATMs, the hearer
unreflectively constructs the sense in which the ATM can be said to
'swallow' the card by adjusting the meaning of the word to the situation
talked about. This is not very different from what goes on when the
meaning of words is enriched so as to fit the specific situation of discourse,
as in the policeman example from §5.3. In both cases, as Langacker
persuasively argued, the situation talked about is 'categorized' by means
of the words which are used to describe it. In enrichment the situation
talked about is a straightforward instantiation of the schema encoded in
the words, which schema therefore gets 'elaborated' through its
application to the situation. That defines what Langacker calls full
schematicity: "the target is compatible with the sanctioning unit [= the
word, with its conventional meaning], and is therefore judged by a speaker
to be an unproblematic instantiation of the category it defines" (Langacker
1987: 68). In contrast, partial schematicity occurs when "there is some
conflict between the specifications of the sanctioning and target
structures, so that the former can be construed as schematic for the latter
only with a certain degree of strain" (Langacker 1987: 69):

Because partial schematicity involves conflicting specifications, the

sanctioning and target structures cannot merge into a single,
consistent conceptualization; in a categorizing judgement of the
form [[SS ---> TS]], the discrepancy between SS and TS keeps them
at least partially distinct. The result is a bipartite conceptualization
including what we recognize as a literal sense (SS) and a figurative
sense (TS). On the other hand, nothing prevents the sanctioning and
target structures from merging into a unified conceptualization when
there is full consistency between their specifications. In the
schematic relationship [[SS ---> TS]], SS is in effect 'swallowed up'
by TS, since all of the specifications of the former are implicit in the
latter, which simply carries them down to a greater level of

101
 precision. (Langacker 1987: 92-93)

The picture that emerges is this. As words are applied, in context, to

specific situations, their meaning is adjusted. Depending on whether the
conventional meaning is fully or only partially schematic for the situation
talked about, adjustment will take one of two forms: sense elaboration
(enrichment), or sense extension (loosening). In sense elaboration the
meaning carried by the words is made more specific through the
interaction with contextual factors. In sense extension, those dimensions
of meaning which stand in conflict to the specifications of the target are
filtered out, but they remain somewhat active and may generate a feeling
of discrepancy between the evoked schema and the sense constructed by
(partially) applying the schema to the situation at hand. That feeling, like
the conflict which underlies it, comes in degrees. Hence there is a
continuum between ordinary cases of sense extension that we don't even
perceive (the ATM swallowing the credit card) and more dramatic cases of
metaphor whose nonliteral character cannot be ignored. 32 The more
noticeable the conflict, the more transparent the departure from t-literal
meaning will be to the language users. Beyond a certain threshold, cases
of sense extension will therefore count as special and nonliteral in the
ordinary sense, despite their p-literal character. They will count as
figurative uses of language (Fig. 5.1).

[Figure 5.1 about here]

Irony and nonserious uses of language are another type of case in

which, I think, the transparency condition is satisfied through something
other than indirectness. To be sure, irony possesses a secondary character
(§5.2): By pretending to say of Paul that he is a fine friend in a situation in
which just the opposite is obviously true, the speaker manages to
communicate that Paul is everything but a fine friend. She shows, by her

32 On this continuum, see D. Sperber and D. Wilson,'Loose talk',

Proceedings of the Aristotelian Society 86, 1986, pp. 153-71.

102
utterance, how inappropriate it would be to ascribe to Paul the property of
being a fine friend. But let us put aside what is thus implied by saying
something ironically — let us concentrate on the primary meaning of the
ironical utterance. In order to understand the utterance at the primary
level one must recognize that the act of asserting that Paul is a fine friend
is staged or simulated rather than actually performed. And that means
that one must discern two 'layers' within the primary meaning of the
utterance: the surface speech act wich the speaker pretends to perform,
and the ironical act of staging the performance of that speech act (Clark
1996: 353-4). This layering, internal to the primary meaning of the
utterance, characterizes 'staged communicative acts', a large family which
includes not only irony but also "sarcasm, teasing, overstatement,
understatement, rhetorical questions, and their relatives" (Clark 1996:
369). Insofar as the duality of layers is internal to the (primary) meaning of
the utterance and is recognized by whoever understands it, the
transparency condition is eo ipso satisfied: the utterance is intuitively
recognized as special, as exhibiting a duality which is absent from
standard ('serious') uses of language. That duality, characteristic of staged
communicative acts, is a third type of duality, distinct both from the 'two-
step procedure' at work in the recovery of secondary meanings, and from
the 'bipartite conceptualization' induced by metaphorical discrepancy. (Of
course, nothing prevents a single utterance from exhibiting the three
types of duality, in various patterns of interaction.)

5.6. Conclusion

The literal/nonliteral distinction turns out to cover two quite different

things:

• For the semanticist, the literal meaning of an expression is the semantic

value which the conventions of the language assign to that expression
('with respect to context', if the expression is indexical). Whenever the
meaning which the expression actually conveys departs from that literal

103
meaning, it is said to be 'nonliteral'.

• In the ordinary sense of the term, nonliteral meaning contrasts with

normal meaning. Nonliteral meaning is special, it involves a form of
deviance or departure from the norm; a form of deviance or departure
which must be transparent to the language users.

Both distinctions are legitimate, but they should be kept separate.

Yet they are standardly confused. The culprit here is the tacit, but very
widespread, assumption that the m-literal meaning of words is what, in
context, they normally express. In semantics (as opposed to psychology),
departures from m-literal meaning tend to be treated as special, on the
pattern of figurative language, nonserious speech or indirect
communication. I think that assumption is mistaken. In context, words
systematically express meanings that are richer than (or otherwise
different from) what the conventions of the language dictate. Contrary to
what formal semanticists tend to assume, the (intuitive) truth-conditions of
our utterances are not compositionally determined by the meanings of
words and their syntactic arrangement, in a strict bottom-up manner. They
are shaped by contextual expectations and world-knowledge to a very
large extent. That is true of all utterances, however 'literal' they are (in the
ordinary sense). If we abstract from those top-down factors, what we get
— the utterance's m-literal meaning — no longer corresponds to the
intuitive truth-conditions which the language users themselves associate
with their own utterances. In other words, there is a gap between the
deliverances of semantic theory and the intuitive content of utterances.
That gap is bridged by pragmatic processes which take place normally and
do not generate 'nonliteral meanings', except in special cases.

104
 Chapter 6

From Literalism to Contextualism

6.1. Five positions

I take it that there currently are five basic positions concerning the role of
context in the determination of truth-conditions. I have already attempted
to refute two of these positions. Before proceeding, it will be helpful to
identify all five, and to take a bird's eye view of the theoretical landscape.
The five positions I am about to list can be ordered on a scale, one
end of which is occupied by the extreme position I call Literalism, and the
other end by another extreme position, which I call Contextualism. The
debate between Literalism and Contextualism was at the forefront of
attention in the philosophy of language of the middle of the twentieth
century. It is widely believed to have been settled (in favour of Literalism)
but I think that is a mistake. The alleged refutations of Contextualism that
have been offered actually refute nothing; while Literalism strikes me as
by and large indefensible. So I think the history of twentieth-century
philosophy of language ought to be rewritten. I will not do so in this book,
however.
What is the Literalism/Contextualism debate exactly about? The
basic question, I think, is whether we may legitimately ascribe truth-
conditional content (the property of "saying" something, of expressing a
thought or a proposition) to natural-language sentences, or whether is it
only speech acts, utterances in context, that have content in a basic,
underived sense. Consider the type of formal language philosophers of the
first half of the century were concerned with. In these languages,
sentences are given an interpretation that is fixed and does not depend on
the context of use. Natural language sentences, by contrast, express a
complete thought (say something definite) only with respect to a context
of utterance — in many cases at least. This difference between natural

105
language and a certain type of formal language is well-known, and no one
has ever attempted to deny it. But there was disagreement as to the
importance of the distinction. Contextualists33 held that the difference
between the two types of language was all-important; natural-language
sentences, according to them, were essentially context-sensitive, and did
not have determinate truth-conditions. Literalists, on the other hand,
believed that the difference between the two types of language could be
abstracted from through a legitimate idealisation.
The literalist idealisation was based on the following claim:

(L) For every statement that can be made using a context-sensitive

sentence in a given context, there is an eternal sentence that can be
used to make the same statement in any context.

To obtain an eternal sentence from a context-sensitive one, one has only

to replace the indexical constituents of the latter by non-indexical
constituents with the same semantic value. Owing to (L), the difference
between natural languages and the formal languages in which the context
of utterance plays no role turns out not to be essential. Using natural
language, we could behave so as to abolish the difference - simply by
choosing to utter only eternal sentences. The reason why we also (and
mainly) use context-sensitive sentences is only that this enables us "to
speak far more concisely than otherwise" (Katz 1977: 20). 34
The contextualists denied (L), if only because they did not really

33 The contextualists I am talking about here are the so-called ordinary

language philosophers: Wittgenstein, Austin, Strawson, etc. (with the

exception of Grice, who was not a contextualist).

34 Note that (L) is much weaker than another principle of effability, namely

(T):

(T) Every entertainable thought may be expressed by means of an

eternal sentence the sense of which corresponds exactly to that

thought.

106
believe that there were eternal sentences. Against eternal sentences, it
can be argued, on contextualist grounds, that fundamental semantic
mechanisms such as reference, quantification, or predication are
essentially context-sensitive: there is no way to refer, or to predicate, or to
quantify without relying on features of the context of utterance. It follows
that sentences, by themselves, don't express propositions.
Whatever we think of Contextualism, there are not many amateurs
of eternal sentences nowadays. It is more or less accepted that natural

Many philosophers have (rightly) argued against (T). For example, Sperber

and Wilson write:

It seems plausible that in our internal language we often fix time and

space references not in terms of universal coordinates, but in terms

of a private logbook and an ego-centred map; furthermore, most

kinds of reference — to people and events for instance — can be

fixed in terms of these private time and space coordinates. Thoughts

which contain such private references could not be encoded in

natural languages but could only be incompletely represented.

(Sperber & Wilson 1986b: 192).

(L) is not subject to this criticism. (L) says only that every statement can
be made using an eternal sentence, not that every thought can be literally

expressed by an eternal sentence. Now, a statement may be of an object,

in the sense that it may be about a certain object without involving a

particular mode of presentation of that object. Such a de re statement

corresponds to a class of thoughts, each involving a particular (and,

perhaps, private) mode of presentation of the object referred to. The fact

that, in the thought, there are private modes of presentation attached to

the objects referred to implies that there are thoughts that cannot be

totally and adequately represented by means of eternal sentences, but

does not imply that there are statements that cannot be made by means

of eternal sentences: statements are public objects at a more abstract

107
language sentences are irreducibly context-sensitive. But there are
fallback positions for Literalism. The strongest fallback position consists in
acknowledging the extent of context-dependence, while insisting that it
still is the sentence which, in virtue of the rules of the langage, expresses
a proposition with respect to the context. This semantic notion of the
content of a sentence (with respect to context) is distinct from the
pragmatic notion of the content of a speech act. For it is the linguistic
rules, not the speaker's beliefs and intentions, which fix the content of the
sentence with respect to context. So Literalism in its modern form holds
that the truth-conditions of a sentence are fixed by the rules of the
language quite independent of speaker's meaning.
So construed, Literalism must be rejected because, as we have seen,
semantic underdeterminacy makes it unavoidable to appeal to speaker's
meaning in determining truth-conditions. Even though there still are
theorists who argue in its favour, I think Literalism is clearly a nonstarter
and I will not be concerned with it in what follows.
The next two fallback positions correspond to weaker forms of
Literalism. According to them, we need to appeal to speaker's meaning in
determining truth-conditions, but we do so only when the sentence itself
demands it. In other words, optional pragmatic processes are not allowed
to affect what is said; what is said obeys the Minimalist constraint (Chapter
1).
The weaker of the two positions in question is what I dubbed the
Syncretic View (Chapter 4). It acknowledges the fact that what is said, in
the intuitive sense, may be affected by primary pragmatic processes of
the optional variety; but it draws a distinction between what is said in the
intuitive sense (the input to secondary pragmatic processes), and what is
strictly and literally said. What is strictly and literally said, at least, obeys
the Minimalist constraint.
The other position is not as weak as that. For reasons that will
become apparent, I call it Indexicalism. It denies that what is said, in the

level than thoughts, and as such do not contain private modes of

presentation.

108
intuitive sense, can be affected by pragmatic processes of the optional
variety. No contextual influences are allowed to affect the intuitive truth-
conditional content of an utterance unless the sentence itself demands it.
This position can be thought of as a research programme: whenever it is
argued that the truth-conditions of an utterance are affected by context in
a top-down manner, the indexicalist feels compelled to re-analyse the
example so as to show that the pragmatic process at issue is an instance
of saturation after all.
So far I have listed three positions. Literalism denies that speaker's
meaning plays any role in fixing the truth-conditions of sentences.
Indexicalism allows speaker's meaning to play a role, but only when there
is a slot to be filled in logical form; no top down influences can affect truth-
conditions. The Syncretic View acknowledges such influences, but limits
their effects to what is said in the intuitive sense, as opposed to what is
strictly and literally said (the minimal proposition).
The next two positions I want to consider fully acknowledge the role
of pragmatic processes, including pragmatic processes of the optional
variety, in the determination of what is said. What is said is taken to be
pragmatically determined, in a way that is sensitive to top-down
influences. In contrast to the Syncretic View, those positions do not find it
necessary to postulate a level of content that is unaffected by top-down
factors. So they fall squarely on the contextualist side. They differ from
each other in their respective attitudes towards the minimal proposition
posited by the Syncretic View. One position, which I call Quasi-
Contextualism, simply considers the minimal proposition as a theoretically
useless entity which plays no role in communication (§4.5). Contextualism
tout court goes much further and denies that the notion even makes
sense.
To sum up, the five positions in the ballpark are:

Literalism
Indexicalism
The Syncretic View

109
 Quasi-contextualism
Contextualism

Literalism and the Syncretic View have been explicitly criticized, and I
have implicitly endorsed Quasi-Contextualism in arguing against the
Syncretic View in Chapter 4. So two positions remain to be introduced and
discussed: Indexicalism, and Contextualism.

6.2. Indexicalism

The pragmatics literature is replete with examples in which what is said in

the intuitive sense seems to be affected by pragmatic processes of the
optional (top-down) variety — what King and Stanley call 'strong pragmatic
effects'. Yet, as Steven Levinson (one of the providers of examples) puts it,
"there will always be doubt about whether a better semantic analysis of
the relevant construction might not accommodate the apparent pragmatic
intrusion in some other way" (Levinson 2000: 214). Indeed, in most cases,
there are different ways of analysing a given example. Some of those
analyses may well suggest that the pragmatic process at issue is not a
primary pragmatic process of the optional variety, but a regular instance
of saturation.
Consider the phenomenon of quantifier domain restriction. I may
utter the sentence 'Most students are male' in order to assert that most
students in my class are male. This is arguably a case in which what is
asserted goes beyond what is linguistically encoded, for I tacitly refer to
my class. Moreover, the additional ingredient in the meaning of the
utterance which is thereby provided seems to be optional, for there is a
possible reading of 'Most students are male' in which no such contextual
restriction of the domain occurs. (I call that an 'absolute' reading for the
quantifier phrase.) So quantifier domain restriction is a prima facie
candidate for the status of 'strong pragmatic effect': a truth-conditionally
relevant aspect of what is asserted seems to result from an optional
pragmatic process. There are many ways of analysing the phenomenon,

110
however. We may consider that the predicate 'student' is contextually
enriched and expresses the ad hoc concept STUDENT_IN_MY_CLASS. We
may also consider the class as an unarticulated constituent of what is said,
like the location of rain in 'It's raining'. But it is also possible to consider
that, in the logical form of the sentence, there is a slot to be contextually
filled: a variable corresponding to the domain of quantification, which
variable can be associated either with the determiner or (more plausibly
perhaps) with the common noun 'student'. On such a view the semantic
analysis of the sentence-type will reveal a hidden quasi-indexical
component, namely the domain variable d:

Mostd (students) (are male)

Most (studentsd) (are male)

The suggestion is that, when the sentence is uttered, a value has to be

contextually assigned to the variable, thereby restricting the domain of
quantification. If we take this line, we treat quantifier domain restriction
as, in effect, a case of saturation (Stanley and Szabo 2000). In this
framework, the cases in which the quantifier phrase is given an 'absolute'
reading are treated as cases in which a value is contextually assigned to
the variable, but intuitively no restriction takes place because the
contextually selected domain is the whole universe (or the entire domain
relevant to the variable, in case we use sorted variables).
Or consider the examples of enrichment with which our discussion of
Minimalism began:

I've had breakfast <this morning>

You're not going to die <from that cut>

The contextually provided ingredients (within angle brackets) seem to

correspond to no element in the sentence; they are putative 'unarticulated
constituents'. But this is not the only possible analysis. According to event
semantics, such sentences quantify over events, hence it is possible to

111
extend the analysis in terms of quantifier domain restriction to them. In
'The pragmatics of what is said' (Recanati 1989) I suggested that the
intuitive difference between 'I've had breakfast' ('restricted' reading) and
'I've been to Tibet' ('absolute' reading) can be handled in those terms:

In both cases what is conveyed by virtue of linguistic meaning alone

is that, in some temporal domain x prior to the time of utterance, 35
there is a certain event, viz. the speaker's having breakfast or his
going to Tibet; but in the first case, the time interval is contextually
restricted to the day of utterance, while in the second case the
relevant interval is more extended and covers the speaker's life (up
to the time of utterance). (Recanati 1989/1991: 104)

If that is so then the intuitive understanding of 'I've had breakfast' in the

relevant context results from saturation rather than free enrichment. 36 The
same sort of thing can be said about the other example. If the sentence is
interpreted as quantifying over events ('you're not going to die' = 'there is
no future event e such that e is your death'), we can say that the
contextually selected domain of quantification is the set of events which
causally follow from the cut. In my 1989 paper I offered a similar analysis
for Sperber and Wilson's favorite example of enrichment, 'It will take us
some time to get there'.
Another type of example for which two sorts of analysis are possible
involves the temporal understanding of conjunction, as in:

35 The domain in question is "a set of events defined by a time interval".

36 The same sort of analysis can be provided even if we don't take the

sentence to quantify over events, but, in virtue of the tense it contains,

over times. 'I've had breakfast' then says that there is a past time t such

that the speaker has breakfast at t. The contextually specified domain of

quantification now is a set of times (rather than a set of events). Like the

event-theoretic analysis, this analysis can be extended to 'You're not going

to die'.

112
They got married and had many children

In chapter 5 I said that the temporal order of the events which is intuitively
part of the truth-conditions corresponds to nothing in the sentence and is
provided through free enrichment. Indeed there is a possible reading of
that sentence (or formally similar sentences) in which no temporal order is
implied. But an alternative analysis in terms of saturation is available
(Carston 1988). Barbara Partee famously suggested that the role of the
past tense is, or may be, to refer to a particular time in the past. 37 Let's
take that for granted. Then, to determine what is said by means of the
sentence 'They got married and had many children', the hearer must
assign a reference to each of the referring expressions, including the past
tense 'got married' and 'had'. Just as pragmatic principles are employed in
ascertaining the referent of 'they', so, Carston says, they are used in
assigning temporal reference. The hearer goes beyond the strict semantic
content of the sentence uttered, and on the basis of contextual
assumptions and pragmatic principles recovers from 'They got married
and had many children' a representation such as 'John and Mary got
married at t and had many children at t+n'.

t is some more or less specific time prior to the time of utterance

and t+n is some more or less specific time, later than t. The
temporal ordering of the events described in the conjuncts is thus
treated as a by-product of the reference assignment process
involved in determining [what is said] (Carston 1988:161).

As I pointed out in 'The pragmatics of what is said', this analysis faces

difficulties when the past tense is replaced by the present perfect, as in
Jonathan Cohen's example ('The old king has died of a heart attack and a
republic has been declared'). That is so because the present perfect can

37 See B. Partee, 'Some structural analogies between tenses and

pronouns in English', Journal of Philosophy 70, 1973, p. 602-3.

113
hardly be considered as referring to a specific time. But if, plausibly
enough, we construe the present perfect as quantifying over times, we can
extend the analysis in terms of quantifier domain restriction to that type of
example also. The sentence will be understood as saying that there is a
time t in the past such that the old king dies of a heart attack at t, and
there is a time t' in the past such that a republic is declared at t'. If the
domain to which the second quantifier is indexed is contextually restricted
to the set of times that are posterior to t, we account for the relevant
temporal reading. So it seems that there always is a way of analysing
Gricean examples of temporally understood conjunctions in terms of
saturation rather than enrichment.
Examples can be multiplied at will. The indexicalist position consists
in systematically offering an analysis in terms of saturation, whenever the
intuitive truth-conditions of an utterance are affected by contextual
factors. As Jason Stanley, the chief representative of this research
program, puts it, "all effects of extra-linguistic context on the truth-
conditions of an assertion are traceable to elements in the actual syntactic
structure of the sentence uttered" (Stanley 2000: 391). To maintain such a
sweeping claim, one must deal with every putative counterexample, in
order to establish, in each case, the conclusion that "we have been given
no reason to abandon the thesis that the only truth-conditional role of
context is the resolution of indexicality, broadly construed" (Stanley 2000 :
401; see also Stanley and Szabo 2000, King and Stanley forthcoming.)
At this point the debate becomes technical (and, fortunately,
empirical). It's a matter of detailed analyses of particular examples (or
classes of examples). Without going into those details, it is fair to say that
the indexicalist starts with a significant disadvantage, for he makes a
universal claim while his opponent only makes an existential claim. For his
opponent to win, it is sufficient to produce one convincing example of a
strong pragmatic effect. But the indexicalist is condemned to deal with all
putative cases, and to show that they are not what they seem to be. A
hopeless task, unless the indexicalist can develop methods, or strategies,
for dealing with large classes of cases at once. Unsurprisingly, such

114
general methods are what the indexicalist is after. Stanley has come up
with one strategy, in particular, that seems very powerful and which he
has applied both to the case of quantifier domain restriction and to
'unarticulated constituents' more generally. In Chapter 7 I will discuss that
strategy and show that it is not as successful as Stanley thinks.

6.3. Contextualism

According to Contextualism — a provocative view which certainly deserves

to be explored — there is no level of meaning which is both (i)
propositional (truth-evaluable) and (ii) minimalist, i.e. unaffected by top-
down factors. As I put it in Direct Reference, "no proposition could be
expressed without some unarticulated constituent being contextually
provided" (Recanati 1993: 260). Such a radical view has been defended by
Charles Travis, John Searle, and a few others. Searle, for example, argues
that a determinate proposition can be expressed only against a
background of unarticulated assumptions. He gives the following example
of unarticulated assumptions at work in understanding a simple utterance:

Suppose I go into the restaurant and order a meal. Suppose I say,

speaking literally, 'Bring me a steak with fried potatoes.' (...) I take it
for granted that they will not deliver the meal to my house, or to my
place of work. I take it for granted that the steak will not be encased
in concrete, or petrified. It will not be stuffed into my pockets or
spread over my head. But none of these assumptions was made
explicit in the literal utterance. (Searle 1992: 180)

Though unarticulated, those assumptions contribute to determining the

intuitive conditions of satisfaction (obedience-conditions, truth-conditions,
etc.) of the utterance. The order 'Bring me a steak with fried potatoes'
does not count as satisfied if the steak is delivered, encased in concrete,
to the customer's house. It is mutually manifest to both the hearer and the
speaker that the speaker intends the ordered meal to be placed in front of

115
him on the restaurant table he is sitting at, etc.
In that sort of case a syncretist will insist that what is strictly and
literally said is free from the relevant assumptions: the order 'Bring me a
steak with fried potatoes' would be literally satisfied if the steak was
delivered, encased in concrete, to the customer's house. (See Soame's
treatment of the 'coffee' example in §4.4.) That is precisely what the
contextualist denies. Another example given by Searle will help to make
sense of the contextualist position. The word 'cut' is not ambiguous, Searle
says, yet it makes quite different contributions to the truth-conditions of
the utterance in 'Bill cut the grass' and 'Sally cut the cake'. That is
because background assumptions play a role in fixing satisfaction-
conditions, and different background assumptions underlie the use of 'cut'
in connection with grass and cakes respectively. We (defeasibly) assume
that grass is cut in a certain way, and cakes in another way. The assumed
way of cutting finds its way into the utterance's truth-conditions:

Though the occurrence of the word "cut" is literal in [both]

utterances..., and though the word is not ambiguous, it determines
different sets of truth conditions for the different sentences. The sort
of thing that constitutes cutting the grass is quite different from,
e.g., the sort of thing that constitutes cutting a cake. One way to see
this is to imagine what constitutes obeying the order to cut
something. If someone tells me to cut the grass and I rush out and
stab it with a knife, or if I am ordered to cut the cake and I run over it
with a lawnmower, in each case I will have failed to obey the order.
That is not what the speaker meant by his literal and serious
utterance of the sentence. (Searle 1980: 222-223)

Now the syncretist will insist that a sentence such as 'Cut the grass'
expresses something that has literal conditions of satisfaction quite
independent of any background assumption; something very abstract,
involving the constant, underspecified meaning of 'cut' and not the
definite sense it takes on particular uses (or types of use). Stabbing the

116
grass with a knife and running over it with a lawnmower are two ways of
literally obeying the order 'Cut the grass', on this view. But the
contextualist stands skeptical. To get something genuinely evaluable, he
claims, i.e. something which enables us to partition possible worlds into
those in which the relevant condition is satisfied and those in which it is
not, we need background assumptions. We cannot specify a determinate
proposition which the sentence can be said literally to express, without
building unarticulated assumptions into that proposition. The best we can
do is to construct a disjunction of the propositions which could be
determinately expressed by that sentence against alternative background
assumptions.
In support of this claim, Searle sets up an example for which no
background assumption is readily available: 'Cut the sun'. What counts as
obeying that order? We don't quite know. The abstract condition we can
associate with that sentence (involving some form of linear separation
affecting the integrity of the sun) is, precisely, too abstract to enable us to
tell the worlds in which the condition is satisfied from the worlds in which it
is not. It is not determinate enough to give us specific truth-conditions or
obedience-conditions.38

6.4. Literalist responses to the contextualist challenge

According to Emma Borg (who follows the lead of other radical

38 In Recanati 1997a, p.120 I gave a real-life example of the phenomenon

Searle is drawing our attention to. In the 1936 movie Desire, by Frank

Borsage, Gary Cooper meaningfully utters the imperative sentence

'Disarm the fricassee'. In context, the obedience conditions of the

utterance are quite clear: we know which state of affairs the addressee

must bring about to obey the order. Without a proper background,

however, we no more know the obedience conditions of this utterance

than we know the obedience conditions of 'Cut the sun'.

117
syncretists,39 such as Herman Cappelen and Ernie Lepore), the fact that we
are unable to specify intuitive conditions of application for the predicate
'cut the sun' does not support the contextualist conclusion that sentences
per se do not have truth-conditions. There is, she claims, a crucial
difference between "knowledge of truth-conditions and the knowledge that
truth-conditions are satisfied". We may know the obedience-conditions of
'Cut the sun' in a purely 'disquotational' manner (i.e. we may know that
'Cut the sun' is obeyed iff the addressee cuts the sun), without knowing
what counts as cutting the sun, in the context at hand. So there is no
reason to deny sentences genuine truth-conditions. The sentence 'Oscar
cuts the sun' does possess truth-conditions; such truth-conditions are
determined by a recursive truth-theory for the language, which issues
theorems such as 'Oscar cuts the sun is true iff Oscar cuts the sun'. We
know those truth-conditions provided we know the language. What we
don't know, simply in virtue of knowing the language, is "a method of
verification for those truth-conditions". This, then, is the syncretist's
ultimate reply to the contextualist. According to the syncretist, the
contextualist is guilty of endorsing a form of (so-called) 'verificationism'.
This move strikes me as an unacceptable weakening of the notion of

39 As King and Stanley (forthcoming) point out, Syncretism — what they

call 'semantic modesty' — "is a position that comes in degrees. On the one

extreme, there are those who hold that most of our intuitions about what

is said by a sentence in a context are affected by strong pragmatic effects.

On the other end of the spectrum, are theorists who hold that while

generally intuitions about what is said by a sentence are reliable guides to

semantic content, there are a restricted range of cases in which strong

pragmatic effects may affect speaker intuitions about what they take to be

the semantic content of a sentence." Cappelen and Lepore are radical

syncretists. They argue for systematically disconnecting the semantic

content which the theorist assigns to a sentence from what is intuitively

said by uttering that sentence (see Cappelen and Lepore 1997).

118
truth-condition. The central idea of truth-conditional semantics (as
opposed to mere 'translational semantics') is the idea that, via truth, we
connect words and the world.40 If we know the truth-conditions of a
sentence, we know which state of affairs must hold for the sentence to be
true; and that means that we are able to specify that state of affairs.41 T-
sentences display knowledge of truth-conditions in that sense only if the
right-hand-side of the biconditional is used, that is, only if the necessary
and sufficient condition which it states is transparent to the utterer of the
T-sentence. If I say 'Oscar cuts the sun is true iff Oscar cuts the sun',
without knowing what it is to 'cut the sun', then the T-sentence I utter no
more counts as displaying knowledge of truth-conditions than if I utter it
without knowing who Oscar is (i.e. if I use the name 'Oscar' deferentially,
in such a way that the right hand side is not really used, but involves some
kind of mention).
The contextualist challenge is likely to elicit another unsatisfactory
response, this time from the indexicalist. To each dimension of
indeterminacy, the indexicalist may argue, there corresponds a slot in
logical form, which must be filled for the utterance to say something
definite. To illustrate that point, let us consider another contextualist
example from Searle (Searle 1983, pp.145-7).
When we ask someone to open the door, the content of the request
goes beyond what is linguistically encoded. Not only is it necessary for the
addressee to identify the relevant door (i.e. to complete or otherwise
enrich the incomplete definite description 'the door'). She must also
determine in what sense the door must be 'opened'. Besides doors and
windows, eyes and wounds can be opened. Now if the addressee 'opened'
the door by making an incision in it with a scalpel, as when opening a

40 See Lewis 1970:xx, Evans and McDowell 1975:xx.

41 In §1.5 I gave an example of the way knowledge of truth-conditions can

be tested, by asking subjects to pair sentences with situations. Pure

'disquotational' knowledge of the sort invoked by Emma Borg does not

suffice to pass that sort of test.

119
wound, she would not have satisfied the request. Still, in a special context,
it could be that the request to open the door must be satisfied precisely by
incising it by means of a scalpel. The manner of opening is thus defeasibly
indicated by context, it is not determinable on the basis of just the
linguistic meaning of the sentence (including the direct object of the verb).
To be sure, we can make it explicit in the sentence itself by introducing
supplementary details, but each addition of this sort cannot fail to
introduce other indeterminacies. If, for example, we add that the door
must be opened 'with a key', we don't specify whether the key must be
inserted into the lock or rather used like an axe to break the door open
(Searle 1992, p.182). However explicit the sentence, there will always be
some aspect of truth-conditional content that is contextually determined
without being explicitly articulated.
At this point, the imagined indexicalist response consists in saying
that, like all verbs, 'open' is associated with a complex frame 42 — the
opening frame — involving a certain number of argument roles: a location
playing the role of INSIDE; another location operating as OUTSIDE; a
BOUNDARY separating the two; a MOVING OBJECT liable to pass from
inside to outside (or the other way round); an OBSTACLE, that is, an entity
preventing the passage of the moving object; an AGENT liable to free the
passage by means of ACTION on the obstacle; an INSTRUMENT serving to
accomplish the action; and so on and so forth. In context, each of the
variables I have enumerated must be assigned a particular value: the
inside, the outside, the obstacle, the site of passage etc., all must be
contextually identified. In the case of 'opening a wound', the inside is the
interior of the wound, the outside is the exterior of the body, the moving
object is the puss or internal secretions of the wound and so on. This
contextual instantiation of the variables is what determines the specific
interpretation assigned to 'open' in a given context, and it is is no different
from what is required for interpreting 'John's boat' or 'He came' or any
other context-sensitive expression in need of saturation. It is therefore

42 The notion of frame which I am using is that elaborated by Fillmore in a

series of papers; for a general presentation see Fillmore 1982, 1985, 1991.

120
unnecessary to modify semantic theory in order to give an account of
Searle's examples; it is enough to extend the list of context-sensitive
expressions, so as to include all verbs (insofar as they are all associated
with frames which comprise a number of argument roles, the fillers of
which must be contextually assigned).
This indexicalist response is no more convincing than the syncretist
response was. Let's admit that the verb 'to open' is associated with the
complex frame I have mentioned. Does that make it an indexical or
context-sensitive expression, whose use triggers, indeed mandates, a
primary pragmatic process of saturation? No. There is an important
difference between the argument roles of a frame and the indexical
variables associated with context-sensitive expressions. Indexical variables
must be contextually assigned values for the expression to acquire a
definite semantic content. If the referent of 'he' in 'He boarded John's boat'
is not contextually specified, or if the relation between John and the boat
remains indefinite, the utterance does not have definite truth conditions.
In contrast, the argument roles of a frame may but need not be assigned
contextual values. The contextual assignment process is optional; it may,
or may not, take place, depending on what is contextually relevant. In
other words, it is the context (not the sentence) which determines which,
among the many argument roles of a given frame, are contextually
assigned particular values, and which remain indefinite (existentially
quantified). In many contexts, it is of no importance whether the door is
opened with a key or in another way; what counts is simply that it is
opened. To be sure, for any given verb (or verb plus syntactic context),
there is a small number of argument roles in the frame for which the
contextual assignment of value is linguistically mandated; but the
indexical response presupposes something much stronger: that the verb
'open' is like an indexical expression, which acquires a definite content
only when the argument roles of the associated frame (all the argument
roles, insofar at they can all be contextually foregrounded) are
contextually assigned values. That is evidently too strong. In a given
context, many of the argument roles which feature in the frame are

121
existentially quantified rather than contextually assigned values. This does
not prevent the verb 'open' from expressing a definite content, in such a
context.
To sum up, for indexicals it is the conventional meaning of the
expression which triggers the process of saturation and makes it
mandatory. With ordinary expressions such as 'open', it is the context, not
the conventional meaning of the expression, which is responsible for
foregrounding certain aspects of the described situation and triggering a
process of contextual specification which goes well beyond what is
linguistically encoded. The process in question is top-down, not bottom up.
It is a pragmatically controlled pragmatic process, rather than a
linguistically controlled pragmatic process, like saturation.

6.5. Where Indexicalism and Contextualism meet

In §6.1. I described Indexicalism as the strongest fallback position for a

literalist forced to acknowledge that speaker's meaning plays a role in
fixing truth-conditional content. According to Indexicalism, it is the
sentence which, via the conventional meaning of the context-sensitive
expressions it contains, triggers and controls the appeal to speaker's
meaning. Speaker's meaning thus plays a role in the determination of
truth-conditional content, but it does so only when the sentence itself sets
up slots to be pragmatically filled (Minimalism). The Syncretic View is a
weaker form of literalism since it concedes that the intuitive truth-
conditions of an utterance may not satisfy the Minimalist constraint, which
holds only for 'what is strictly and literally said' (the minimal proposition).
Next comes Quasi-Contextualism, which differs from the Syncretic View in
downplaying the minimal proposition. For the quasi-contextualist, the
minimal proposition literally expressed by the sentence exists in semantic
heaven but plays no role in the actual process of communication. Finally,
Contextualism denies the very existence and possibility of the minimal
proposition, insofar as it is determined in a purely bottom-up manner:
without pragmatic intrusions of the optional (top-down) variety, no

122
determinate proposition could be expressed.
Even though it is useful and illuminating to order the five positions
on the scale from Literalism to Contextualism, as I have done, their
relations are more complex than a simple ordering can suggest. In
particular, there is a dimension along which Indexicalism and
Contextualism are close neighbours.
One must not forget that the gist of Literalism is an attempt to
minimize context-sensitivity, while Contextualism takes it to be an
essential feature of natural language (as opposed to formal languages).
From this point of view, the indexicalist position is somewhat ambiguous.
On the one hand, by rejecting contextual influences on content that are
not linguistically controlled, the Indexicalist does his best to preserve the
literalist doctrine that linguistic entities are genuine bearers of content,
context-sensitivity notwithstanding. On the other hand, by multiplying
indexical variables, the indexicalist widens the gap between sentence
meaning and propositional content, in the spirit of Contextualism.
Indexicalism can thus be seen as a natural ally of Contextualism. 43

43 In 'The Pragmatics of what is said' I argued that the Minimalist Principle

can serve as a methodological weapon in the hand of the contextualist:

Since, for a defender of the Minimalist Principle, what is said departs

from what the sentence means only insofar as there is in the

meaning of the sentence a slot to be contextually filled, he is led to

posit new slots, new dimensions of semantic indeterminacy, every

time the following condition obtains: the traditional slots (identity of

the speaker, time of utterance, etc.) have been filled, yet it seems

that the meaning of the sentence still underdetermines what is said.

(...) It follows that there are more slots to be filled for a sentence to

express a complete proposition, from the point of view of a

minimalist, than there is from the point of view of someone who

rejects the Minimalist Principle. (...) I conclude that the Minimalist

Principle is an incentive to maximize context-sensitivity; it leads its

123
 Likewise, Contextualism can be construed as an extreme form of
Indexicalism that generalizes context-sensitivity so as to rule out eternal
sentences not merely de facto but de jure. It's not just that there are
expressions whose meaning is schematic and involves contextual
variables; it is linguistic meaning in general which suffers from a form of
indeterminacy which makes it unfit to carry content save against a rich
contextual background. Owing to that indeterminacy, some form of
enrichment or contextual elaboration becomes mandatory for the
sentence to express a definite proposition. The reason why Contextualism
and Indexicalism become so close is precisely that, with Contextualism,
the distinction between mandatory and optional pragmatic processes is
somewhat blurred.
According to Contextualism, the meaning of words and/or phrases
(whether indexical or not) is not determinate enough to yield even a
minimal proposition. In chapter 9 I will consider various arguments in
favour of that claim, and various contextualist positions corresponding to

defenders to widen the gap between linguistic meaningfulness and

full propositionality. From a contextualist point of view, this is a good

reason for maintaining the Minimalist Principle as far as possible. (...)

What I have just said shows that it is a mistake to claim, as

Carston does, that the Minimalist Principle is partly responsible for

the usual underrating of the gap between sentence meaning and

what is said. Carston believes that partisans of the Minimalist

Principle "assume that the domain of grammar, sentences, and the

domain of truth-conditional semantics, propositions, are essentially

the same" (Carston 1988:164). But this is not at all the case. Quite

the contrary, defenders of the Minimalist Principle take the meaning

of a sentence to be far less "propositional", much more

underdetermined as far as truth-conditions are concerned, than is

ordinarily supposed. (Recanati 1989/1991: 117-8)

124
these arguments. One extreme position denies that words have meanings
in anything like the traditional sense. Another, less extreme view takes
takes polysemous words such as 'open' or 'cut' as paradigmatic and
generalizes the form of underdeterminacy they reveal. According to both
of these views, the meaning of a predicate is not a determinate concept: it
is a 'semantic potential', which only determines a concept against a rich
pragmatic context.44 At this point Contextualism becomes hard to
distinguish from (an extreme form of) Indexicalism. Nearly all words are
like indexicals, whose contents must be contextually determined. 45 As
Sperber and Wilson say, "words are used as pointers to contextually
intended senses" (1998: 200).46

44 A third contextualist position to be discussed in Chapter 9 construes the

process of semantic composition as, to a large extent, pragmatically

sensitive, in such a way that complex phrases must be assigned a sense,

in context, even if the constituent words all possess a definite sense.

45 For an explicit statement to that effect, see Georgia Green, "Some

Remarks on How Words Mean" (197xx).

46 The difference between the contextualist positions mentioned in the

text and the view of Sperber and Wilson is that Sperber and Wilson take
words also to encode concepts. Sperber and Wilson are quasi-

contextualists. They accept that ordinary words have contents, in and by

themselves, and that sentences containing them express propositions

(once indexicals etc. have been assigned semantic values) — even though

they take these contents and these propositions to be communicationally

irrelevant:

Many words seem to encode not a full-fledged concept but what

might be called a pro-concept. (...) As with pronouns, their semantic

contribution must be contextually specified for the associated

utterance to have a truth-value. (...) All words behave as if they

encoded pro-concepts: that is, whether or not a word encodes a full

125
 Chapter 7

Indexicalism and the Binding Fallacy

7.1. Mandatory v. Optional

What characterizes contextual ingredients of the optional variety is the

fact that their contextual provision is not mandatory — it is not required in
virtue of a linguistic convention governing the use of a particular
construction (or class of constructions). In context, it may be that that
ingredient is ‘required’; but then it is required in virtue of features of the
context, not in virtue of linguistic properties of the expression-type. A
contextual ingredient is mandatory in the relevant sense, and is provided
through saturation, only if in every context such an ingredient has to be
provided (precisely because the need for saturation is not a contextual
matter, but a context-independent property of the expression-type). This,
then, is the criterion we must use for deciding whether a contextual
ingredient results from an optional pragmatic process or from saturation:
Can we imagine a context in which the same words are used normally, and
a truth-evaluable statement is made, yet no such ingredient is provided?
To illustrate the contrast between the two types of pragmatic
process, let us consider the phenomenon of 'null instantiation' (to use
Fillmore's terminology), where the direct object of a transitive verb is not
syntactically realized, or at least not overtly. There are two sorts of case,

concept, the concept it is used to convey in a given utterance has to

be contextually worked out. (Sperber and Wilson 1998: 185)

In contrast, a radical contextualist denies that words have full-fledged

contents or encode concepts. All words — or, more cautiously, nearly all

words — encode only 'pro-concepts' (semantic potentials) and serve as

pointers to intended senses. In Thoughts and Utterances, pp. 359-64,

Robyn Carston tentatively amends relevance theory along those lines.

126
which must be sharply distinguished (Fillmore 1986, Fillmore and Kay
1994). In indefinite null instantiation (INI), the argument role
corresponding to the direct object is existentially quantified instead of
being assigned a particular value. In that type of case the suppression of
the direct object arguably changes the semantic value of the verb: it
denotes not the original two-place relation, but a property generated by
existentially quantifying the object argument-role of the original relation.
As Quine pointed out, from any n-place predicate P, one can generate an
n-1 place predicate by applying to P an operator he calls ‘Derelativization’,
which he describes as follows:

(Der P) x1...xn-1 iff there is something xn such that Px1...xn

If ‘P’ is a two-place predicate, ‘Der P’ will be a genuine one-place

predicate, denoting a property rather than a relation (Quine 1960: 229-
231). In the case of 'eat' and other verbs subject to INI, suppressing the
direct object in surface syntax amounts semantically to the same result as
applying ‘Der’ to the original two-place predicate. Intransitive 'eat' thus
denotes the property of eating (Eat1), namely the property one has in
virtue of filling the Eater argument-role in some instance of the dyadic EAT
relation (Eat2). One eats1 iff there is something that one eats2.
In definite null instantiation (DNI) the situation is quite different. Like
'eat', 'hear' and 'notice' are transitive verbs which have an intransitive use:
they can be used without a direct object, as in 'I heard' or 'I noticed'. When
that is the case, however, there must be something definite, in context,
which the speaker is said to have heard or noticed. To understand what is
said, an interpreter must identify that thing. This shows that the objectless
verb still denotes the original two-place relation, even though the second
argument-role is not realized in surface syntax. 47 Unless a particular filler is
contextually assigned to the unrealized argument role, the utterance

47 In the situation-theoretic framework I will sketch in Chapter 8, an

alternative account of DNI is available. The missing argument can be

construed as belonging to the 'circumstance of evaluation' rather than to

127
remains semantically incomplete. Thus 'I heard' or 'I noticed' cannot be
interpreted as meaning that the speaker has heard or noticed something
or other. This is what it means to say that, in contrast to 'eat', 'hear' is not
subject to INI, but to DNI. ‘John eats’ is true iff there is something that John
eats2. The fact that the relevant thing does not have to be specified
supports the conclusion that intransitive ‘eat’ denotes the property of
eating1. In contrast, ‘x noticed’ is definitely not true iff there is something
y which x noticed. Rather, it is true, for some contextually specified thing
y, iff x noticed that thing. The relevant y has to be specified for the
utterance to be truth-evaluable.48 That is an instance of saturation, even
though the constituent to be contextually provided is 'unarticulated' in
surface syntax.49
To sum up, it is a conventional property of the English verb ‘eat’ that

the 'articulated content'. Thus, instead of maintaining that the objectless

verb still denotes the original two-place relation, we can say that the

objectless DNI-verb is an indexical one-place predicate : the property it

denotes is relativized to, and varies with, the 'circumstantial' constituent

that is contextually provided. Thus in a context c 1 intransitive 'notice' will

denote the property of noticing the accident, while in another context c2 it

will denote the property of noticing the man. See my 'Unarticulated

constituents', p. 312, footnote 11 for an analogous suggestion concerning

another type of implicit argument.

48 On the contrast between the two sorts of verbs, see Barbara Partee,

'Compositionality', in F. Landman and F. Veltman (eds.), Varieties of Formal

Semantics, Foris 1984, 281-311, p. 299, Fillmore 1986, Prandi 19xx,

Groefsema 1995. As Cresswell pointed out (after Partee), we observe the

same contrast between two sorts of relational nouns: «Unlike nouns like

mother and sister whose default cases are existentially quantified the

default cases of enemy and of representative appear to have the second

argument supplied contextually» (Cresswell 1996: 39).

128
(i) it can be used intransitively (in contrast to ‘devour’), and (ii) when so
used it does not require saturation of the unarticulated argument-role.
That argument role is best thought of as suppressed: the two-place
relation has been replaced by a property. Likewise, it’s a conventional
property of the English verb ‘finish’ that (i) it too can be used intransitively
(in contrast to ‘complete’), but (ii) when so used it does require saturation
of the unarticulated argument-role. The latter is not suppressed, though it
remains implicit.
The fact that the suppressed argument role of intransitive 'eat' does
not require saturation does not entail that a value cannot be contextually
provided for that role. What the DNI/INI contrast suggests is only this: if a
value is provided for the suppressed argument role of 'eat', it is provided
through an optional pragmatic process of enrichment. This is different from
the mandatory provision of a contextual value for the implicit argument-
role in need of saturation, in cases of DNI such as 'finish' or 'notice'. Verbs
subject to DNI carry something like a free variable in semantic structure,
which must be assigned a definite value in context. Verbs subject to INI do
not carry such a free variable, hence the contextual provision of a value
for the suppressed argument role cannot be treated as an instance of
saturation, but as a case of enrichment.
Let me give an example. We can imagine a context in which ‘Look!
He’s eating!’ would be understood as stating not merely that the individual
denoted by ‘he’ is eating something or other, but that he is eating a
certain salient piece of food, e.g. a dangerous mushroom which has been
the focus of attention for some time. The eaten object would then be
contextually provided without being linguistically required, since the
intransitive verb only denotes the property and does not require
saturation. In such a case the contextual specification of the eaten object
would result from free enrichment. But in the above cases of DNI ('I heard',

49 Insofar as it is mandatory and provided through saturation, the

constituent is unarticulated only in a weak sense. For the distinction

between two senses of 'unarticulated', see Recanati 2002: 307-313.

129
'I finished', 'I noticed'), the contextual provision of the relevant ingredient
is a bottom-up pragmatic process — a variety of saturation — rather than
a top-down pragmatic process of free enrichment. This is established by
the fact that the provision of an object filling the second argument role is
required in every context, not just in the contexts in which it turns out to
be relevant.

7.2 Two Criteria

I have argued that, whenever a contextual ingredient of content results

from a pragmatic process of the optional variety, we can imagine contexts
in which the lack of such an ingredient would not prevent the sentence
from expressing a complete proposition. This gives us a criterion for telling
apart cases in which a contextual ingredient results from saturation and
cases in which it does not:

Optionality Criterion
Whenever a contextual ingredient of content is provided through a
pragmatic process of the optional variety, we can imagine another
possible context of utterance in which no such ingredient is provided yet
the utterance expresses a complete proposition.

Systematically used, that criterion may give surprising results. Thus it is

common to hold that the sentence 'It's raining' expresses a complete
proposition only if a location is contextually provided. If that's right, then
the contextual specification of the location of rain is an instance of
saturation, much like the contextual provision of the thing heard in 'I
heard'.50 Using the Optionality Criterion, however, I have established that

50 Thus Ken Taylor writes:

The view which I favor supposes that the verb ‘to rain’ has a

lexically specified argument place which is marked THEME and

that this argument place takes places as values. This is a way of

saying that the subatomic structure of the verb ‘to rain’ explicitly

130
the location of rain is not provided through saturation; for there are
contexts in which the sentence ‘It is raining’ expresses a complete
proposition, even though no location is contextually provided as that which
the utterance concerns (§1.3).
Applying the Criterion is not always easy, however. For it is not
always easy to tell whether or not a contextual ingredient is actually
provided. In §6.2 I gave an example of that difficulty. There are two ways
to look at contextual restrictions of the domain of quantification. First
option: we say that a contextual ingredient is provided only when the
domain of quantification is implicitly restricted. The cases in which the
quantifier ranges over the entire domain of discourse, without contextual
restriction, are treated as cases in which no pragmatic process takes place
— the sentence is simply taken at face value. The fact that there are such
'absolute' readings of quantified sentences can then be considered as
showing that the process of contextual domain restriction is not
mandatory but optional. Second option: we consider that a primary
pragmatic process of saturation must take place to assign a value to the
domain variable in logical form. Absolute readings themselves result from
a particular assignment of value to the variable, but that is not visible

marks rainings as a kind of change that places undergo. (...) Thus

though:

It is raining

is missing no syntactically mandatory sentential constituent,

nonetheless, it is semantically incomplete. The semantic

incompleteness is manifest to us as a felt inability to evaluate the

truth value of an utterance of [that sentence] in the absence of a

contextually provided location (or range of locations). This felt need

for a contextually provided location has its source, I claim, in our

tacit cognition of the syntactically unexpressed argument place of

the verb ‘to rain’. (Taylor 2001 : 53)

131
because of the nature of the assignment: the entire domain is assigned to
the variable, hence no contextual restriction intuitively occurs. A
contextual ingredient is nonetheless provided. — The Optional Criterion,
by itself, cannot help us with the choice between the two options. It's only
after the choice has been made that we can apply the Optionality
Criterion. That is, we must decide, first, whether or not a contextual
ingredient is provided when the quantified sentence is given the 'absolute'
reading. It's only if we give a negative answer that we can use the
Optionality Criterion to establish that quantifier domain restriction is not
an instance of saturation.
Are there other criteria? According to Jason Stanley, there is one that
is most helpful: the Binding Criterion. Stanley's idea is the following.
Whenever a contextual ingredient is provided as a result of saturation,
there is something like a free variable in logical form, which must be
contextually assigned a value. If a contextual ingredient does not result
from saturation, but is pragmatic through and through, there is no variable
in logical form to which that ingredient corresponds. It is 'unarticulated' in
the strongest possible sense. But that entails that no 'binding' of the
relevant constituent is possible:

Since the supposed unarticulated constituent... is not the value of

anything in the sentence uttered, there should be no reading of the
relevant linguistic constructions in which the unarticulated
constituent varies with the values introduced by operators in the
sentence uttered. Operators in a sentence only interact with
variables in the sentence that lie within their scope. But, if the
constituent is unarticulated, it is not the value of any variable in the
sentence. Thus, its interpretation cannot be controlled by operators
in the sentence. (Stanley 2000: 410-411)

For binding to occur, Stanley argues, there must be a bindable variable in

the sentence to which the operator is prefixed; but if there is such a
variable, then the contextual ingredient (when there is one) ought to be

132
seen as the value contextually assigned to that variable in the process of
saturation, rather than as an 'unarticulated constituent' corresponding to
nothing in the sentence and resulting from 'free enrichment'. This, then, is
the criterion Stanley uses:

Binding Criterion
A contextual ingredient in the interpretation of a sentence S results from
saturation if it can be ‘bound’, that is, if it can be made to vary with the
values introduced by some operator prefixed to S.

Using that criterion, Stanley attempts to show that quantifier domain

restriction is an instance of saturation. For there are cases in which the
domain of a quantifier may be bound by another quantifier, as in the
following examples:

(1) In most of John's classes, he fails exactly three Frenchmen.

(2) In every room in John's house, every bottle is in the corner
(3) Whatever John does, most of the class falls asleep

"In each of these examples", Stanley and Szabo say,

the domain of the second quantifier expression varies with the

values of the first quantifier expression. For example, the proposition
intuitively expressed by an utterance of (1) is the proposition that,
for most x such that x is a class of John's, John failed three
Frenchment in x. Thus the domain of the quantifier expression 'three
Frenchmen' varies with the value of the variable introduced by the
quantifier 'most'. (Stanley and Szabo 2000: 243)

Since there is no binding without a bindable variable, Stanley and Szabo

conclude that "in the logical form of quantified sentences, there are
variables whose values, relative to a context, are (often restricted)
quantifier domains" (258). It follows that the contextual provision of a
restricted domain of quantification is an instance of saturation.

133
 7.3. The indexicalist challenge

Sometimes the Binding Criterion and the Optionality Criterion converge

and give the same verdict. Take syncategorematic predicates such as
'small', or verbs subject to Definite Null Instantiation, such as 'notice':

(4) John is small <for a basketball player>

(5) John noticed <the mistake>

The Optionality Criterion tells us that the contextual provision of the

comparison class (for 'small') or of the thing noticed (for intransive uses of
'notice') is in both cases an instance of saturation because there is no
possible context of utterance for (4) or (5) in which no such ingredient
would be contextually provided yet the sentence would still express a
complete proposition. Using the Binding Criterion we reach the same
conclusion, for binding can occur in both cases. Stanley gives the following
example:

(6) Most species have members that are small

Here, instead of being contextually specified, the comparison class varies

with the values introduced by ‘most species’: each species in turn serves
as comparison class for smallness. Likewise, Partee (1989) points out that
we can bind the ‘object’ of ‘notice’ even though it is unarticulated in
surface syntax:

(7) Every secretary made a mistake in his final draft. The good secretary
corrected his mistake. Every other secretary did not even notice.

The two criteria converge also with respect to examples like

(8) John is home

134
No proposition is expressed unless it is determined whose home is in
question (John's home, or the speaker's home – not to mention other
possibilities). Moreover, there is no possible context of utterance for that
sentence where it means that John is at someone or other's home.
Saturation is required. The Binding Criterion confirms this verdict. We can
say

(9) Everybody went home

and mean that each of the persons in question went to his or her home.
On that interpretation the implicit argument varies with the values
introduced by the 'everybody'. According to the Binding Argument, this
shows that there is a variable in logical form, which can be bound or
contextually assigned a value, as the case may be. When a value is
contextually assigned, as in (8), this is an instance of saturation.
Life is not as simple as that, however. There are also cases in which
the two criteria yield conflicting results. By the Optionality Criterion, as we
have seen, the location of rain which is implicit in 'It's raining' is provided
through an optional process of free enrichment; yet, Stanley points out, it
can be bound:

(10) Everywhere I go, it rains

In such a statement the location of rain is undertood as varying with the

places introduced by the quantifier ‘Everywhere I go’. The statement
means that:

For every location l such that I go to l, it rains in l (when I am there)

For such binding to occur, Stanley says, there must be a free variable l in
the sentence ‘it rains’. That variable can either be bound (as in (10)), or be
contextually given a value. Whatever location may be contextually
provided for the rain is therefore not a genuine 'optional' ingredient, but
the contextual value of a free variable in logical form. The contextual

135
provision of a specific location therefore counts as an instance of
saturation, contrary to the verdict delivered by the Optionality Criterion
Similarly, the temporal location of the breakfast event in Sperber
and Wilson's example ('I've had breakfast') seems to result from
enrichment according to the Optionality Criterion; yet it can be bound by a
quantifier.51 We can say:

(11) No luck. Each time you offer me lunch, I’ve had a very large breakfast.

The temporal location of the breakfast event now systematically varies

with the temporal values introduced by ‘each time you offer me lunch’. It
follows that there is a variable in logical form. But if there is, then the
alleged instance of enrichment is an instance of saturation after all. So the
argument goes.
Unsurprisingly, Stanley uses the Binding Criterion in arguing for
Indexicalism. He appeals to data of the sort originally collected by Barbara
Partee (1989) to show that, whenever an 'optional' ingredient has been
postulated to account for the intuitive meaning of an utterance, one can
intuitively ‘bind’ it, i.e. make it vary according to the values introduced by
some operator. It follows that "we have been given no reason to abandon
the thesis that the only truth-conditional role of context is the resolution of
indexicality, broadly construed" (Stanley 2000 : 401). Stanley therefore
maintains that "all truth-conditional context-dependence results from
fixing the values of contextually sensitive elements in the real structure of
natural language sentences" (Stanley 2000: 392). There are no 'optional'
contextual ingredients in truth-conditional content.

7.4. Is the Binding Criterion reliable?

For the 'breakfast' example, we could perhaps accept Stanley's conclusion.

For that example is very similar to cases of quantifier domain restriction

51 Stanley does not actually discuss that example, but his general

argument applies nonetheless.

136
(§6.2); and for such cases it is not easy to apply the Optionality Criterion in
a non-question-begging manner. When the Optionality Criterion is not
reliable, it seems reasonable to appeal to the other criterion. Still, the
conflict between the criteria (in cases such as 'It is raining') creates an
obvious problem, and should make us cautious; especially in view of the
use which Stanley makes of the Binding Criterion. For he uses it to deny
the very existence of optional ingredients in truth-conditional content.
Before accepting such an extreme conclusion, we must assure ourselves
that the Binding Criterion is reliable. Is it really?
I think it is not, for it works too well. It obliges us to treat as provided
through saturation not only contextual elements which can plausibly be
regarded as values of variables in logical form, as well as elements for
which at least the question arises, but also many contextual elements for
which that sort of treatment is simply out of question. This is a serious
weakness which should lead one to doubt the reliability of the Binding
Criterion. Let me give two striking examples.
The first example, which I already mentioned in Chapter 5, is
adapted from David Rumelhart (1979: 78):

(12) The policeman stopped the car

In interpreting this utterance we make certain assumptions concerning the

way the car was stopped by the policeman. On the most natural
interpretation we assume that the policeman issued appropriate signals to
the driver, who stopped the car accordingly. But if we know, or suppose,
that the policeman was actually driving the car in the reported scene, we
will understand his stopping of the car very differently from the way we
understand it when we assume that he was regulating the traffic. Quite
different ‘manners of stopping’ are involved in the two cases. Those
implied manners of stopping are part of the way we understand the
utterance but they are additional aspects of the interpretation,
linguistically optional hence external to what is said by minimalist
standards. What is said in the minimal sense is only that the policeman
stopped the car in some way or other. The specific manner of stopping is

137
provided through ‘free enrichment’.
Here, as in the case of ‘It’s raining’, the Optionality Criterion tells us
that the contextually provided manner of stopping is provided through free
enrichment. For we have no trouble imagining a context in which no such
manner of stopping would be contextually specified. Moreover, in contrast
to the ‘rain’ case, there is a wide consensus among theorists that the
contextually provided manner of stopping in such an example is a
pragmatic embellishment of the interpretation which is of no more concern
to semantics than our tendency, as interpreters, to imagine the policeman
dressed in a certain way. Indeed I think that everybody, including Stanley,
would agree that in the policeman case the contextually provided
constituent is pragmatic through and through. It is not part of the
proposition literally expressed in the minimalist sense (what is said min). Yet
the argument from binding shows that, even in that case, the contextually
provided constituent is linguistically articulated by a variable. For we can
say things like

(13) However he did it, the policeman stopped the car

(14) In some way or other, the policeman stopped the car

meaning:

For some manner of stopping m, the policeman stopped the car in manner
m.

If we apply the Binding Criterion, we shall have to conclude that the

contextually provided manner of stopping is articulated and determined
through a bottom-up process of saturation, like the reference of indexicals.
The absurdity of this conclusion argues against the Binding Criterion.
The other example is even more striking. Remember the utterance:
‘Look! He is eating!’ We imagined a context in which a salient mushroom
was understood as being the thing eaten. That the contextually provided
constituent results from free enrichment follows from the fact that
intransitive ‘eat’, as Stanley himself accepts (2000, p. 401, fn. 14),

138
denotes the property of eating, abstracted from the relation of eating-
something by suppressing the second argument role. That is a case of
'indefinite null instantiation' in which no contextual specification of the
thing eaten is required in virtue of the semantics of the verb (in contrast to
what happens in cases of 'definite null instantiation', such as 'I noticed' or
'I heard). Still, intuitively, binding is possible. We can say:

(15) John is anorexic, but whenever his father cooks mushrooms, he eats.

On a natural interpretation, we understand that John eats the mushrooms

his father has cooked. Intuitively, a form of binding is operative here; for
the food eaten by John covaries with the food cooked by his father. Such
examples show that intuitive binding, per se, does not entail the existence
of a free variable in logical form. The Binding Criterion, on which Stanley’s
argument rests, must be rejected.
It remains to be shown what exactly is wrong with the Binding
Criterion, and with Stanley's argument in support of Indexicalism. I will
provide a detailed analysis of what I call the 'Binding fallacy' in §7.6. First,
however, we need a bit of stage-setting, which the next section will
provide.

7.5. Variadic functions

When someone says 'It's raining', there often is a tacitly understood

location, such that the utterance is true if and only if it's raining at that
location. The location can also be made explicit, by means of modifiers
such as ‘in Paris’, ‘here’, or ‘everywhere I go’. Looking at those modifiers
can help us understand the status of the contextually provided ingredient
in cases in which the location remains implicit.
The modifiers in question are syntactically optional. They make a
predicate out of a predicate. If we start with a simple predicate, say ‘rain’,
we can make a different predicate out of it by ajoining an adverb such as
‘heavily’ or a prepositional phrase such as ‘in Paris’. Thus we go from ‘It’s
raining’ to ‘It’s raining heavily’ to ‘It’s raining heavily in Paris’.

139
Semantically, I suggest that we construe the modifier as contributing a
certain sort of function which I call a variadic function. A variadic function
is a function from relations to relations, where the output relation differs
from the input relation only by its decreased or increased adicity. Adding a
predicate modifier (adverb or a prepositional phrase) to a predicate
expressing a n-ary relation Rn thus results in a complex predicate
expressing an n+1-ary relation, in which the n+1th argument is a
circumstance: a time, a location, a manner, or what not.
A relation can be represented as a set of interconnected argument-
roles, as in frame semantics. Thus the Eat relation contains two argument-
roles : the Eater and the Food. The effect of an ‘expansive’ variadic
function of the sort contributed by adverbial modifiers is to add an
argument-role. The output relation therefore contains the same argument-
roles as the input relation, plus the extra argument-role provided by the
variadic function. For example, in the statement ‘John eats in Paris’ the
prepositional phrase ‘in Paris’ contributes a variadic function which maps
the property of eating, ascribed to John in the simpler statement ‘John
eats’, onto the dyadic relation Eat_in (x, l ) between an individual and a
location. That relation is predicated of the pair <John, Paris> in the more
complex statement. Note that the prepositional phrase ‘in Paris’
contributes not only the variadic function, but also the argument (Paris)
which fills the extra argument-role.52
There are also ‘recessive’ variadic functions that decrease the
valence of the input relation by suppressing an argument-role. 53 In English,
various alternations such as the passive alternation and the intransitive
alternation (Levin 1993) can be described in such terms. The operation of

52 As McConnell-Ginet puts it, such phrases "have a dual role of

augmenting the predicate to which they attach and of providing an

argument for the augmented predicate" (McConnell-Ginet 1982: 171).

53‘Recessive’ comes from Lucien Tesniere, a pioneer in that area (Tesniere

1969). ‘Expansive’ comes from Dowty (cited in McConnell-Ginet 1982:

168), and is also reminiscent of Bach’s ‘expansion’.

140
passivation (‘John kisses Mary’  ‘Mary is kissed’) suppresses the
argument-role corresponding to the subject of the active sentence,
whereas intransitivation (‘John eats the apple’  ‘John eats’) has the effect
of suppressing the argument-role corresponding to the direct object of the
verb. Quine’s ‘Der’ operator, which I mentioned earlier, does something
similar: it decreases the adicity of the input relation R n by existentially
quantifying the nth argument. But I think recessive variadic functions do
not, by themselves, involve existentially quantifying the suppressed
argument-role. Thus if we start with ‘John spoke to Mary’, we can suppress
the argument-role of Adressee and generate ‘John spoke’. This does not
entail that John spoke to someone. Whether or not the suppression of an
argument-role, effected through a recessive variadic function, amounts to
existentially quantifying that argument-role depends upon the nature of
the argument-roles that remain unsuppressed. To fill the Speaker
argument-role one need not have an addressee. One can speak without
speaking to anyone. But to fill the Eater argument-role, arguably, there
must be something (edible) that one eats: No Eater without an Eatee. The
suppressed argument-role therefore remains in the background via the
Eater argument-role which, for metaphysical reasons, cannot be entirely
dissociated from it. That is why ‘John is eating something’ can be inferred
from ‘John is eating’. That, according to me, is a nonformal, ‘metaphysical’
inference similar to that from ‘Mary is dancing’ to ‘Mary is dancing
somewhere’.54
The variadic functions that increase the valence of the input relation

54 It is arguably for the same sort of reasons that the agent remains

lurking in the background even after the Agent argument-role has been

suppressed through passivation. The fact that ‘The ship was sunk

voluntarily/to collect the insurance’ is possible, while ‘The ship sank

voluntarily/to collect the insurance’ is not, is sometimes taken to show that

in the passive sentence the Agent argument-role has not been suppressed

but remains covertly present at some level of syntactic structure (Stanley

2002: 152-3). Chomsky argues that the argument-role is present lexically

141
through the addition of a circumstance to the set of its argument-roles can
be represented by means of an operator (or rather, a family of operators)
‘Circ’. When applied to an n-place predicate P, ‘Circ’ produces an n+1
place predicate (‘Circ P’). There will be as many Circ-operators as there are
argument-roles which can be added to the set of argument-roles of the
input relation. There will be a temporal Circ-operator, a locational Circ-
operator, etc., depending on the nature of the extra argument-role. Which
Circ-operator is at issue will be indicated by means of a subscript. For

(in the valence of the verb) even though, syntactically, it is not (Chomsky

1986: 32-35). As far as I am concerned, I favour the hypothesis that the

argument-role is metaphysically implied rather than linguistically realized.

A passive sentence such as ‘The ship was sunk’ results from passivation of

the transitive ‘sink’, as in ‘John sank the ship’. Now the ship plays two roles

in (the relations described by) sentences such as ‘The ship was sunk’ or

‘John sank the ship’ : the P-role (‘patient’ or thing acted upon), and the U-

role (undergoer of change of state). ‘John sank the ship’ means John did

something to the ship that resulted in the ship’s sinking, where the two

tokens of ‘the ship’ correspond to the P-role and the U-role respectively. In

‘The ship sank’, the ship plays only the U-role. Now nothing can fill the P-
role unless there is an agent doing something. Hence expressions like

‘voluntarily’ or ‘to collect the insurance’, which qualify the manner of

acting or the goal of an agent, can be adjoined to sentences like ‘The ship

was sunk’, since the P-role which occurs in the denoted relation is

metaphysically tied to the (linguistically unrealized) Agent role. In contrast,

the expression cannot be adjoined to sentences like ‘The ship sank’,

because no action is denoted and no agent is even ‘lurking in the

background’. (The explanation I have just offered is very tentative, of

course. There may be decisive syntactic evidence that the argument-role

has not really been suppressed through passivation. But the mere fact that

‘The ship was sunk to collect the insurance’ is good while ‘The ship sank to

142
example, the operator ‘Circlocation’ contributed by locative modifiers (such
as the prepositional phrase ‘in Paris’) will map e.g. the Eat relation to the
Eat_in relation by adding a Location argument-role:

Circlocation (Eats (x)) = Eats_in (x, l)

As I pointed out, a modifier such as ‘in Paris’ does not merely

increase the valence of the input relation by adding a new argument-role;
it also provides the extra argument needed to fill that argument-role. ‘John
eats in Paris’ should therefore be represented as follows:

Circlocation: Paris (Eats (John)) = Eats_in (John, Paris)

Like the prepositional phrase ‘in Paris’, the Circ-operator thus completed
does two things: map the Eat relation to the Eat_in relation by adding a
Location argument-role; and supply a particular value (Paris) for that role. 55

7.6 The Binding Fallacy

In ‘Everywhere I go, it rains’ a variable is bound by the quantifier

‘everywhere I go’. The sentence says that, for every place l such that I go
to l, it rains in l. Stanley concludes that, when ‘It rains’ is understood with
respect to a contextually provided location, that location is provided

collect the insurance’ is bad is not sufficient to establish such a conclusion,

contrary to what Stanley suggests.)

55 According to McConnell-Ginet, who puts forward a similar proposal,

adverbs such as ‘slowly’ do not contribute an argument filling the extra

argument-role, but they existentially quantify the new argument-role while

contributing a property of its values (McConnell-Ginet 1982). If we accept

McConnell-Ginet’s idea, ‘John eats slowly’ will be analysed as follows:

Circrate: slow (Eats (John)) = (r ) (Slow (r) & Eats_at_rate (John, r)

That is, there is a rate r which is slow, such that John eats at that rate.

143
through saturation. The sentence ‘It rains’ really is the sentence ‘It rains in
l’, where the unpronounced location variable can either be bound or be
contextually assigned a value. Fully spelled out, Stanley’s indexicalist
argument against optional ingredients runs as follows:

1. Contextualists say that in the simple statement ‘It rains’, the tacitly
understood location of rain is unarticulated and results from an optional
pragmatic process of free enrichment.
2. In ‘Everywhere I go it rains’, binding occurs: the location of rain varies
with the values introduced by the quantifier ‘everywhere I go’.
3. There is no binding without a bindable variable.
4. Therefore, ‘It rains’ involves a variable for the location of rain.
5. It follows that the contextualist is mistaken: in the simple statement ‘It
rains’, the location of rain is articulated. It is the (contextually assigned)
value of a free variable in logical form, which variable can also be bound
(as in the complex sentence ‘Everywhere I go, it rains’).

The argument is fallacious because of a crucial ambiguity at stage 4.

When it is said that ‘It rains’ involves a variable (because binding occurs),
which sentence ‘It rains’ is at issue? One may well accept that in the
complex sentence ‘Everywhere I go it rains’, the (open) sentence on which
the quantifier operates involves a location variable which the quantifier
binds: ‘For every place l such that I go to l, it rains in l’. That indeed
follows from step 3. But in order to reach the conclusion at step 5, we need
something stronger: 4 must be understood as claiming that the location
variable is also involved when the sentence ‘It rains’ is uttered in isolation.
Stanley’s argument therefore relies upon an unstated premiss, namely the
following:

(SUP) In ‘Everywhere I go it rains’, the sentence on which the quantifier

‘everywhere I go’ operates is the very sentence ‘It rains’ which can also be
uttered in isolation (and whose usual interpretation involves a contextually
provided ingredient).

If we accept (SUP) it follows that the variable which is bound in the

complex sentence has got to be present also, unbound, in the simple

144
sentence ‘It rains’. Whoever accepts the analysis of adverbial modification
in terms of variadic functions must reject (SUP), however.
According to the variadic analysis, the phrase ‘everywhere I go’ does
not merely contribute what binds the variable, it also contributes the
variable itself, i.e. the extra argument-role for a location. The phrase
‘everywhere I go’ has a dual role exactly like that of any prepositional
phrase. Consider ‘in Paris’. In ‘In Paris it rains’, the prepositional phrase ‘in
Paris’ contributes both (a) a variadic function which adds an extra
argument-role to the set of argument-roles of the input predicate ‘rain’,
and (b) an argument which fills the role. This duality is quite transparent
since the prepositional phrase consists of two items: a preposition which
determines the type of the extra argument-role, and a name which
specifies what fills the role. When the prepositional phrase is an
‘intransitive preposition’ like ‘here’, it is less obvious that it plays two
semantic roles, but it does so nonetheless. In ‘It rains here’, the locative
adverb ‘here’ contributes a variadic function which increases the valence
of the expressed relation, and it also contributes a specific location which
fills the extra argument-role. We find the same duality when the phrase is
quantificational instead of being singular. In ‘Everywhere I go, it rains’, the
phrase ‘everywhere I go’ contributes both the expansive (adicity-
increasing) variadic function and the operator which binds the extra
argument-role. From the point of view of the variadic analysis, therefore,
the proper representation of ‘Everywhere I go it rains’ is:

[For every place l such that I go to l ] (in l (it rains))

What the quantifier operates on here is the subformula ‘in l (it rains)’,
whose free variable it binds. In that subformula we do find a variable for a
location. The simple sentence ‘It rains’ does not correspond to that
subformula, however, but to the sub-subformula ‘it rains’, which does not
contain a free variable for a location. Stanley’s argument goes through
only if we conflate two different things: the open sentence on which the
quantifier operates, and the simple sentence ‘It rains’ to which the phrase
‘everywhere I go’ has been adjoined. On the variadic analysis, they are

145
clearly distinguished.

7.7. Conclusion: the failure of Indexicalism

Stanley uses the Binding Criterion in attempting to establish that, in

examples like ‘It rains’, the contextually provided location is not a optional
ingredient of truth-conditional content, as the contextualists claim, but
results from contextually assigning a value to a free variable. The same
argument can be used in indefinitely many cases to show that what was
thought to be an optional ingredient of truth-conditional content is not
really. Indeed, the Binding argument is the major weapon Stanley uses in
his indexicalist assault on Contextualism. It is this argument which Stanley
and Szabo use to show that quantifier domain restriction is an instance of
saturation.
In this chapter we have seen that Stanley's attempt fails. It fails
because the Binding Argument rests on a fallacy. The construction ‘rain +
locative prepositional phrase’ expresses a relation in which there is indeed
an empty argument-slot for a location. When the prepositional phrase is
quantified, as in ‘Everywhere I go it rains’, the variable representing that
empty slot is indeed bound. But in the simple construction, without
prepositional phrase, there is neither an empty argument-slot for a
location nor a free variable. That is so because the quantified prepositional
phrase (QPP) does more than bind the variable; it also contributes the
variadic function ‘in l’ which maps the relation ‘rain’ to ‘rain_in_l’. If we
disregard the prepositional phrase and abstract from its contribution, we
suppress the variadic function and the free variable that goes with it.
I conclude that 'It rains' no more contains a free variable for a
location than 'he is eating' (with intransitive 'eat') contains a free variable
for the thing eaten. When, in context, an implicit location is contextually
supplied, or when some specific piece of food (e.g. the dangerous
mushroom) is tacitly understood as being the thing eaten, the contextual
ingredient thus provided is not assigned to a free variable; it does not
come to fill a pre-existing gap in semantic structure, but enriches the

146
interpretation of the utterance, in a typical top-down manner. This is
similar to the sort of enrichment that takes place in (12) (Rumelhart's
policeman example).
With quantifier domain restriction also the Binding Argument fails. As
Stanley and Szabo themselves mention,

one might respond [to the argument] by conceding that in sentences

such as (1)-(3) there is a quantifier domain variable that is bound by
the initial quantifier, but deny that in an 'unembedded' sentence
such as

(16) John failed exactly three students

there is a quantifier domain variable present. According to this

response, a variable is associated with quantifier expressions only in
the special case of bound readings such as (1)-(3). (Stanley and
Szabo 2000: 244)

This exactly parallels the response I made to the Binding Argument as

applied to the 'rain' example. Now Stanley and Szabo think they can
dispose of this response, in the quantifier domain case. They write:

Consideration of facts from ellipsis serves to dispose of this

response. Consider the discourse:

(17) John has failed exactly three Frenchmen. In fact, in most

classes John has taught, he has.

The natural reading of the second sentence in (17) is that in most

classes x such that John has taught x, he has failed exactly three
Frenchmen in x. However, if there is no quantifier domain variable
present in the initial sentence in (17), then there is no way of
deriving the natural reading of the second sentence.

147
 The second sentence of (17) is a standard case of syntactic
ellipsis (verb phrase ellipsis). According to standard theories of such
ellipsis, the predicate 'failed exactly three Frenchment' in the first
sentence is copied or reconstructed in the final syntactic structure of
the second sentence. If there is no quantifier domain variable
available for binding in the predicate 'failed exactly three
Frenchmen' in the first sentence of (17), then the result of copying or
reconstructing it in the logical form of the second sentence will also
not contain a bindable variable, in which case there will be no way to
derive its natural reading... Therefore, on the assumption that
standard theories of syntactic ellipsis are correct, there is a bindable
variable for quantifier domains present even in sentences such as
(16). (Stanley and Szabo 2000: 244-5.)

In this passage, however, Stanley and Szabo commit the Binding Fallacy
once again. Let us see why.
I suggested that the quantified prepositional phrase ‘everywhere I
go’ in ‘Everywhere I go, it rains’ has a dual role. It does not merely
contribute what binds the variable, it also contributes the variable itself,
i.e. the extra argument-role for a location. From this point of view the
proper representation of ‘Everywhere I go it rains’ is:

[For every place l such that I go to l ] (in l (it rains))

What the quantifier operates on here is the subformula ‘in l (it rains)’,
whose free variable it binds. The simple sentence ‘It rains’ does not
correspond to that subformula, but to the sub-subformula ‘it rains’, which
does not contain a free variable for a location.
The same analysis can be applied to the case of quantifier domain
restriction. On this analysis the second sentence of (17) will be
represented as:

[For most classes x such that John has taught x] (in x (John has failed
exactly three Frenchmen))

148
What the quantifier operates on here is the subformula 'in x (John has
failed exactly three Frenchmen)', whose free variable it binds. The first
sentence of (17) does not correspond to that subformula, however, but to
the sub-subformula occurring within the rightmost parentheses. In other
words, the quantified prepositional phrase 'In most classes he has taught'
contributes not only the quantifier 'for most classes x such that John has
taught x', it also contributes the variadic function 'in x' with its free
variable. All the material that is underlined in the above formula, including
the domain variable, corresponds to what the QPP contributes. So it is a
fallacy to conclude, as Stanley and Szabo do, that "if there is no quantifier
domain variable present in the initial sentence in (17), then there is no
way of deriving the natural reading of the second sentence." There is an
easy way of deriving that reading: one has merely to assume that the QPP
contributes the variable as well as the quantifier which binds it. On this
analysis, the result of copying or reconstructing the first sentence of (17)
in the logical form of the second sentence will not contain a bindable
variable, but that's fine since the bindable variable is contributed
independently by the QPP.
Of course, it does not follow that quantifier domain restriction is not
an instance of saturation. The only thing that follows is that Stanley's and
Szabo's argument in favour of an indexicalist treatment in terms of
saturation is fallacious and must be disregarded. Contrary to what they
claim, they have not "demonstrated the existence of quantifier domain
variables" (Stanley and Szabo 2000: 245). It remains an open question
whether quantifier domain restriction is best treated as an instance of
enrichment or as an instance of saturation. The Optionality Criterion does
not settle the issue because, as we have seen, it gives no clear result in
that sort of case. In other cases, however, the Optionality Criterion works
reasonably well; well enough for us to conclude (pace Stanley) that there
are cases in which some ingredient of truth-conditional content results
from a pragmatic process of the optional variety.

149
 Chapter 8

Circumstances of evaluation

In this book we have been concerned with the contribution context makes
to truth-conditional content. How essential is that contribution, and how
much controlled by linguistic conventions? Those are two of the main
questions we have dealt with, in an attempt to revive the debate between
Literalism and Contextualism. In this chapter we will consider a
fundamental dimension under which context contributes to truth-
conditional content: the so-called 'circumstance of evaluation'.
Truth-evaluation (or semantic evaluation more generally) requires
not merely a content to evaluate, but also a 'circumstance' against which
to evaluate that content. As Austin once put it, "it takes two to make a
truth". The circumstance of evaluation is not an aspect of the content to
be evaluated, but an entity with respect to which that content is
evaluated. Still, according to the theory of situations to be introduced in
this chapter, the circumstance of evaluation is an aspect of content in a
broader sense of 'content'. And that aspect of content is irreducibly
contextual.

8.1 Modality

The notion of circumstance of evaluation is familiar from modal logic. In

modal logic, propositions are evaluated relative to 'possible worlds'. The
possible worlds are necessary to truth-evaluation, but they are not
themselves represented in the propositions that we evaluate. Thus 'I am
French' is true, with respect to a world w, iff I am French in w; but the
sentence 'I am French' only talks about me and the property of being
French. The world of evaluation is not a constituent of the content to be
evaluated.
One can bring the world into the content by making the statement

150
more complex. The complex sentence 'Possibly, I am French' tells us that
in some possible world I am French. The modal statement I make by
uttering that sentence is about possible worlds, not merely about me and
the property of being French. In hybrid logic (a variety of modal logic), one
can even make statements 'referring' to specific possible worlds: 'In w, I
am French'.56 But the worlds that are thus introduced into the content of
the complex statement (via modal operators such as 'Possibly' or 'in w')
are used in evaluating the simple statement 'I am French', i.e. the
statement that is embedded within the modal statement. The modal
statement itself is evaluated with respect to possible worlds, and it shares
with the simple statement the property that the world with respect to
which it is evaluated is not itself represented in the statement under
evaluation.57
To appreciate the unarticulated character of the circumstance of
evaluation in the modal framework, it is worth looking at what happens
when we (standardly) translate a modal statement into first-order logic, by
explicitly quantifying over possible worlds. Thus translated 'Necessarily p'
becomes 'w p(w)', 'Possibly p' becomes $w p(w), etc. All complete
sentences are transformed into predicates (of worlds). A simple categoric
statement such as 'Rain is wet' will be represented as 'p(w)', where 'p' is
the proposition that rain is wet transformed into a predicate of worlds, and
'w' is a free variable to which the actual world is contextually assigned as
default value.
The big difference between the modal statement and its standard
extensional translation is that, in the extensional framework, the
circumstance of evaluation (the world) becomes a constituent of content.
The contrast between content and circumstance is lost. This is too bad, for
that contrast makes a lot of sense. To evaluate a sentence, we determine

56 See P. Blackburn, 'Hybrid Logic Manifesto',

57 The possible worlds which the modal statement talks about are

themselves possible with respect to the world relative to which the modal

statement happens to be evaluated.

151
whether the state of affairs it describes obtains in some 'reality' which
serves as circumstance of evaluation. But that reality — the actual world,
say — is not itself, or at least doesn't have to be, among the constituents
of the state of affairs in question, i.e. among the entities that are talked
about and articulated in the content of the proposition. The world comes
into the picture for purposes of evaluation, but the thoughts that are
evaluated need not be metaphysically elaborated thoughts about the
world. Indeed the users of the language need not even have the ability to
entertain such thoughts. Only the theorist needs to be able to talk about
the world of evaluation, in her metalanguage. The thoughts that are
evaluated 'concern' the world, but they need not be 'about' it in the sense
in which they are about the entities which they represent. (See Perry 1986
for the distinction between 'concerning' and 'being about'.)
Let us consider a simple language without modal operators or other
means of talking about worlds; let us go further and assume that the users
of the language don't possess the reflective abilities necessary for thinking
about modal issues. They entertain only nonmodal thoughts such as 'Rain
is wet'. The possible-worlds semanticist who studies their language will still
need to think and talk about the possible worlds relative to which the
sentences of the language are evaluated; but, contrary to what the
standard extensional translation suggests, mention of the possible worlds
in question will be confined to the theorist's metalanguage.
Now suppose the users of the object-language become sophisticated
and start thinking about metaphysical issues. Suppose they come to talk
and think about what is actually the case as opposed to what might be the
case. Such modal talk can be formally represented in two ways, as we
have seen: by using sentence operators, or by explicitly quantifying world
variables in the object-language. If we use the modal framework and
introduce modal operators such as 'actually' or 'possibly', nothing will be
changed for the fragment of the language that does not involve those
operators. The sentence 'Rain is wet' will still be a simple, modally
innocent sentence. The language will simply have been enriched by the
introduction of new resources enabling us to construct more complex

152
sentences. But if we use the standard extensional framework and
represent modal sentences ('It might be that...', 'Actually...') by means of
explicit quantification over possible worlds, as suggested above, then,
unless special precaution is taken to avoid that consequence, a change of
language takes place, not merely an enrichment. In the new language, all
sentences (including simple sentences) now contain a hidden argument-
place for a world. Modal innocence is lost forever.
I think this move is (almost) as damaging as the previous one — the
ascription of thought and talk about possible worlds to modally innocent
subjects. Even if the users of the language are sophisticated enough and
can think about modal issues, it is misleading to suggest that they always
think and talk about such issues even when they entertain simple
thoughts or utter simple sentences such as 'Rain is wet'. By forcing us to
construe e.g. the assertion that rain is wet as involving a covert argument-
place which the actual world fills, the extensional translation blurs the
cognitively important distinction between the simple, modally innocent
assertion 'Rain is wet' and the modal assertion 'Actually, rain is wet'. To
maintain that distinction, we have to see modal sentences as constructed
from simple sentences by the application of modal operators to them. In
this way we can analyse the ability to use and understand modal
sentences as resting on two distinct abilities: the ability to use and
understand simple sentences; and the ability to imagine other possible
worlds and to contrast the actual world with them. The first ability is
independent from the second: we can use and understand simple
sentences (e.g. 'Rain is wet') even if we lack the ability to think thoughts
about the actual world (in Perry's sense of 'about').
The important thing, I said, is to see modal sentences as constructed
from simple sentences by means of operators. Now this is something we
can do even if we want to represent modal talk extensionally. The
apparatus of variadic functions enables us to do just that. In §7.6 I
analysed 'Everywhere I go it rains' as resulting from the application of a
locative variadic operator to the sentence 'It rains'. That operator does two
things. First, it modifies the adicity of the predicate in the sentence it

153
applies to: it adds an extra argument-place for a location, which can be
represented by a free variable. Second, it introduces a restricted quantifier
which binds that variable. The operator can be paraphrased as 'for every
location l such that I go to l, in l it is the case that'. 'Necessarily it rains'
can be represented in the same hybrid way, by applying to the sentence 'It
rains' a sentence operator which can be rendered as: 'for every world w, in
w it is the case that'. Since the variable 'w' is introduced by the variadic
operator, we don't have to treat the emergence of modalities as a radical
change in the language, but simply as an enrichment of it; an enrichment
which does not affect the simple (nonmodal) sentences, hence preserves
modal Innocence.

8.2. Time and tense

The difference we have found between two ways of representing modality

can be found also between two ways of representing tense, one which
preserves temporal innocence in simple sentences and one which does
not.
In tense logic, tense is represented by means of sentence operators. 58
Alternatively, tenses can be represented by adding extra argument-places
for times.59 If we choose the latter course, it is no longer possible to
consider adjectives such as 'warm' or 'yellow' as denoting properties; they
have to be considered as denoting relations — relations between the

58 Barbara Partee (1973) says that examples like 'I did not turn off the

stove' (in which reference is made to a specific time) speak against a

treatment in terms of operators, because modal operators can't capture

the referential nature of (some uses of) tenses. But the

referential/quantificational issue is orthogonal to the question, whether or

not we should use operators. Even if standard modal operators are

quantificational rather than referential, nothing prevents the introduction

of 'referential' operators in the modal framework. See Prior 19xx, Cresswell

19xx, Blackburn 19xx.

154
objects which have the alleged properties and the times at which they
have them. As Michael Dummett has pointed out, this relational approach
significantly departs from our habitual way of thinking:

We think of adjectives such as "warm", "smooth", "slender" and so on

as denoting properties; properties that a thing may have at one time,
and not at another, but nevertheless properties rather than relations
between objects and times. And this goes with the way in which we
come to understand such adjectives. To know what it is for someone
to be my nephew, I have first to learn what it is for anyone to be the
nephew of any given person. But we do not begin by learning in what
relation an object must stand to an arbitrary time for it to be warm or
wet at that time, and then, having learned what time is referred to by
the adverb "now", derive from this a grasp of what it is for it to be
warm now. Rather, we first learn what it is for something to be warm,
wet, smooth or slender, that is to say, for the predicate "is warm
(wet, smooth, slender)" to be applicable to it, where the verb "is" is in
the true present tense. From this we advance to an understanding of
what is meant by saying of an object that it was or will be warm, etc.,
at some other time. The advance is made by our acquiring a general
grasp of the past and future tenses. That is to say, to understand
"was warm" or "will be warm", we apply to our prior understanding of
what is meant by saying that something is warm our general
comprehension of what it is to speak of how things were or will be at
another time. In so doing, we are in effect treating the tenses (and
other indications of time) as operators applied to sentences in the
present tense of which we have previously acquired an
understanding, just as the tense-logical semantics treats them. We

59 There is a third option: tenses can be represented as temporal

predicates of events. If we choose Davidson's analysis of adverbial

modification instead of the variadic analysis, this is a natural move to

make.

155
 could not learn the language in any other way. (Existence, Possibility,
and Time, ms pp. 16-17)

Dummett's complaint about the relational treatment of tenses parallels my

complaint about the extensional rendering of modal talk. The relational
treatment threatens temporal innocence, just as overt quantification over
world variables (without variadic functions) threatens modal innocence.
In the temporal case there is a possible objection, due to the fact
that tense is (to put it crudely) obligatory in English. Since it is, one might
argue that there is no such thing as nontemporal talk, hence no such thing
as temporal innocence. According to that line of argument, there is a
contrast between time and modality: there are simple, nonmodal
sentences, whose characteristics must admittedly be preserved and
captured; but there are no simple, nontemporal sentences in our
language.
From the tense-logical point of view, that objection is misguided. The
present tense is not a tense like the past or the future. It is more primitive
and, in a sense, temporally neutral. Someone can think 'It is hot in here'
even if she has no notion of time whatsoever, hence no mastery of the
past and the future. If this is right, mastery of genuine temporal talk rests
on two distinct abilities: the ability to use and understand simple
sentences (i.e. sentences in the present) and the ability to think about
times and to constrast the past and the future with the present. As in the
case of modality, the first ability is independent from the second.
It is true that, when we say or think 'It is hot in here', we talk (or
think) about what is presently the case; we characterize the situation at
the time of utterance. Yet this is not part of what the sentence itself
expresses. The content of the sentence, from the tense-logical point of
view, is a function from times to truth-values. When the sentence is
uttered, the function is applied to the time of utterance. That is so whether
the sentence is in the present or any other tense. Even if I say 'It has been
hot' or 'It will be hot', I characterize the time of utterance (and, in relation
to it, some earlier or later time). The time of utterance, which the sentence

156
is used to characterise, is the time with respect to which we evaluate the
sentence. The best thing I can do here it to quote Prior, the founder of
tense logic:

If tenses are formed by attaching prefixes like 'It has been the case
that' to the present tense, or to a complex with a present tense
'kernel', it is not always true to say that what is in the present tense
is understood as a characterisation of the time of utterance; rather,
it characterises whatever time we are taken to by the series of
prefixes. The presentness of an event, we may say, is simply the
occurrence of the event, and that is simply the event itself. But
every complete tensed sentence characterises the time of utterance
in some way or other, and other times only through their relation to
that one. (A. Prior, Worlds, Times and Selves, p. 30)

To sum up, the time of utterance is not represented, it does not feature in
the content of tensed sentences; it only comes into the picture as the
circumstance with respect to which the content of a tensed sentence is
evaluated.
Another objection, voiced by Evans, concerns the fact that a tensed
sentence like 'It is hot', 'It has been hot' or 'It will be hot' is not evaluable
as true or false, unless we are given a particular time. In the absence of a
time specification, the sentence is only 'true-at' certain times and 'false-at'
others. Such a sentence, therefore, is semantically incomplete by Frege's
lights:

A thought is not true at one time and false at another, but it is either
true or false, tertium non datur. The false appearance that a thought
can be true at one time and false at another arises from an
incomplete expression. A complete proposition or expression of a
thought must also contain a time datum. (quoted in Evans, Coll.
Papers, p. 350)

157
As Evans points out, the problem of semantic incompleteness does not
arise in the modal case. Even if a thought is said to be 'true at' one world
and 'false at' another, as in modal logic, this does not prevent it from
being true (or false) tout court. It is true tout court iff it is true-at the
actual world. But the 'thought' that it is hot cannot be evaluated as true or
false tout court. In the absence of a contextually supplied time it can only
be ascribed relative, 'truth-at'-conditions. Only a particular, dated
utterance of such a sentence can be endowed with genuine truth-
conditions. What this shows is that the time of utterance is part of the
(complete) content of the utterance;60 hence it cannot be expelled out of
the content and treated like the world of evaluation. So the objection goes.
According to Dummett, Evans's objection to Prior is based on a
misunderstanding. Prior was concerned only with sentence-types and their
contents. The content of a sentence-type is a function from times to truth-
values, hence a sentence-type has only relative truth-conditions: it is true
at some times and false at other times. To introduce a notion of absolute
truth, one thing we can say (though not, according to Dummett, what Prior
himself would say)61 is that, when a sentence is uttered, the function which
is its content is applied to some contextually provided time (typically, the
time of utterance). The time in question serves as circumstance of
evaluation for the utterance: the utterance is true tout court iff the
sentence is 'true-at' the contextually provided time. As Dummett points
out,

The variable truth-value and the absolute truth-value attach to

60 Or, in a Fregean framework, part of the expression of such a content.

61 "The simplest way to introduce a notion of absolute truth", Dummett

writes in the same manuscript, "is to follow the analogy with possible

words semantics and stipulate a type sentence to be true simpliciter just

in case it is true-now. Tense-logic, in the hands of its inventor, could be

regarded, without violation of its principles, as a semantics exclusively of

statements uttered at one particular time".

158
 different things; it is the type sentence that is true at one time, false
at another, but the utterance that is true or false simpliciter
(Dummett, p. 44)

Since there are two distinct levels, corresponding to the sentence-

type and the utterance, there is no harm in taking the utterance to
possess a 'content' also (contentu), distinct from that of the sentence
(contents). For example, we can treat the utterance as expressing a
structured proposition consisting of (i) the contextually provided time as
subject, and (ii) the content of the sentence-type, predicated of that time.
But if we do so, we must acknowledge the unarticulated nature of the
'subject' in the contentu of tensed utterances. As Prior says, "tensed
propositions are understood as directly or indirectly characterising the
unmentioned time of utterance" (Worlds, times and selves, p. 30). Hence
there is a trade-off: if we want to restrict ourselves to what is linguistically
articulated, we must focus on the content s, which is 'semantically
incomplete' by Frege's lights — it corresponds to the content of a predicate
rather than to that of a complete sentence in a logically perfect language.
If, following Frege, we want to focus on the complete content of the
utterance, that which makes it truth-evaluable in absolute terms, we must
acknowledge the role played in that content (content u) by unarticulated
constituents corresponding to the circumstances in which the content s is
evaluated.

8.3. Situations

Let us take stock. For purposes of semantic evaluation we need a

circumstance as well as a content. Even Frege, who was unconcerned by
modalities and thought of the actual world as the only world there is, was
aware of that fact. He took fictional sentences to be unevaluable, for the
following reason: since the author of a fictional statement does not
attempt to characterize the actual world, we are given a content without
any circumstance of evaluation for it. The obvious conclusion to draw from

159
Frege's remarks on fiction is that, to get a truth-value, a content is not
sufficient; we need to connect that content with the actual world, via the
assertive force of the utterance, in virtue of which the content is presented
as characterizing that world. Frege was aware not only that we need a
circumstance in addition to a content, but also that the circumstance is
not, and cannot be, an aspect of the content articulated in the sentence. If
a sentence lacks the force of a serious assertion, because the speaker
does not attempt to characterize the actual world but is engaged in a
different enterprise (e.g. poetry), making the content of the sentence more
complex by means of operators such as 'it is true that' will not change the
situation. Whether or not an utterance is serious and characterizes the
actual world is a pragmatic matter — a matter of 'force', not a matter of
content (in the narrow sense of 'content').
Once it is admitted that we need a circumstance over and above the
content to be evaluated, we can part from Frege and, following Prior,
tolerate contents that are not 'semantically complete' in Frege's sense, i.e.
endowed with absolute truth-conditions. We can, because the
circumstance is there which enables the content to be suitably completed.
Thus the content of tensed sentences is semantically incomplete, yet the
circumstance (the time) relative to which such a sentence is evaluated is
sufficient to complete it. It follows that we must distinguish two levels of
content. The content we evaluate with respect to the circumstance is the
contents; it may, but need not be, semantically complete by Frege's lights.
What is semantically complete in any case is the content u. It consists of
the contents and the circumstance with respect to which the content s is
evaluated.
Situation theory as I understand it62 follows those ideas to their

62 By 'situation theory' here I do not mean the official doctrine expounded

in Barwise and Perry's Situations and Attitudes (1983), but a body of ideas

centered around the notion of 'Austinian proposition' (Barwise 1989,

Barwise and Etchemendy 1987, Recanati 1997b, 1999, 2000; see also

Perry 1986 where some of these ideas originate). My version is, I admit,

160
consequences. It generalizes and systematizes them, in two main
directions:

1. There is no reason why only times and worlds should be accepted as

features of the circumstance of evaluation. Why not also, for example,
locations? If I say 'It's raining', the location is unarticulated, but it is
relevant qua feature of the circumstance of evaluation: what I say (or
think) is true iff it's raining at the contextually provided location. Why not
also consider the agent of the speech act (the speaker) or the thought act
(the thinker) as (part of) the circumstance of evaluation, to handle the
cases in which the content to be evaluated is a property of agents which
the speaker or thinker self-attributes (see Lewis, "Attitudes de dicto and de
se", 1979)? Why not extend the notion also to ordinary objects? If, talking
about my car, the mechanics tells me, 'The carburettor is in good condition
but there is a problem with the front wheels', my car is a crucial feature of
the circumstance of evaluation. It is true (or false) of my car that the
carburettor is in good condition, etc. The same thing could have been said
of another car, but as things turn out it is my car which figures in the
contentu of the mechanics's utterance.
Rather than list all the features which may figure in a circumstance
of evaluation, let us follow Barwise, Perry and others and use the word
'situation' to denote any entity or complex of entities which can play that
role. Anything counts as a situation provided, for some sentence S, it
makes sense to ask whether or not what S expresses is true in it (or 'of it'
or 'at it' or 'with respect to it'). Ordinary situations — restricted portions of
the actual world — are, of course, the paradigmatic case of a situation in
this generalized sense.

2. When the content of the sentence is semantically incomplete, it is the

utterance which is the proper bearer of (absolute) truth-value. Thus tensed
sentences only have relative truth-values, they express 'relativized
propositions' (Perry 1979), and we need to shift to utterances to get

fairly idiosyncratic.

161
absolute truth-values and absolute propositions. Traditional theorists think
that with sentences that are not relevantly context-sensitive and whose
content is not semantically incomplete, there is no need to invoke a
double layer of content. The content of the sentence, insofar as it has an
absolute truth-value, is the only thing we need.
Situation theory rejects that viewpoint. In situation theory, the
content of a sentence (whatever the sentence) is a function from
situations to truth-values. Hence the relativity of truth, construed as a
property of sentences: the same sentence may be true relative to a
situation and false relative to another one. That is so even if the sentence
itself is not relevantly context-sensitive or semantically incomplete. Even
when the sentence is truth-evaluable in the absolute sense — when it is
'semantically complete' by Frege's lights — situation theory says there is a
principled distinction between the content s of the sentence and the
contentu of the utterance. In such a case, the content s will be a 'classical'
proposition (a function from possible worlds to truth-values), and the
contentu will contain a situation in addition to that proposition. What the
utterance 'says' is that the situation in question supports the proposition in
question. It follows that two distinct evaluations are possible, in such
cases. We can evaluate the sentence itself (i.e. evaluate the proposition
with respect to the actual world), or we can evaluate the utterance, that is,
evaluate the proposition with respect to the situation figuring in the
contentu.
I can't refrain from quoting my favourite example here (from Barwise
and Etchemendy 1987). Commenting upon a poker game I am watching, I
say: 'Claire has a good hand'. What I say is true, iff Claire has a good hand
in the poker game I am watching (at the time of utterance). But suppose I
made a mistake and Claire is not among the players in that game.
Suppose further that, by coincidence, she happens to be playing poker in
some other part of town and has a good hand there. Still, my utterance is
not intuitively true, because the situation it concerns (the poker game I am
watching) is not one in which Claire has a good hand. But we can say that
the sentence is true, or at least true at the time of utterance: for it says

162
that Claire has a good hand, and Claire has a good hand (somewhere). The
unarticulated constituent which distinguishes the content u from the
contents makes all the difference here, and it accounts for our intuitive
classification of the utterance as non-true.

8.4. Saturation or enrichment?

The situation-theoretic framework I have sketched can, hopefully, be used

to handle some of the examples of enrichment discussed in previous
chapters:

(1) I've had breakfast <this morning>

(2) You're not going to die <from that cut>
(3) It's raining <here>

We may construe the contentu of (2) as involving as unarticulated subject

the cut-finger situation in which the child finds himself. The utterance is
true, because that situation is not one in which the child is going to die.
(That is: it is not the case that, in that situation, there is a time t posterior
to the time of utterance t* such that the addressee dies at t. There is
bound to be a time t posterior to t* at which the addressee dies, but it is
presumably not to be found in that particular, temporally restricted
situation, corresponding to the cutting event and its immediate
consequences.) Similarly, (1) posits the past existence of a breakfast
event, and is evaluated with respect to a certain day. It is therefore true iff,
on that day, there is a time t anterior to the time of utterance t* such that
the speaker has breakfast at t.63
I keep assuming that the above examples are instances of free

63 Admittedly, that is not even a sketch of an analysis. Detailed work is

needed to substantiate the claim that (1) and (2) can be appropriately

handled in this sort of way. In particular, the relations between the various

'coordinates' of the circumstance of evaluation must be exactly specified.

163
enrichment (rather than instances of saturation), but this still is an open
question. Contrary to what one might think, the situation-theoretic
framework can accommodate both options. It all depends on whether we
think that the relevant circumstantial constituent is necessary for the
utterance to express a complete proposition, i.e. a function from possible
worlds to truth-values.
Consider (3). If we take the sentence 'It is raining' to express a
complete proposition once a time has been contextually provided, then the
provision of the unarticulated location is an instance of enrichment. The
sentence expresses the proposition that there is rain (at the relevant time)
and this is evaluated relative to a specific location. Insofar as the sentence
expresses a complete proposition (with respect to a given time), that
proposition can also be evaluated 'directly', that is, it can be evaluated
with respect to the actual world. Thus evaluated the proposition comes out
true, since there is rain (at the present time) in the actual world. So we
find the two levels of truth-conditions talked about earlier, corresponding
to the contentu and the contents. On the other hand, if we think that we
need to be contextually given a place and not merely a time in order to
get a function from possible worlds to truth-values, then we hold that the
contextual provision of a location is a matter of saturation. The content s is
now taken to be semantically incomplete even after a time has been
contextually provided: with respect to a given world-time pair, 'It's raining'
only expresses a function from places to truth-values. This is like the case
of tensed sentences, whose contents, as we have seen, are semantically
incomplete. The distinction between the two levels of content — the
contents and the contentu — still holds, whether or not the content s is
semantically complete. The provision of the relevant circumstantial
constituent is an instance of saturation when the content s is incomplete,
and it is an instance of enrichment when it is complete. The same two
options are presumably available for (1) and (2), in the situation-theoretic
framework.
The contextual restriction of the domain of quantification can also be
handled as the provision of a circumstance of evaluation, without begging

164
the question whether this is an instance of enrichment or saturation. First
option: We take a sentence such as 'Everybody is tired' to determine a
function from possible worlds to truth-values (again, once a time has been
contextually provided) — a function that yields truth with respect to all
worlds in which everybody is tired. That function can then be evaluated
with respect to a particular situation, in which case it will yield truth iff
everybody in that situation is tired. Second option: we decide that the
contextual provision of a time is not sufficient — an appropriate domain for
the quantifier, that is, a particular set of people, must also be contextually
provided for the sentence to express a complete proposition. Whether we
take the first or the second option, in both cases a particular set of people
can be contextually provided via the situation of evaluation and serve as
domain for the quantifier. But its contextual provision will be considered
as an instance of enrichment if we take the first option, and as an instance
of saturation if we take the second option.

8.5. Sub-sentential circumstances

A well-known objection to the situation-theoretic treatment of quantifier

domain restriction concerns cases in which there are more than one
quantifier, and possibly more than one contextual domain, in a given
sentence. The sentence 'Every sailor waved to every sailor' can be used to
mean that every sailor on the ship waved to every sailor on the shore
(Stanley and Williamson 1995). Since we can have only one circumstance
of evaluation for that sentence, how can we deal with this multiplicity of
contextual domains? This is also the objection raised by Scott Soames to
the situation-theoretic treatment of incomplete definite descriptions
(Soames 1986).64 As I pointed out in 'Domains of Discourse' (Recanati
1996), the answer to this difficulty consists in multiplying situations of
evaluation. This can be done by extending to sub-sentential expressions
the notion of a circumstance of evaluation. Thus, following Kuroda (1982), I

64 Soames, S. (1986), Incomplete Definite Descriptions. Notre Dame

Journal of Formal Logic 27: 349-75.

165
have argued that every predicate requires a situation of evaluation
(Recanati 1987, forthcoming). Let me briefly sketch a treatment of
quantifier domain restriction within that framework.
In a standard conception, predicates denote sets of objects, and
quantifiers denote relations between sets of objects. In 'Every F is G',
'every' denotes that relation which holds between the set of Fs and the set
of Gs just in case the former is included in the latter. Now predicates
denote sets of objects only relative to situations. It follows that, before we
can evaluate a quantificational statement, we must evaluate the
predicates with respect to situations so as to determine the sets of objects
which serve as arguments to the quantifier.
When I say that predicates denote sets of objects only relative to
situations, I intend this as a rather trivial point. Some objects are red in a
given situation, which may no longer be red in a different (say, temporally
posterior) situation. So the set of red objects is variable and depends upon
the situation at stake, even if we do not vary the domain from one
situation to the next. There may be properties which stick to their objects
in the sense that, if an object has them in a situation, it must have them in
any situation to the domain of which that object belongs. Even if there are
such sticking properties, still the predicates which correspond to them will
possibly denote different sets of objects in different situations because the
domain of objects itself can vary from one situation to another.
So predicates require situations for their evaluation. The same thing
holds for sentences, as we have seen. Whether a sentence is true or false
depends upon how things are in the relevant situation of evaluation. Now
the main predicate in that sentence — the predicate which corresponds to
the topmost verb-phrase — will always be evaluated with respect to the
situation of evaluation for the sentence in question. But the other
predicates in the sentence will possibly involve distinct situations of
evaluation ('resource situations', in the terminology of Barwise and Perry
1983).
Consider the simple sentence: 'Every student laughs'. There is
simply no possibility of a divergence between the situation with respect to

166
which the sentence is evaluated and the situation with respect to which
the main predicate, 'laughs', is evaluated. That means that, if the sentence
is evaluated with respect to some situation s, the set of laughers which
serves as second argument to the quantifier 'every' will be the set of
laughers-in-s. In contrast, the predicate which occurs as part of the noun-
phrase, 'student', can be evaluated with respect to any situation: it may be
the situation of evaluation s (in which case every student-in-s is said to be
among the laughers-in-s), it may be the situation of utterance c (in which
case every student-in-c is said to be among the laughers-in-s), or it may be
any auxiliary situation which happens to be sufficiently salient for either
linguistic or extralinguistic reasons.
In this framework, we can handle the problem of quantifier domain
restriction (and the analogous problem of incomplete definite descriptions)
via the provision of a situation of evaluation, or of a sequence of such
situations. Again, this does not beg the question whether the provision of a
situation for the evaluation of a given predicate is an instance of
enrichment or of saturation. It's an instance of saturation if, like Stanley
and Szabo, we take a situation variable (or something like that) to be
associated with every common noun in logical form. But we may also
consider that the provision of a situation of evaluation is not the
assignment of a value to a variable in logical form, but an optional process
which may or may not take place. Whenever that process does not take
place, the situation of evaluation for the predicate can only be the
situation of evaluation for the sentence. That means that if the sentence is
semantically complete and we evaluate it directly relative to the actual
world, the set of objects denoted by the predicate will be the set of objects
which satisfy the predicate in the actual world.

8.6. Conclusion

As we have seen at the beginning of this chapter, the possible worlds

relative to which, in modal logic, sentences are evaluated are not aspects
of the contents to be evaluated, but remain external to those contents.

167
The same thing holds when time is the relevant circumstance of
evaluation. Yet there is a difference between time and modality: the type
of content that is evaluated with respect to a time in tense logic is not
semantically complete in the sense in which the content which is
evaluated with respect to a possible world is. This difference was
emphasized by Evans in his attack on tense logic. But the proper response
to make, on behalf of tense logic, is the drawing of a distinction between
two levels of content. The content which a sentence articulates is one
thing; the content of an utterance of that sentence, with respect to a given
circumstance of evaluation, is another thing. Only the second, richer type
of content is a legitimate bearer of truth-value, by Fregean lights. It follows
that the truth-conditions of a tensed utterance crucially involves the
circumstance of evaluation, which is not an aspect of the content
articulated in the sentence. Hence there is a fundamental aspect of truth-
conditional content that is not articulated in the uttered sentence.
The theory of situations I briefly introduced generalizes the lessons
from tense logic. In this theory every utterance articulates a certain
content and expresses a richer type of content (an 'Austinian proposition')
consisting of the articulated content and the situation with respect to
which that content is (intended to be) evaluated. The complete, Austinian
proposition expressed by an utterance is true iff the situation of evaluation
supports the articulated content. That content is typically not a classical
proposition but a relativized proposition — a function from rich
circumstances to truth-values.
In this framework there is an 'unarticulated constituent' in the
(complete) content of every utterance, namely the circumstance of
evaluation. Admittedly, there are different types of unarticulated
constituent. 65 In many cases a circumstantial constituent counts as
'unarticulated' only in a weak sense. That is so whenever the constituent,
without being explicitly represented in the articulated content, is
nevertheless mandatory because that content is not semantically
complete by Frege's lights. So the time of evaluation (i.e. the time relative

65 See footnote 47.

168
to which an utterance is evaluated) is unarticulated because, as Prior says,
it is 'unmentioned' in the utterance, yet its contextual provision results
from a form of saturation: such a constituent must be provided in order to
get a complete proposition, i.e. a function from possible-worlds to truth-
values. In situation theory, however, the necessity to provide the relevant
circumstantial constituent is exactly like the necessity to provide a
possible world in order to determine a truth-value. In general, contents
need circumstances for their evaluation. Once this is acknowledged, the
door is open to relativized contents of all sorts, and the privilege given to
classical propositions (functions from possible worlds to truth-values) over
relativized propositions (functions from rich circumstances to truth-values)
looks somewhat arbitrary.66 Be that as it may, a distinction must be drawn
in all cases between the articulated content (content s) and the complete
content of the utterance (contentu); and the irreducibly contextual nature
of the circumstance that features in the content u must be acknowledged.
Insofar as this position generalizes the context-dependence of truth-
conditional content, it can be categorized as 'contextualist'.
Still, one might argue, the theory I have sketched posits a level of
content that is very close to the linguistic meaning of the sentence: the
contents. Insofar as it posits such a level, the theory of situations sounds
more similar to the Syncretic View than to Contextualism or even Quasi-
Contextualism. — I find this reaction natural, yet I maintain that the theory
of situations is on the side of Contextualism. I concede that there are
similarities with the Syncretic View, but they are more superficial than it
looks.
The main analogy consists in the fact that, in both cases, we find two

66 Classical propositions fulfill two functions: they are the bearers of

absolute truth-values, and they are the object of the attitudes. In the

situation theoretic framework, the Austinian proposition fulfils the first

function, and the articulated content (relativized proposition) fulfils the

second function, for which classical propositions are in any case ill-suited.

So it is unclear that we need classical propositions.

169
levels of content: the proposition expressed by the sentence, and a richer
proposition containing unarticulated constituents. 67 Taken literally, 'You're
not going to die' is true iff the addressee is not going to die. This
corresponds to the minimal proposition of the Syncretic View (what is
strictly and literally said), and to the content s of situation theory. That
content yields falsity in all possible worlds in which the addressee is not
immortal. So the utterance counts as false ('literally false') if we evaluate
the contents directly with respect to the actual world. If we take the
intended situation of evaluation into account and evaluate the complete

67 In 'Situations and the structure of content' (Recanati 1999) I

emphasized that analogy:

[In situation theory] an utterance is true iff the situation s it refers to

supports the fact  it expresses. The complete truth-conditional

content of an utterance therefore is:

s |= 

This complete content, or Austinian proposition, is distinct from the

proposition  on the right hand side of the support sign '|='. So there

are two levels of content, two propositions which every utterance

expresses: the fact  which is stated, and the Austinian proposition to

the effect that the situation of reference supports that fact.

There is much in common between the right hand side in the

Austinian proposition (I call it the 'nucleus') and minimalist content.

(...) The Austinian framework, therefore, helps us to implement the

compromise view imagined by Salmon and Bach. As in the latter we

have four levels, with the notion of 'what is said' split in two:

• sentence meaning
• what is said1: nucleus

• what is said2: Austinian proposition

• what is implicated

(Recanati 1999: xx)

170
Austinian proposition (contentu), the utterance counts as true iff the
addressee is not going to die in that particular situation (resulting from the
cut). This truth-condition corresponds to the intuitive truth-conditions of
the Syncretic View.
Despite this convergence between the two views, there are
important differences. The articulated content (right-hand-side of the
Austinian proposition) is unlike the minimal proposition of the syncretist in
two respects:

1. As we have seen, the articulated content of the situation theorist

need not be a complete proposition; it may be a relativized
proposition. In contrast, the minimal proposition must be a complete
proposition, because the 'minimalist' criterion that is used in
individuating it involves truth-evaluability. 68
2. More important in the context of the present discussion, there is
absolutely no reason why the articulated content (content s) of the
situation theorist should not be affected by primary pragmatic
processes of the optional variety.69 Typically, the situation that is
contextually provided for the evaluation of an utterance will

68 Note, however, that some syncretists use a different criterion, and can
therefore accommodate relativized propositions. Thus for Bach (who uses

the 'IQ test' as criterion), what is literally said need not be a complete

proposition, it may be a 'propositional radical'. Soames has recently come

to endorse a similar view (see his "Naming and Asserting", in Z. Szabo ed.,

Semantics and Pragmatics, Oxford University Press).

69 Hence it would be misleading to keep describing this content as 'the

content of the type sentence', to use Dummett's words. The content s is as

much a property of the utterance as the content u. The contents is the

articulated content of the utterance, while the content u is the complete

content of the utterance (including the intended circumstance of

evaluation).

171
contribute to shaping the content that is evaluated with respect to
that situation: the meaning of words will naturally be enriched and
adjusted to fit the intended situation of evaluation. So the
articulated content of the situation theorist is definitely not the
minimal proposition posited by the Syncretic View.

172
 Chapter 9

Contextualism: How far can we go?

9.1. The modulation of sense

In context the meaning of words is adjusted or 'modulated' (Cruse 1986:

50-53) so as to fit what is being talked about. Sense modulation is
essential to speech, because we use a (more or less) fixed stock of
lexemes to talk about an indefinite variety of things, situations and
experiences.70 Through the interaction between the context-independent
meanings of our words and the particulars of the situation talked about,
contextualised, modulated senses emerge, appropriate to the situation at
hand. The meaning of a word can thus be made contextually more
specific, or it may, on the contrary, be loosened and suitably extended, as
in metaphor. It may also undergo 'semantic transfer', etc.
According to many authors among those who have studied the
phenomenon (e.g. Cohen 1986, Lahav 1989, Pustejovsky 1995),
modulation is the process whereby the meaning of a given word is affected
by the meanings of other words in the same sentence. Thus the meaning
of the adjective 'light' is affected by the meaning of the noun it modifies: a
light lunch is not light in quite the same sense in which a piece of luggage
is said to be light (on this example, see Michael Pelczar, "Wittgensteinian
Semantics", Nous 2000). According to Jonathan Cohen, this is one of the
big differences between natural language and formal languages: "artificial
languages satisfy an insulationist account whereas natural languages
require an interactionist one" (1986, p.224). Cohen characterizes
insulationalist and interactionist accounts as follows:

70 "Since the universe never repeats itself exactly, everytime we speak

we metaphorize" (Bolinger, Aspects of Language, p. xx).

173
 According to the insulationist account the meaning of any one word
that occurs in a particular sentence is insulated against interference
from the meaning of any other word in the same sentence. On this
view the composition of a sentence resembles the construction of a
wall from bricks of different shapes. The result depends on the
properties of the parts and the pattern of their combination. But just
as each brick has exactly the same shape in every wall or part of a
wall to which it is moved, so too each standard sense of a word or
phrase is exactly the same in every sentence or part of a sentence in
which it occurs...
Interactionism makes the contradictory assertion: in some
sentences in some languages the meaning of a word in a sentence
may be determined in part by the word's verbal context in that
sentence... On this view the composition of a sentence is more like
the construction of a wall from sand-bags of different kinds. Though
the size, structure, texture and contents of a sand-bag restrict the
range of shapes it can take on, the actual shape it adopts in a
particular situation depends to a greater or lesser extent on the
shapes adopted by other sand-bags in the wall, and the same sand-
bag might take on a somewhat different shape in another wall or in a
different position in the same wall. By exploiting local context in this
way a language can be much more prolific of semantic variety than
insulationalism can give it credit for being. ('How is conceptual
Innovation Possible', Erkenntnis 25-2, 1986, pp. 223-4).

I agree with Cohen that the semantics of natural language is not

insulationist. As I have been at pains to emphasize, the meaning of the
whole is not constructed in a purely bottom up manner from the meanings
of the parts. The meaning of the whole is influenced by top-down,
pragmatic factors, and through the meaning of the whole the meanings of
the parts are also affected. So we need a more 'interactionist' or even
'Gestaltist' approach to compositionality. 71 But I think Cohen takes too

71 The idea of 'Gestaltist compositionality' comes from Victorri 1999.

174
narrow a view of the sort of interaction that sense modulation results from.
Let us go back to the example discussed by Michael Pelczar:

In these sentences:

'There was a light breeze from the south.'

'This is a light package.'
'The fire won't last with such light fuel.'

— 'light' expresses a content having to do with intensity, weight, and

density, respectively. And this clearly results from the fact that in the
first sentence 'light' is used to speak of a breeze, whereas in the
second it is applied to describe a package, and in the third, to some
fuel for a fire. (Pelczar 2000: 488)

According to Cohen, the sense of 'light' is modulated through its

interaction with the meaning of the head noun ('breeze', 'package',
'fuel').72 I think this is not quite right. The meaning of the head-noun is not
what matters most. It is, as Pelczar emphasizes, what the word "is used to
speak of", what it "is applied to describe", that matters. Nothing prevents
us from using 'light', in the same phrase 'light fuel', to characterize the
weight of the fuel rather than its density, if this is what is relevant to us in
the context at hand. We can even imagine a context in which 'light', as
applied to a package, would characterize the intensity of the sound it
emits.
To take another example from Pelczar's study, the word 'get' takes
on different senses — denotes different relations — depending on what fills
the second argument-place of the relation. Thus 'Smith has gotten the

72 This is what Lahav calls the 'noun-dependence' of adjectives (and

verbs), a feature that, according to him, threatens the Principle of

Compositionality. See his 'Against Compositionality: the Case of

Adjectives', Philosophical Studies 57: 261-79, 1989.

175
virus' means that he has contracted the virus, while in 'I need to get some
eggs' 'get' has what Pelczar calls the acquisitional sense. Yet, as Pelczar
points out, we can easily imagine a context in which 'Smith has gotten the
virus' means that Smith has acquired, i.e. successfully collected a sample
of, the virus. The same remarks can be made with respect to the word
'cut', whose meaning is modulated differently in 'John cut the grass' and
'John cut the cake'. As Searle points out, if I ask John to cut the grass, and
he cuts it the way one cuts a cake (or if I aks him to cut the cake, and he
cuts it the way one cuts grass), he has not complied with my request; he
has not done what I asked him to do. Still, we can imagine a context in
which in asking John to cut the grass, I would mean that he is to cut it the
way one cuts a cake:

Suppose you and I run a sod farm where we sell strips of grass turf
to people who want a lawn in a hurry... Suppose I say to you, "Cut
half an acre of grass for this customer"; I might mean not that you
should mow it, but that you should slice it into strips as you could cut
a cake or a loaf of bread. (Searle 1980: 224-5).

So it's not just the interaction of the meaning of the word with the
linguistic context that is relevant. The really important factor is the
discourse topic. The 'internal' interaction of word-meanings within the
same sentence or discourse is only a reflection of the more general
phenomenon: the interaction between word meaning and the situation the
words are used to talk about.

9.2. The semantic relevance of modulation

As the virus examples shows, modulation is not just the post-semantic

addition of a truth-conditionally irrelevant 'shade of meaning' to the core
meaning of words. Modulation is truth-conditionally relevant, because the
words' contribution to truth-conditional content (or to satisfaction-
conditions more generally) is not their pre-modulation meaning but their

176
modulated senses. Thus different truth-conditions correspond to the two
senses which 'get' may take on through modulation: the sense of 'acquire'
or that of 'contract'. In the acquisitional sense 'Smith has gotten the virus'
will be true if Smith has collected a sample of the virus, even if Smith has
not thereby contracted the virus in question. But unless he has contracted
the virus, it will be false that he has 'gotten the virus', in the other sense.
So modulation results in a truth-conditional difference. Thanks to that
difference, it is possible to say 'He has gotten the virus, without getting
the virus', consistently — provided the sense of 'get' is modulated
differently on its two occurrences.
Since modulation has an impact on truth-conditions, it cannot be
relegated to the 'post-semantic' periphery. But there is another way of
denying modulation semantic relevance: we can relegate it to the 'pre-
semantic' limbo by construing it as a variety of disambiguation. The word
'get', one might argue, has several senses, one of which must be
contextually selected on any particular use. Since disambiguation is not a
semantic but a pragmatic process, there is no reason why modulation (a
variety of disambiguation) should matter to semantics.
Like the post-semantic dismissal, this strategy fails. For there to be
the sort of disambiguation that would support the pre-semantic dismissal,
a pre-established list of discrete senses for 'get' must be somehow given,
from which the contextually relevant sense can be selected. But, as Herb
Clark and many others following him have stressed, modulated senses
result from a process of 'generation' or 'creation' rather than of selection
(Clark 1992, chapters 10 and 11; Pustejovsky 1995). The fundamental
difference between sense generation and sense selection is that
generation is productive: new senses can be generated, in a creative
manner. As a result words can take on an indefinite variety of possible
senses. (Another important difference between straigthforward ambiguity
and modulation is the non-discrete nature of the senses generated
through modulation: as several authors point out, there is a continuum of

177
possible semantic values for a polysemous word such as 'get'. 73)
Can we say at least that the phenomenon of modulation only
concerns a restricted class of words, namely those that are 'polysemous'
and possess a family of related senses? No. First, as Cohen points out,
polysemy is the norm rather than the exception ('Ambiguity in truth-
theoretic semantics', 1985, p. 131). To be sure, not all polysemous words
are as polysemous as ultrafrequent words such as 'get'. Polysemy comes
in grades, and it could be argued that modulation is associated with a high
degree of polysemy: highly polysemous words, it may be said, have an
abstract core meaning which calls for elaboration or modulation, and from
which the various senses of the polysemous expression are generated. 74
This does not entail that non-polysemous words are not subject to
modulation, however. Polysemy arguably proceeds from modulation (since
the array of senses of a polysemous expression reflects conventionalized
patterns of modulation), not the other way round. So the proper response
to make is that modulation can affect any ordinary word, whether or not it
is (highly) polysemous.
A paradigm example of modulation is the use of colour terms such
as 'red' (Travis 1981: 16; Lahav 1989: 264). The conditions of application
for 'red' depend upon (inter alia) what sort of thing the colour adjective is
applied to. As Lahav puts it, the conditions that a table has to meet in
order to be describable by 'red table' are not the same conditions that a
house or a book or a bird has to meet in order to be describable by 'red
house' or 'red book' or 'red bird' (1989 262-4). 75 Yet 'red' is not especially
polysemous; it is just an ordinary adjective, whose conditions of

73 See Fuchs (1991) on French 'pouvoir', Atkins (1993) on 'safety', and

Grandy 1990 (pp. 570-571) on 'prime'.

74 See Alfonso Caramazza and Ellen Grober, 'Polysemy and the structure

of the subjective lexicon', in C. Rameh (ed.), Semantics: Theory and

Application. Georgetown University Round Table on Language and

Linguistics, Washington DC: Georgetown University Press, 1977?, pp. 181-

207

178
application can be varied or modulated in context, like that of any ordinary
adjective. Or take the verb 'to like' which we find both in 'He likes John's
sister' and in 'He likes the food John cooks'. It's not the same sort of 'liking'
that is involved in both cases, and we can imagine that someone who likes
John's sister in one sense (say, the culinary sense — for that you have to
imagine a weird context) does not like her in the other sense. Yet 'like' is
just like any other verb. So the phenomenon of modulation is not a special
feature of a restricted class of expressions. In general, the conditions of

75 "For a bird to be red (in the normal case), it should have most of the

surface of its bodly red, though not its beak, legs, eyes, and of course its

inner organs. Furthermore, the red color should be the bird's natural color,

since we normally regard abird as being 'really' red even if it is painted

white all over. A kitchen table, on the other hand, is red even if it is only

painted red, and even if its 'natural' color underneath the paint is, say,

white. Morever, for a table to be red only its upper surface needs to be

red, but not necessarily its legs and its bottom surface. Similarly, a red

apple, as Quine pointed out, needs to be red only on the outside, but a red

hat needs to be red only in its external upper surface, a red crystal is red

both inside and outside, and a red watermelon is red only inside. For a
book to be red is for its cover but not necessarily for its inner pages to be

mostly red, while for a newspaper to be red is for all of its pages to be red.

For a house to be red is for its outside walls, but not necessarily its roof

(and windows and door) to be mostly red, while a red car must be red in its

external surface including its roof (but not its windows, wheels, bumper,

etc.). A red star only needs to appear red from the earth, a red glaze

needs to be red only after it is fired, and a red mist or a red powder are red

not simply inside or ouside. A red pen need not even have any red part

(the ink may turn red only when in contact with the paper). In short, what

counts for one type of thing to be red is not what counts for another."

(Lahav 1989: 264)

179
application for words can be varied or modulated in context. That is the
phenomenon a proper theory of language and communication must
account for.

9.3. Four approaches

Insofar as it takes as input the conventional, linguistic meaning of an

expression, and delivers as output the modulated sense that is the
expression's contribution to truth-conditional content, modulation is in the
same ballpark as indexicality. Cohen stresses the similarity between the
two phenomena:

It is an enormous convenience that the same word can often be

uttered in one or other of several different though related senses.
Instead of having to learn a very much larger number of words, each
with fixed and context-independent meanings, we can learn a
relatively small number of words with variable meanings... The
economy of resources that is thus achieved by natural language is
obviously comparable with the economies that are achieved by
variability of reference in the case of proper nouns and other
indexical expressions as an alternative to the fixed references
available with the use of co-ordinate systems, individual constants,
etc. (Cohen 1985: 132)

There is a fundamental difference between indexicality and modulation,

however. Despite its semantic relevance, modulation is a pragmatic
process in the fullest possible sense: it is a pragmatically controlled
pragmatic process, rather than a linguistically controlled pragmatic
process (like saturation). Neither enrichment, nor loosening, nor transfer,
nor any other of the mechanisms at work in modulation seems to require,
on the side of the input, a 'slot' or gap in semantic structure demanding to
be filled and triggering the search for an appropriate filler. In contrast to
saturation, which proceeds from the bottom up, modulation seems to be

180
fundamentally top-down.
In this section I will introduce four types of approach to modulation,
which may be placed on the scale from Literalism to Contextualism
(Chapter 6). Insofar as modulation is a cover term for what, in this book, I
have called 'primary pragmatic processes of the optional variety', there is
no account of modulation that is literalist, since the very phenomenon of
modulation tells against Literalism. Still, among the possible accounts of
the phenomenon, some are closer than others to the literalist pole, and
some are closer than others to the contextualist pole.
The first view to be discussed is the natural view to hold for a
syncretist or a quasi-contextualist. I call it the Strong Optionality view (SO).
It takes the primary pragmatic processes that are involved in modulation
to be optional in a strong and straightforward sense. They take place, for
contingent reasons, but they might as well not take place. In a suitable
context, the senses expressed by the words would be, simply, the senses
they possess in virtue of the rules of the language.
Consider standard examples such as:

She took out her key and opened the door

There is a big lion in the middle of the piazza

The pragmatic process that enriches the meaning of the first utterance so
as to convey both a sense of temporal order (giving to 'and' the sense of
'and then') and a notion of the intrument used in opening the door (giving
to 'opened the door' the sense of 'opened the door with the key') — that
process might also not take place. As Grice emphasized, such pragmatic
suggestions are always cancellable, explicitly or contextually (Grice 1989:
39).76 Once the pragmatic suggestion has been cancelled, what the words

76 The speaker could cancel the suggestion explicitly by saying: 'She took

out her key and opened the door – not with the key, though, and not in

that order. She took out her key while opening the door (which was

unlocked) with the other hand'.

181
contribute to truth-conditional content is their bare linguistic senses. The
same thing can be said about the second example. 'Lion' can be
understood in the representational sense: the 'lion' that is talked about
can be a representation (e.g. a statue) of a lion, not a real lion. That
reading results from a primary pragmatic process that can be variously
described.77 However we describe it, its optional character is obvious:
nothing prevents that sentence from being understood literally, as talking
about a real lion.
The three other views I want to discuss are all contextualist. They all
ascribe to modulation a form of necessity or ineliminability, to the point of
blurring the mandatory/optional contrast. Discussing them will enable us
to distinguish more or less radical varieties of Contextualism. The last view
I shall consider corresponds to the most radical form: Meaning
Eliminativism
Like SO, the Pragmatic Composition view (PC) accepts that the
literal, input sense undergoing modulation could, in a suitable context, be
the expressed sense. So it construes the pragmatic process of modulation
as optional. But it construes it as optional only with respect to the word
whose sense is modulated. If we consider not words in isolation, but the
complex expressions in which they occur, we see that the primary
pragmatic processes of modulation are not always contingent and
dispensable, but often essential. Even though the linguistic meaning of a
given word (or the semantic value that results from saturation) could be
the expressed sense, still the process of semantic composition, i.e. the
putting together of that sense with the semantic values for other

77 According to Cohen (1978, 1985), the representational sense is

generated by 'cancelling' the feature +animate from the rich lexical

meaning of 'lion'; according to Grandy (1990), it is generated by enriching

the structured character of the utterance with an unarticulated function

from objects to their representations; according to Walton (1990), the

circumstance of evaluation for the predicate 'lion' is shifted from the actual

world to the world of the pretense.

182
expressions, cannot proceed unless appropriate adjustments take place so
as to make the parts fit together within an appropriate whole. On this view
words have meanings which could go directly into the interpretation,
without modulation, but it is the composition process that forces
modulation to take place, or at least invites it: often the meanings of
individual words do not cohere by themselves, and can be fitted together
only by undergoing a process of mutual adjustment.
Remember the examples I gave in chapter 2. If we take it as
axiomatic that only sounds can be heard, then, in 'I hear the piano', either
the sense of 'hear' or that of 'the piano' must be modulated for the
sentence to make sense (§2.6).78 Or remember the Gary Cooper example,
'disarm the fricassee', or Searle's example, 'Cut the sun', from §6.3. To
make sense of such sentences we must imagine (or retrieve from memory)
a possible scenario in which the semantic values of the words fit together. 79
That imaginative exercise involves elaborating what the meaning of the
words give us — going beyond that linguistic meaning and, for example,
interpreting 'disarm' in the specific sense of: 'take the gun out of' or
'remove the gun from'.
Or consider the paradigmatic example I gave in the last section: the
adjective 'red'. There is no particular incompleteness in the linguistic
meaning of that word — it means something like 'being of the colour red'
or 'having the colour red'. Insofar as 'red' refers to a specific colour (and it
does) this, it seems, expresses a definite property — a property that could,
in principle, go into the interpretation of a sentence in which the adjective
'red' occurs. (For example: 'Imagine a red surface.') But in most cases the
following question will arise: what is it for the thing talked about to count

78 Note that there may be no fact of the matter as to which particular word

or word-meaning undergoes modulation, in examples like those of Chapter

2. This is due to the holistic character of modulation.

79 See A. Sandford and S. Garrod, 'The role of Scenario Mapping in Text

Comprehension', Discourse Processes 26, 1998, pp. 159-90; D. Rumelhart,

'Some problems with the notion of literal meanings'.

183
as having that colour? Unless that question is answered, the utterance
ascribing redness to the thing talked about (John's car, say) will not be
truth-evaluable. It is not enough to know the colour that is in question
(red) and the thing to which that colour is ascribed (John's car). To fix the
utterance's truth-conditions, we need to know something more —
something which the meanings of the words do not and cannot give us: we
need to know what it is for that thing (or for that sort of thing) to count as
being that color. What is it for a car, a bird, a house, a pen, or a pair of
shoes to count as red? To answer such questions, we need to appeal to
background assumptions and world knowledge. Linguistic competence
does not suffice: pragmatic fine-tuning is called for.
To sum up, on the view (PC) under discussion, even if the semantic
value of a word is fixed by language (and context, if saturation is
necessary), composing it with the semantic values for other words often
requires help from above. It is semantic composition which has a
fundamentally pragmatic character.80 So there is a sense in which
modulation is necessary, but that is not quite the sense in which
saturation is. With saturation there is a semantic gap and an instruction to
fill the gap — both the gap and the instruction being part of the linguistic
meaning of the expression. With modulation, there need be no gap and
there is no instruction to search for some contextual filler. The expression
means something, and that meaning could go into the interpretation — so
modulation is optional — but to determine a suitable sense for complex
expressions, we need to go beyond the meaning of individual words and
creatively enrich or otherwise adjust what we are given in virtue purely of
linguistic meaning. We must go beyond linguistic meaning, without being
linguistically instructed to do so, if we are to make sense of the utterance.
As Rumelhart writes, utterance understanding involves "searching memory
for a schema to account for the described event" (p. 85). I can't resist
quoting Rumelhart's methodological comment here:

80 See Benveniste 1958 for a clear statement of that idea. See also

Rumelhart, 'Some problems with the notion of literal meaning', p. 83.

184
 This approach is, I believe, quite different from the 'standard'
appraoch. The standard view emphasizes the 'bottom up' process of
constructing meaning from smaller component meanings.
Nonlinguistic knowledge comes into play after the set of possible
meanings has been selected. My approach suggests that
comprehension, like perception, should be likened to Hebb's
paleontologist (Hebb 1949) who uses his beliefs and knowledge
about dinosaurs in conjunction with the clues provided by the bone
fragments available to construct a full-fledged model of the original.
In this case the words spoken and the actions taken by the speaker
are likened to the clues of the paleontologist, and the dinosaur, to
the meaning conveyed through those clues. On this view the
processing is much more 'top down' in that internal hypotheses are
actively imposed on the observed utterances. (p. 85) 81

On the third type of view — the Wrong Format view (WF) —

modulation is ineliminable for deeper reasons. It's not just semantic
composition which requires adjustment and modulation of word meaning.
Individual word meanings themselves are such that they could not go
directly into the interpretation. They don't have the proper format for that.
They are either too abstract and schematic, in such a way that elaboration
or fleshing out is needed to reach a determinate content (Guillaume
1929 :132-134, Ruhl 1989); or they are too rich and must undergo
'feature-cancellation' (Cohen 1971, 1977, 1978), or some other screening
process through which some aspects will be backgrounded and others
focussed on. Note that there are versions of this view which take the

81 I am not suggesting that Rumelhart himself advocated the Pragmatic

Composition view. I think his contextualism was more radical than that. As

the above passage shows, he held that semantic composition is, to a large

extent, a top down process. But he would presumably not have accepted

the claim that the meaning of a linguistic expression could go directly into

the interpretation, without modulation.

185
meaning of a word to consist both in some abstract schema in need of
elaboration and a large store of encyclopedic representations most of
which must be screened off as irrelevant, on any particular use (Langacker
1987).
According to WF modulation is mandatory, hence it becomes harder
to distinguish it from saturation. This is the point where Indexicalism and
Contextualism meet (§6.6). An Indexicalist will argue that the schematicity
of word meaning invoked by (some versions of) WF just is the presence in
that meaning of free variables or parameters. So there is no modulation
really, only saturation. A contextualist will reply, as I did in §6.4, that the
alleged parameters do not behave like standard indexical variables. 82
The last view of interest to us, Meaning Eliminativism (ME), is a sort
of WF pushed to the extremes. According to WF, words have meanings,
but those meanings don't have the proper format for being recruited into
the interpretation of utterances; they are not determinate senses but

82 In his contextualist manifesto, 'Is Snow White?', Julius Moravcsik points

out that contextual indices of the sort used to deal with indexicals have

two important properties. First, they "simply limit or relativize truth to a

frame of application within which variations in extension and truth do not

arise". Second, "for most of the standard indices we can find individual
lexical items to which these are linked". Neither of these conditions holds

in the case of modulation. The phenomenon "is not linked to a specific sub-

vocabulary of English. Practically the whole descriptive vocabulary is

affected". Moroever, "the various practical contexts create rather than

restrict extension. It is not as if we could survey all of the possible future

extension-creating contexts that our future and possible interactions with

the world can create. Thus the intuitive idea that underlies the practice of

relativizing truth is missing in our type of case. The expanding ranges of

application and contexts cannot be described in a rigorous way, while this

is possible in the cases in which the introduction of indices is legitimate"

(p. 9-10).

186
overly rich or overly abstract 'semantic potentials' out of which
determinate senses can be constructed. ME goes farther in the same
direction. It denies that words (qua linguistic types) have 'meanings' in
anything like the traditional sense — not even abstract or knowledge-rich
meanings, as in WF. Meanings for types undergo wholesale elimination, in
favour of the senses contextually expressed by particular tokens. ME is
implicit in the writings of some of the early contextualists (Austin,
Wittgenstein, etc.), and before comparing it to WF, it is worth taking a look
at their view.

9.4. Truth-conditional unstability: from Waismann's 'open texture'

to Searle's 'background'

A good place to start is Waismann's classic paper 'Verifiability', and more

specifically the passage where he introduces the notion of 'open texture'. 83
That passage is representative of a view (and a type of example) which
Waismann shared with Austin and Wittgenstein, and which has been
revived recently by contemporary contextualists such as Travis and Searle.
According to this view, truth-conditional content is essentially unstable and
context-dependent.
What Waismann says in introducing open texture can be
paraphrased as follows. It seems that we have no problem assigning truth-
conditions to ordinary sentences. To do so we describe a state of affairs,
i.e. a type of situation, the obtaining of which is necessary and sufficient
for the sentence to be true. That seems easy enough to do. Thus we think
we know the truth-conditions of 'There is a cat next door', or 'This is gold'
(or 'This is a man'): we can specify a state of affairs s such that the
utterance is true if and only if s obtains. But this is an illusion — we can't
really. Given an utterance u and the state of affairs s which is its alleged
truth-maker, it is always possible to imagine a world in which s obtains, yet
it is not the case that u is true (with respect to that world).

83 Friedrich Waismann, 'Verifiability', in A. Flew (ed.), Logic and Language,

1st series, Oxford: Basil Blackwell, 1951, pp. 119-123.

187
 To show that, one has only to embed the state of affairs s within a
larger situation, by providing further details about an imagined world in
which s obtains. If the world in question is sufficiently unlike our world (for
example, if the 'cat' talked about turns out to speak Latin, or grows to a
fantastic size, or changes into a fish), we shall be at a loss when it comes
to deciding whether the statement 'There is a cat next door' is actually
true, with respect to that world, even though the state of affairs we initially
specified obtains. Is a world in which there is an animal next door exactly
like a cat in all respects except that it speaks Latin a world in which there
is a cat next door? We don't know, because "most of our empirical
concepts are not delimited in all possible directions" (120). Or, to take
another classic example:

The notion of gold seems to be defined with absolute precision, say

by the spectrum of gold with its characteristic lines. Now what would
you say if a substance was discovered that looked like gold, satisfied
all the chemical tests for gold, whilst it emitted a new sort of
radiation? 'But such things do not happen'. Quite so; but they might
happen, and that is enough to show that we can never exclude
altogether the possibility of some unforeseen situation arising in
which we shall have to modify our definition. Try as we may, no
concept is limited in such a way that there is no room for any doubt.
We introduce a concept and limit it in some directions; for instance,
we define gold in contrast to some other metals such as alloys. This
suffices for our present needs, and we do not probe any farther. We
tend to overlook the fact that there are always other directions in
which the concept has not been defined. And if we did, we could
easily imagine conditions which would necessitate new limitations. In
short, it is not possible to define a concept like gold with absolute
precision, i.e. in such a way that every nook and cranny is blocked
against entry of doubt. That is what is meant by the open texture of
a concept. (Waismann 1951: 120)

188
 We find something analogous to Waismann's embeddings in Searle's
writings on the 'background' (e.g. Searle 1978, 1980). Searle (1978)
enquires into the truth-conditions of 'the cat is on the mat'. It's easy to
describe the sort of state of affairs that would make the sentence true
(with respect to a particular assignment of values to indexical
expressions). But once we have described such a state of affairs, we can
embed it within an extraordinary situation:

Suppose that the cat and the mat are in exactly the relations
depicted only they are both floating freely in outer space, perhaps
outside the Milky Way galaxy altogether... Is the cat still on the mat?
And was the earth's gravitational field one of the things depicted...?
What I think is correct to say as a first approximation in answer to
these questions is that the notion of the literal meaning of the
sentence "The cat is on the mat" does not have a clear application,
unless we make some further assumptions, in the case of cats and
mats floating freely in outer space. (Searle 1978: 211.)

Now Waismann has an explanation for the fact that, in extraordinary

situations, 'words fail us', as Austin puts it. Here it goes:

If I had to describe the right hand of mine which I am now holding

up, I may say different things of it: I may state its size, its shape, its
colour, its tissue, the chemical compound of its bones, its cells, and
perhaps add some more particulars; but however far I go, I shall
never reach a point where my description will be completed:
logically speaking, it is always possible to extend the description by
adding some detail or other. Every description stretches, as it were,
into a horizon of open possibilities: however far I go, I shall always
carry this horizon with me. (...) [This] has a direct bearing on the
open texture of concepts. A term is defined when the sort of
situation is described in which it is to be used. Suppose for a
moment that we were able to describe situations completely without

189
 omitting anything (as in chess), then we could produce an
exhaustive list of all the circumstances in which the term is to be
used so that nothing is left in doubt; in other words, we could
construct a complete definition, i.e. a thought model which
anticipates and settles once and for all every possible question of
usage. As, in fact, we can never eliminate the possibility of some
unforeseen factor emerging, we can never be quite sure that we
have included in our definition everything that should be included,
and thus the process of defining and refining an idea will go on
without ever reaching a final stage. (Waismann 1951: 121-3)

Waismann's point is simple enough. A term is used in, or applies to,

situations of a certain type — situations with which we are acquainted. (Or
at least: our mastery of the term is somehow connected to our having
experienced the type of situation in question.) To define a term is to
describe the type of situation in question. The problem is that the
situations in question, like Waismann's right hand or any aspect of
empirical reality, cannot be completely described: they possess an
indefinite number of features, some of which may never have been
noticed and perhaps will never be noticed by anyone. When we describe
an empirical situation, we make certain features explicit, but an indefinite
number of other features remain implicit and constitute a sort of hidden
'background'. (Here I use Searle's term, on purpose.) Now the applicability
of a term to novel situations depends on their similarity to the source
situations, i.e. to the situations by association with which the term has
acquired the meaning it has. For the term to be (clearly) applicable, the
target situation must be similar to the source situations not only with
respect to those features which easily come to mind and constitute the
'explicit' definition of the term (the 'tip', to use the iceberg metaphor), but
also with respect to the hidden background. If the two situations
considerably diverge with respect to the latter, it's unclear whether or not
the term will be applicable, even though the explicit conditions of
satisfaction are satisfied.

190
 Other examples can be provided where the divergence between the
source situation and the target situation affects not only the hidden
background but also the explicit part (the tip of the iceberg). Thus Searle
(1980 p. 230-1) imagines that 'Snow is white' is uttered as a description of
the following situation:

Suppose that by some fantastic change in the course of nature the

earth is hit by an astronomical shower of radiation that affects all
existing and future water molecules in such a way that in their
crystalline form they reflect a different wave length when in sunlight
from what they did prior to the radiation shower. Suppose also that
the same shower affects the human visual apparatus and its genetic
basis so that snow crystals look exactly as they did before. Physicists
after the shower assure us that if we could see snow the way we did
before, it would look chartreuse but because of the change in our
retinas, which affects our observation of snow and nothing else, snow
looks the same color as ever and will continue to do so to ensuing
generations... Would we say that snow was still white? (Searle
1980:230).

Putnam's twin-earth thought-experiment has the same structure. It's a

case of divergence between source situations and target situation, where
the divergence affects the tip and not merely the the hidden background.
For if we are to define water, we shall say that it's a liquid with such and
such phenomenal properties, and that it's chemical structure is H20, in the
same way in which the whiteness of snow has to to both with its reflecting
a certain wavelength AND with its looking to us a certain way. We can
imagine an extraordinary situation where the two things that go together
as part of the tip are actually divorced from one another. In such cases we
don't know what to say because there is some measure of similarity
between the source situations and the target situation, but we have no
contextual clue as to which dimension of similarity matters.

191
 9.5 Ostensive definitions

Waismann's argument for the unstability of application conditions can be

objected to on two grounds. First, it presupposes 'descriptivism', i.e. the
idea that the only way to define words is to do so descriptively. Second, it
confuses the semantic and the epistemological.
(First objection) All that Waismann's argument establishes, arguably,
is that words can't be defined in purely descriptive terms. But this does
not imply that words cannot be defined at all, or that they do not possess
definite (stable) conditions of application. The reason why words can't be
defined in purely descriptive terms is that they have an irreducibly
referential dimension: they, as it were, point to real situations in the world
— situations which have an indefinite number of features and cannot be
exhaustively described. Why not simply incorporate this referential
dimension into our statements of truth- and application-conditions? Why
not follow Putnam and define a cat as an animal belonging to the same
species as this specimen (or those specimens, where the specimens in
question are normal cats found in the local environment)? Why not define
gold as this metal (while pointing to a piece of gold)? By explicitly
incorporating some ostensive reference to the actual environment into the
definition of predicates like 'gold' and 'cat', it seems that we can overcome
the alleged unstability of truth-conditions. 'There is a cat next door' will be
true, in the imagined situation (where a catlike animal next door speaks
Latin), provided the catlike animal belongs to the same species as our cats
(the cats to be found in ordinary situations).
(Second objection) Of course, we do not know whether or not the
strange animal we imagine would be considered a cat by the scientists,
just as we don't know whether or not a gold-like metal emitting a new sort
of radiation would count as gold. But our epistemic limitations do not
prevent the word 'gold' (or the word 'cat') from having a definite content
and a definite extension: gold is anything that is the same metal as this,
and a cat is any animal of the same species as those. Our epistemic
limitations, so much emphasized by Waismann, have no bearing on the

192
properly semantic issue.
These objections are well-taken but they are not decisive. In
particular, they do not threaten Waismann's conclusion (or at least, my
conclusion) regarding the unstability of truth- and application-conditions.
They would threaten that conclusion if, by appealing to ostensive
definitions, we could determine a stable content (i.e. stable conditions of
application) for the word thus defined. But we cannot. For a stable content
to be determined by an ostensive definition à la Putnam, the dimension of
similarity to the demonstrated exemplars must itself be fixed.84 But what
fixes the dimension of similarity? On Waismann's picture, words are
associated with situations of use, that is all. To apply the word to or in a
novel situation, that situation must be similar to the source situations; but
we cannot survey in advance all the possible dimensions of similarity
between the source situations and possible target situations: open texture
again.
According to Putnam (1975), the predicate 'water' means something
like: 'same-L as the transparent, odorless, thirst-quenching stuff to be
found in lakes and rivers in the local environment', or more simply 'same-L
as that' (pointing to a sample of water). If we change the demonstrated
stuff (i.e. the liquid which satisfies the stereotype of water in the local
environment) we thereby change the extension of 'water'. 'Water',
therefore, has an indexical component, Putnam says. Yet there is another
form of context-sensitivity in play here, which Putnam's ostensive
definitions do not properly capture: the dimension of similarity itself is not
given, but contextually determined. In some contexts the chemical
composition of the demonstrated stuff will be relevant, in other contexts

84 As McKay and Stern point out, "to determine the extension of the

natural kind term, we need, in addition to the sample, at least some

indication of the breadth of the term — the respects in which other

individuals [in the extension] must be related to the sample' (Thomas

McKay and Cindy Stern, 'Natural kind terms and standards of

membership', Linguistics and Philosophy 3, 1979, /27-34;/ p. 27).

193
only functional properties will be relevant. 85 Accordingly XYZ will count as
water in some contexts simply because it's that liquid which is odorless,
colorless, quenches thirst, and can be found in lakes and rivers. If the
conversation bears on the issue of what currently fills a certain bottle, milk
or water, then plainly the answer 'it's water' will be true, even if it's XYZ
rather than H20 (Chomsky 2000 : 41,). Putnam himself expresses
awareness of this point in the Meaning of Meaning (Harnish ed. p.243):

x bears the relation same-L to y just in case (1) x and y are both
liquids, and (2) x and y agree in important physical properties....
Importance is an interest-relative notion. Normally the 'important'
properties of a liquid or solid, etc., are the ones that are structurally
important; the ones that specify what the liquid or solid, etc., is
ultimately made out of... From this point of view the important
characteristic of a typical bit of water is consisting of H20. But it may
or may not be important that there are impurities... And structure may
sometimes be unimportant; thus one may sometimes refer to XYZ as
water if one is using it as water...
Even senses that are so far out that they have to be regarded as a
bit 'deviant' may bear a definite relation to the core sense. For
example I might say 'did you see the lemon', meaning the plastic
lemon. A less deviant case is this: we discover 'tigers' on Mars. That
is, they look just like tigers, but they have a silicon-based chemistry
instead of a carbon-based chemistry... Are Martian 'tigers' tigers? It

85 A Kripkean/Putnamian might object that we are concerned only with

what words mean in our language, not with what they mean in other

possible situations. Now in our language water is H20, and the truth-

conditions of 'This is water' is that this stuff is H20. But this objection is

mistaken. I am talking about OUR language. Our language is such that we

can imagine contexts in which it would be true to say 'This is water' of

some stuff if the stuff in question had certain phenomenal/functional

properties, whether or not the thing in question was H20.

194
 depends on the context.

We now have in hand all the ingredients necessary to define the

theoretical position I want to ascribe to the (radical) contextualist: the
irreducibly referential dimension of meaning; the central role of similarity
in determining extension (hence truth-conditions); and finally the context-
dependence of similarity relations. With those ingredients, we can concoct
an extreme view very much in the spirit of Austin and Wittgenstein:
Meaning Eliminativism.

9.6. Meaning Eliminativism

According to WF, the sense expressed by an expression must always be

contextually constructed on the basis of the (overly rich or overly abstract)
meaning, or semantic potential, of the word type. Just as the reference of
an indexical expression is not linguistically given but must be contextually
determined, the sense of an ordinary expression is not linguistically given
but must be constructed. In that framework there still is a role for the
linguistic meaning of word types: it is the input (or one of the inputs) to
the construction process.
The difference between Meaning Eliminativism (ME) and WF is that,
according to ME, we don't need linguistic meanings even to serve as input
to the construction process. The senses that are the words' contributions
to contents are constructed, but the construction can proceed without the
help of conventional, context-independent word meanings.
Note that, according to a trivial extension of WF, the linguistic
meaning of a word is not merely the input to the process of semantic
modulation: it is also the output of a process of induction through which
the child, or anyone learning the language, abstracts the meaning of the
word from the specific senses which it expresses, or seems to express, on
the observed occasions of use. It is a truism that the child or language
learner starts not with pre-formatted linguistic meanings, but with actual
uses of words and the contextualised senses that words assume on such

195
uses. So both contextualised senses and context-independent linguistic
meanings are input, and both are output, in some construction process.
The linguistic meaning of a word type is the output of an abstraction
process; that process takes as input the contextualised senses used as
evidence by the language learner. On the other hand, the linguistic
meaning of a word type also serves as input to the modulation process
which yields as output the contextualised sense of the word on a particular
occasion of use (Figure 9.1).

[Figure 9.1 about here]

ME purports to simplify WF by suppressing the intermediary step

(linguistic meaning) and computing directly the contextual sense which an
expression assumes on a particular occasion of use on the basis of the
contextual senses which that expression had on previous occasions of use
— without ever abstracting, or needing to abstract, 'the' linguistic meaning
of the expression type.86 This amounts to merging the two construction
processes: the abstraction of meaning from use, and the modulation of
meaning in use (Figure 9.2). According to ME, there is a single process of
abstraction-modulation which takes as input previous uses of the
expression and yields as output the contextual sense assumed by the
expression on the current use.

[Figure 9.2 about here]

On the resulting picture, words are not primitively associated with

abstract 'conditions of application', constituting their conventional

86 For a detailed psychological model supporting ME, see D. Hintzman,

'"Schema Abstraction" in a Multiple-Trace Memory Model', Psychological

Review 93: 411-28, 1986, and 'Judgments of Frequency and Recognition

Memory in a Multiple-Trace Memory Model', Psychological Review 95: 528-

51, 1988.

196
meaning (as on the Fregean picture). The conditions of application for
words must be contextually determined, like the reference of indexicals.
What words, qua linguistic types, are associated with are not abstract
conditions of application, but rather particular applications.
In the spirit of Wittgenstein, consider what it is for someone to learn
a predicate P. The learner, whom I'll call Tom, observes the application of P
in a particular situation S; he associates P and S. At this stage, the
semantic potential of P for Tom is the fact that P is applicable to S. In a
new situation S', Tom will judge that P applies only if he finds that S'
sufficiently resembles S. To be sure, it is possible for S' to resemble S in a
way that is not pertinent for the application of P. The application of P to S'
will then be judged faulty by the community, who will correct Tom. The
learning phase for Tom consists in noting a sufficient number of situations
which, like S, legitimate the application of P, as opposed to those, like S',
which do not legitimate it. The semantic potential of P for Tom at the end
of his learning phase can thus be thought of as a collection of legitimate
situations of application; that is, a collection of situations such that the
members of the community agree that P applies in or to those situations.
The situations in question are the source-situations. The future
applications of P will be underpinned, in Tom's usage, by the judgement
that the situation of application (or target-situation) is similar to the
source-situations.
In this theory the semantic potential of P is a collection of source-
situations, and the conditions of application of P in a given use, involving a
given target-situation S'', are a set of features which S'' must possess to
be similar to the source-situations. The set of features in question, and so
the conditions of application for P, will not be the same for all uses; it is
going to depend, amongst other things, on the target-situation. One
target-situation can be similar to the source-situations in certain respects
and another target-situation can be similar to them in different respects.
But the contextual variability of the conditions of application does not end
there. Even once the target-situation is fixed, the relevant dimensions for
evaluating the similarity between that situation and the source-situations

197
remain under-determined: those dimensions will vary as a function of the
subject of conversation, the concerns of the speech participants, etc.
One particularly important factor in the contextual variation is the
relevant 'contrast set'. As Tversky has pointed out, judgements of
similarity are very much affected by variations along that dimension. 87 If
we ask which country, Sweden or Hungary, most resembles Austria
(without specifying the relevant dimension of similarity), the answer will
depend on the set of countries considered. If that set includes not just
Sweden, Hungary and Austria but also Poland, then Sweden will be judged
more like Austria than Hungary; but if the last of the four countries
considered is Norway and not Poland, then it is Hungary which will be
judged more like Austria than Sweden. The explanation for that fact is
simple. Poland and Hungary have certain salient geopolitical features in
common which can serve as basis for the classification: Hungary and
Poland are then put together and opposed to Austria and Sweden. If we
replace Poland by Norway in the contrast set a new principle of
classification emerges, based on the salient features shared by Norway
and Sweden: in this new classification Hungary and Austria are back
together. Tversky concludes that judgements of similarity appeal to
features having a high 'diagnostic value' (or classificatory significance),
and that the diagnostic value of features itself depends on the available
contrast set.
So the set of similarity features on which sense depends itself
depends upon the relevant contrast set, and the relevant contrast set
depends upon the current interests of the conversational participants. It
follows that one can, by simply shifting the background interests ascribed
to the conversational participants, change the truth-conditions of a given
utterance, even though the facts (including the target-situation) don't
change, and the semantic values of indexicals remain fixed. Charles Travis
has produced dozens of examples of this phenomenon of truth-conditional
shiftiness over the last thirty years, and his examples often involve

87 See Amos Tversky, ‘Features of Similarity’, Psychological Review 84,

1977, pp. 327-52.

198
manipulating the relevant contrast-set.88
In this framework the background-dependence of truth-conditions
emphasized by Searle is accounted for by appealing to the global
character of the similarity between target-situation and source-situations.
As Waismann stresses the source-situations are concrete situations with an
indefinite number of features. Some of these features are ubiquitous and
their diagnostic value in a normal situation is vanishing. 89 They belong to
the most general and immutable aspects of our experience of the world:
gravity, the fact that food is ingested via the mouth, etc. When we specify
the truth conditions of a sentence (for example the sentence 'The cat is on
the mat'), or the conditions of application of a predicate (for example the

88 The following example (inspired from Austin) is taken almost at random

from a list of Travis-examples compiled by a student of mine:

Fred is walking with his young nephew beside a pond where a decoy duck

is floating. Pointing to the decoy, he says, "That's a duck". Again we might

ask whether what he said is true or false. But again, the above description

is not enough for us to tell. If Fred has just finished laughing at a

sportsman who blasted a decoy out of the pond, and if he has been trying

to show his nephew how to avoid similar mistakes, then what he said is
false. But suppose that Fred and his nephew are attending the annual

national decoy exhibition, and the boy has been having trouble

distinguishing ducks from geese. Then what Fred said may well be true. It

would also be true had Fred said what he did in pointing out the fact that

all the other ducks were poor copies (perhaps on the order of Donald

Duck). (Travis 1975: 51)

89 See ‘Features of similarity’, p. 342: "The feature 'real' has no diagnostic

value in the set of actual animals since it is shared by all actual animals

and hence cannot be used to classify them. This feature, however, aquires

considerable diagnostic value if the object set is extended to include

legendary animals, such as a centaur, a mermaid or a phoenix."

199
predicate 'on' in that sentence), we only mention a small number of
features — the 'foreground' features — because we take most of the
others for granted; so we do not mention gravity, we presuppose it.
Nevertheless, gravity is one of the features possessed by the situations
which are at the source of the predicate 'on'; and there is an indefinite
number of such features. These background features of the source-
situations can be ignored inasmuch as they are shared by the situations of
which we can want to speak when we utter the sentence; but if we
imagine a target-situation where the normal conditions of experience are
suspended, and where certain background features of the source-
situations are not present, then we shatter the global similarity between
the target-situation and the source-situations. Even if the target-situation
has all the foreground features which seem to enter into the 'definition' of
a predicate P, it suffices to suspend a certain number of background
features in order to jeopardize the application of P to the target-situation.
That shows that the semantic potential of P is not, as in Fregean
semantics, a set of conditions of application determined once and for all,
but a collection of source-situations such that P applies to a target-
situation if and only if it is relevantly similar to the source-situations.
A caveat: as Searle himself emphasizes, the fact that the target-
situation does not possess certain background features of the source-
situations does not automatically entail the non-applicability of the
predicate P. It can be that the background features which the target-
situation does not possess (for example gravity) are contextually irrelevant
and do not affect the application conditions of the predicate. 90 For the

90 It is easy to imagine a context with respect to which it would be

definitely true to say 'the cat is on the mat' of the gravitationless situation

described by Searle:

For example, as we are strapped in the seats of our spaceship in outer

space we see a series of cat-mat pairs floating past our window. Oddly,

they come in only two attitudes. From our point of view they are either as

depicted in Figure 1 or as would be depicted if figure 1 were upside down.

200
same sort of reason, the possession by the target-situation of what I have
called the foreground features of the source-situations is no more a
necessary condition for the application of the predicate than it is a
sufficient condition. For a predicate (or a sentence) to apply to a target-
situation that situation must resemble the source-situations under the
contextually relevant aspects. So a predicate can apply even if the target-
situation differs markedly from the source-situations, as long as, in the
context and taking into account the contrast set, the similarities are more
significant than the differences. Thus, in certain contexts, as Putnam
notices, the predicate 'lemon' will apply to plastic lemons, or the word
'water' to XYZ.

9.7. Conclusion

Meaning Eliminativism gets rid of abstract meanings for types, in favour of

particular uses. The contextualised sense carried by the word on a
particular use depends upon similarity relations between that use of the
word (as applied to the 'target-situation') and past uses of the same word
(as applied to the 'source-situations'). The features of similarity that are
perceived are so context-dependent (and in particular, contrast-
dependent) that the truth-conditional sense of the word is susceptible to
vary, and to vary more or less indefinitely, even after we have fixed (i) the
collection of past situations of use, (ii) the target-situation in which or to
which the word is currently applied, and (iii) the reference of indexicals
and the semantic values of all expressions in need of saturation.
Meaning Eliminativism is, as I said, the most extreme form of
Contextualism. It is probably too extreme, but what strikes me most in that
framework is its surprising viability. The knock-down objections such an
extreme view is supposed to raise (and does raise in the classroom) do not
look so devastating when explicitly stated. To close this chapter, let me
briefly consider two standard objections to the elimination of linguistic

"Which is it now?", I ask. "The cat is on the mat", you answer. Have you

not said exactly and literally what you meant? (Searle 1978 p.212).

201
meaning.
The first objection (which I may have heard about fifty times) runs
like this. Granted, the senses expressed by words are affected by
contextual factors; they are 'modulated'. But we must treat modulation
effects as arising from the interaction of context with context-independent
linguistic meanings. If, instead, we go eliminativist and say that the
context does everything, while the words themselves contribute nothing,
then it does not matter any more whether the word one utters is 'red' or
'rectangle'. But it makes a big difference which word we actually utter. This
shows that Meaning Eliminativism goes too far: it throws out the baby with
the bathwater.
This objection presupposes that, for the eliminativist, "the words
themselves contribute nothing". But that is not true. There is something
which words do contribute, and which is sufficient to account for the
difference it makes when we substitute a word for another one. That thing
is not a 'meaning' in the traditional sense: it is what I called the 'semantic
potential' of words. 'Red' and 'rectangle' have very different semantic
potentials. The semantic potential of a word is the collection of past uses
on the basis of which similarities can be established between the source
situations (i.e. the situations such uses concerned) and the target situation
(the situation of current concern). So there is, indeed, interaction between
the 'context' and what words themselves contribute. Context does not do
everything, even on the most extreme form of Contextualism.
According to the second objection, there is actually more stability in
semantic content than ME suggests. Not only do words have relatively
stable conditions of application, across uses; there is also interpersonal
stability: language users converge in their judgments regarding conditions
of application or truth-conditions. How will that stability be accounted for?
With regard to the first part of the objection the contextualist pleads
guilty: the unstability and shiftiness of truth-conditional content across
uses he takes to be an empirical fact, sufficiently established by the
wealth of examples and thought-experiments provided by the authors I
have mentioned. Still, according to the objection, the contextualist goes

202
too far in that direction. But the contextualist will deny that that is so. The
degree of unstability achieved will depend upon details of the model, and
can be varied so as to fit experimental data. (See Hintzman 1986, 1988 for
suggestions.) — As for the second part of the objection, the contextualist
will be happy to grant the interpersonal stability of truth-conditional
content. That stability can be accounted for on psychological rather than
linguistic grounds. Thus Bolinger notes that

speakers not only share the same code but also share the ability to
see the same resemblances between what their code already
designates and what they would like it to designate, and so to make
the old forms reach out to new meanings. That is how language
breaks free of its rigidity. (D. Bolinger, Aspects of Language).

In any event, there are all sorts of versions of the contextualist

position which are intermediate between WF and ME. Like Rumelhart and
most lexical semanticists, we may consider that words are associated not
so much with (a collection of) particular, episodic situations of use, but
with more abstract schemata corresponding to types of situations. Insofar
as there still is a referential or quasi-referential dimension in virtue of
which, through these schemata, words point to actual situations of use,
the basic mechanisms posited by ME remain in play. Or we may favour a
model in which both episodic traces of situations of use and abstract
schemata are stored and exploited in contextually determining truth-
conditional content. Whichever option we take the important contextualist
idea should not be misconstrued. How much abstraction takes place is not
the issue, for the contextualist. The important idea — which forms the core
of Austin's theory of truth — is twofold: first, words are associated with
worldly situations (at whatever level of abstractness and fine-
grainedness), i.e. entities at the level of reference rather than the level of
sense; second, the content (sense) of words depends upon context-
sensitive similarities between those situations and the target-situation in
which or to which the word is applied.

203