If you have difficulty printing this file, please click here for the Adobe Acrobat (.pdf) version.

 

 

Unarticulated Constituents

 

François Recanati

CREA, Paris

recanati@poly.polytechnique.fr

 

 

Summary

 

I. Truth-conditional pragmatics

1.         Two kinds of interpretation

2.         What is wrong with the standard picture

3.            Primary pragmatic processes

4.         From Radical Literalism to Radical Contextualism

 

II. A literalist response

5.            Stanley's attack on TCP

6.            Versions of Minimalism

7.         The argument from binding

 

III. Testing for unarticulatedness

8.         Bogus unarticulated constituents

9.            Unarticulated constituents are never mandatory

10.       'It is raining'

11.            Variadic functions

 

IV The argument from binding: a refutation

12.       The Binding Criterion

13.            Existential closure by default?

14.       The Binding Fallacy

 

V. 'Bound' unarticulated constituents

15.       The Weak Binding Criterion

16.       Indirect binding via unarticulated functions

17.            'Wh+ever'-phrases

18.       To bind or not to bind

19.       Free enrichment: the 'syntactic' interpretation

 

 

PART 1

TRUTH-CONDITIONAL PRAGMATICS

 

 

I. TWO KINDS OF INTERPRETATION

 

Most semanticists, whether philosophers or linguists, view the division of labour between semantics and pragmatics as follows. 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. So the usual story goes.

            As Grice emphasized, speaker's meaning is 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 meaning 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 typically goes beyond, or otherwise diverges from, what the uttered sentence literally means. In such cases the hearer must rely on background knowledge to determine what the speaker means — what his or her communicative intentions are — on the basis of what he or she actually says.

            On this view two distinct and radically different processes are jointly involved in the interpretation of linguistic utterances: semantic interpretation and pragmatic interpretation. They are standardly described as follows:

 

• Knowing a language is like knowing a theory by means of which one can deductively establish the truth-conditions of arbitrary sentences of that language. Semantic interpretation consists in applying that theory to a particular sentence of the language so as to determine its truth-conditions on the basis of the references of its parts and the way they are syntactically combined.

 

• Pragmatic interpretation is a totally different process. It is not concerned with language per se, but with human action. When someone acts, whether linguistically or otherwise, 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 (for example, a communicative intention) in terms of which we can make sense of the action.

 

Two important characteristics of pragmatic interpretation, as opposed to semantic interpretation, stand out and must be stated from the outset.

            First, pragmatic interpretation is possible only if we presuppose that the agent is rational. To interpret an action, we have to make hypotheses concerning the agent's beliefs and desires; hypotheses in virtue of which it can be deemed rational for the agent to behave as she does. This feature of pragmatic interpretation I will refer to as its hermeneutic character. It strikingly contrasts with the algorithmic, mechanical character of semantic interpretation (as standardly conceived).

            Second, and relatedly, pragmatic interpretation is always defeasible. 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. This, again, contrasts with the non-defeasible, monotonous character of semantic interpretation.

            A third contrast worth stating concerns the role of context in semantic and pragmatic interpretation. Because of its defeasibility — what Stainton (forthcoming) calls its 'all-things-considered' character — there is no limit to the amount of contextual information that can in principle affect pragmatic interpretation. But context comes into play in semantic interpretation only to help determine the reference of those few expressions whose reference is not fixed directly by the rules of the language but is fixed by them only 'relative to context'. The context at issue is a small package of factors involving only very limited aspects of the actual situation of utterance: who speaks, when, where, to whom, and so forth. In contrast, the context relevant to determining what the speaker means is all-inclusive. Any aspect of the total world in which the utterance takes place (not to mention the 'possible worlds' projected by the beliefs, intentions, etc., of the language users) is part of the context which can affect pragmatic interpretation. In a nutshell: Anything can affect pragmatic interpretation (as opposed to semantic interpretation, which is 'informationally encapsulated').^[1]

            From what has just been said, it follow that there are two notions of context: a narrow and a broad one, corresponding to semantic and pragmatic interpretation respectively. As Kent Bach puts it:

 

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 context pragmatic.^[2]

 

When the (narrow) context comes into play to determine the semantic values of indexicals, it does so in the algorithmic and non-hermeneutical manner which is characteristic of semantic interpretation as opposed to pragmatic interpretation. The narrow context determines, say, that 'I' refers to John when John says 'I' quite irrespective of John's beliefs and intentions. As Barwise and Perry write, "even if I am fully convinced that I am Napoleon, my use of ‘I’ designates me, not him. Similarly, I may be fully convinced that it is 1789, but it does not make my use of ‘now’ about a time in 1789" (1983: 148).

            That, then, is the standard picture. It has come under sustained attack during the last fifteen years, and an alternative picture has been put forward: Truth-conditional pragmatics (TCP). My aim, in this paper, is to defend TCP against a counter-attack mounted by Jason Stanley in his paper 'Context and Logical Form'.^[3]

 

II. WHAT IS WRONG WITH THE STANDARD PICTURE

 

According to the standard picture, knowledge of a language enables interpreters to deductively assign truth-conditions to arbitrary sentences of that language. To be sure, if the sentence contains an indexical expression, its truth-conditional content will depend upon the context; mere knowledge of the language will not be sufficient for truth-conditional interpretation. But the context relevant to content-determination in such cases is the narrow context. To fix the truth-conditional content of even indexical utterances, there is no need to engage in 'all-things-considered' reasoning, no need to take the speaker's beliefs and intentions into consideration. That is the gist of the standard picture. Knowledge of the language and, if necessary, of the (narrow) context suffices for truth-conditional interpretation. Considerations pertaining to the speaker's beliefs and intentions become relevant only when we are to (nondemonstratively) determine what the speaker means by or in saying what she says. In other words: semantic interpretation delivers truth-conditions; pragmatic interpretation determines aspects of utterance meaning over and above truth-conditions.

            It is this claim that has been called into question. It has been argued that pragmatic interpretation is necessary to get truth-conditions in the first place. Pragmatic processes are involved not only to determine what the speaker means on the basis of what she says, but also to determine what is said, insofar as this is distinct from the meaning of the sentence type.

            Recall that, on the standard view, the reference of indexicals is determined automatically on the basis of a linguistic rule, without taking the speaker's beliefs and intentions into consideration. Now this may be true of some of the expressions which Kaplan classifies as 'pure indexicals' (Kaplan 1989), but it is certainly not true of those which he calls '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.

            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 incorporate a sequence of 'speaker's intended referents' into the narrow context, in such a way that the n^th demonstrative in the sentence will refer to the n^th 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 pure indexicals, while in fact we can only determine the speaker's intended referent (hence the narrow context relevant to the interpretation of the utterance) by resorting to pragmatic interpretation and relying on the wide context.

            We encounter the same problem even with expressions like 'here' and 'now' which Kaplan classifies as pure indexicals (rather than demonstratives). Their semantic value is said to be 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 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, formal tricks notwithstanding, pragmatic interpretation has a role to play in determining the content of the utterance.

            The alleged automaticity of content-determination and its independence from pragmatic considerations is an illusion due to an excessive concern with a sub-class of 'pure indexicals', namely words such as 'I', 'today' etc. But they are only a special case — the end of a spectrum. In most cases the reference of a context-sensitive expression is determined on a pragmatic basis. That is true not only of standard indexical expressions, but also of many constructions involving something like a free variable. For example, a possessive phrase such as 'John's car' arguably means something like the car that bears relation R to John. The free variable R 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 dependence upon the wide context is a characteristic feature of 'semantically indeterminate expressions', which are pervasive in natural language. Their semantic value varies from occurrence to occurrence, 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.

 

III. PRIMARY PRAGMATIC PROCESSES

 

The pragmatic processes that are involved in the determination of what is said, and which justify talk of 'truth-conditional pragmatics', I call primary pragmatic processes. Qua pragmatic processes, they have all the properties characteristic of pragmatic interpretation: in particular, defeasibility and dependence upon the wide context. (This last property I will henceforth refer to as 'background-dependence', so as to distinguish it from the narrower form of context-dependence illustrated by pure indexicals such as the pronoun 'I'.) According to truth-conditional pragmatics, when an utterance is made and a certain truth-conditional interpretation emerges for that utterance, it does so as a result of pragmatic processes that can be affected by any change in the wide context ('background'); it follows that one can, by imaginatively altering the background, affect the truth-conditions of the utterance even though one leaves the narrow context untouched.^[4] A given truth-conditional interpretation for an utterance can therefore always be revised in the light on additional background information.

            There are two classes of primary pragmatic processes: top-down and bottom-up. Before presenting that distinction, I should first say something about what distinguishes primary pragmatic processes in general from the more traditional sort of pragmatic processes evoked in the Gricean literature: secondary pramatic processes, as I call them.

            Secondary pragmatic processes presuppose that something has been said (some proposition expressed). They are 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 M-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 implicatures'. 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.

            In contrast to secondary pragmatic processes, primary pragmatic processes do not presuppose that some proposition has been identified or determined; they are involved in the very determination of what is said. Their input is not a complete proposition, but the linguistic meaning of the sentence, of which the language users need not be consciously aware. (In contrast, as we have just seen, participants in the speech process are aware of both what is said — the input to secondary pragmatic processes — and of what the speaker implies by saying it, as well as of the inferential connection between them.)

            I said above that there are two types of primary pragmatic processes. The determination of the reference of indexicals and, more generally, the determination of the content of context-sensitive expressions is a typical bottom-up process, i.e. a process triggered (and made obligatory) by a linguistic expression in the sentence itself. For example, if the speaker uses a demonstrative pronoun and says 'She is cute', the hearer must determine who the speaker means by 'she' in order to fix the utterance's truth-conditional content. Similarly, if the speaker uses the genitive construction 'John's car', the hearer must determine which relation R is meant to hold between John and the car at issue. The expression itself acts as a variable in need of contextual instantiation; it sets up a slot which the interpreter has to fill. It is in that sense that the pragmatic process at issue here — 'saturation', as I call it — is a bottom-up process. But there are other primary pragmatic processes which are not bottom-up. Far from being triggered by an expression in the sentence, they take place for purely pragmatic reasons.

            To give a standard example, suppose someone asks me, at about lunch time, whether I am hungry. I reply: 'I've had a very large breakfast'. In this context, my utterance conversationally implies that I am not hungry. In order to retrieve the implicature, the interpreter must first understand what is said — the input to the secondary pragmatic process responsible for implicature generation. That input is the proposition that the speaker has had a very large breakfast... when? No time is specified in the sentence, which merely describes the posited event as past. On the other hand, the implicature that the speaker is not hungry could not be derived if the said breakfast was not understood as having taken place on the very day in which the utterance is made. Here we arguably have a case where something (the temporal location of the breakfast event on the day of utterance) is part of the intuitive truth-conditions of the utterance yet does not correspond to anything in the sentence itself.^[5] If this is right, then the temporal location of the breakfast event is an unarticulated constituent of the statement made by uttering the sentence in that context.

            Such unarticulated constituents, which are part of the statement made even though they correspond to nothing in the uttered sentence, result from a primary pragmatic process of free enrichment — 'free' in the sense of not being linguistically controlled. What triggers the contextual provision of the relevant temporal specification in the above example 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. Both types of process are primary since they contribute to shaping the intuitive truth-conditions of the utterance, which intuitive truth-conditions serve as input to secondary pragmatic processes.

            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.

 

IV. FROM RADICAL LITERALISM TO RADICAL CONTEXTUALISM

 

Many theorists think one should not allow top-down processes, which are pragmatic through and through, to affect the proposition expressed by an utterance. In order to be part of what is literally said, they claim, a pragmatically provided constituent must at least correspond to something in the sentence. It  must be 'articulated'. This constraint is what, in previous writings, I referred to as (Pragmatic) Minimalism.^[6]

 

Minimalism

What is said is affected by the bottom-up process of saturation but not by top-down processes such as free enrichment.

 

In those writings I argued against Minimalism, on the following grounds. Once pragmatic interpretation is allowed to play a role in the determination of what is said, it is somewhat arbitrary to set limits to its operation, as Minimalism attempts to do. From a psychological point of view, we cannot separate those aspects of speaker's meaning which fill gaps in the mental 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. They are indissociable, mutually dependent aspects of a single process of pragmatic interpretation (see Recanati 1995 for an illustration of this interdependence).

            The suspicion has arisen in several quarters that the quarrel between Truth-conditional pragmatics and Minimalism may well be verbal rather than substantive. To some extent, I agree (see section V below). If the notion of 'what is said' we are trying to characterize is meant to capture the intuitive truth-conditions of an utterance qua input to secondary pragmatic processes, then it must be acknowledged that what is said, in that sense, incorporates unarticulated constituents and is therefore affected by free enrichment. (Or so it seems.) But this does not prevent us from defining another notion of what is said, conforming to Minimalism. In Recanati 1999 I have used subscripts to distinguish the two notions, and I will do so here again. Let 'what is said_min' be the proposition expressed by an utterance when the effects of top-down pragmatic processes such as free enrichment have been discounted, in accordance with Minimalism; and let 'what is said_int' correspond to the intuitive truth-conditions of the utterance, which may well result from the operation of such processes. Both what is said_min and what is said_int are shaped by pragmatic interpretation, but not to the same extent. If I am right what is said_int is affected by top-down processes such as free enrichment, whereas the only pragmatic processes that are allowed to affect what is said_min are those that are triggered by something in the sentence itself.

            The view according to which there are two equally legitimate notions of 'what is said', each corresponding to a distinct level in the interpretation of an utterance, I have dubbed the Syncretic View in Recanati 2000. It is a reasonable and balanced position. It is actually a mixture of two views: Truth-conditional pragmatics (TCP), and Minimalism or rather a particular version of it (S*-MIN). TCP consists of theses (i) and (ii), while S*-MIN consists of thesis (iii). The Syncretic View is the conjonction of (i), (ii), and (iii).

 

TCP

(i) Pragmatic interpretation is needed to contextually determine the content of context-sensitive expressions. In other words: saturation is a full-fledged pragmatic process.

(ii) The intuitive truth-conditions of many utterances are also affected by top-down pragmatic processes such as free enrichment.

 

S*-MIN

(iii) The proposition literally expressed by the utterance (what is said_min) is affected by the bottom-up process of saturation but not by top-down processes such as free enrichment.^[7]

 

            Beside TCP and S*-MIN, which are compatible and whose conjunction yields the Syncretic View, there are two extreme positions. Like the Syncretic View, Radical Contextualism incorporates TCP, but what it adds to it is not S*-MIN but its denial. Radical Contextualism denies that there is a 'minimal' proposition resulting from saturation but unaffected by free enrichment (section VI). No less extreme is Radical Literalism. It incorporates S*-MIN, as the Syncretic View does, but what it adds to it is not TCP, but its negation. Radical Literalism denies both theses (i) and (ii). It denies that the content of context-sensitive expressions depends upon what the speaker means, and holds instead that it is determined by linguistic rules (with respect to the narrow context). It denies even more strongly that the intuitive truth-conditions of an utterance are ever affected by top-down processes of pragmatic interpretation. In other words, Radical Literalism offers a full-fledged defense of the Standard Picture against TCP.

            As some of my readers may know, I find Radical Contextualism appealing and I have mentioned it with favour in a couple of places (Recanati 1993: 259-60; Recanati forthcoming); but I will not attempt to defend that extreme position here. What I want to do, as I said at the beginning of the paper, is to defend TCP against the literalist counter-attack mounted by Jason Stanley.

 

 

PART 2

A LITERALIST RESPONSE

 

 

V. STANLEY'S ASSAULT ON TCP

 

Stanley objects to TCP on the grounds that it makes truth-conditional interpretation similar to "the kind [of interpretation] involved in interpreting kicks under the table and taps on the shoulder" (Stanley 2000: 396). That, indeed, is the gist of TCP. To refute TCP, one must dispose of both (i) and (ii). Stanley discusses only (ii) in his paper, so his defense of the Standard Picture against TCP is at best incomplete. On his behalf, one may argue that, perhaps, he is prepared to concede (i) to TCP. Perhaps he is prepared to grant that pragmatic interpretation plays a role in the determination of truth-conditional content, provided the effects of pragmatic interpretation on truth-conditional content are restricted to assigning values to context-sensitive elements in the expression uttered. Let us assume that that is his actual position (as seems likely). Then an advocate of TCP has an easy reply: the concession is sufficient to establish the conclusion that truth-conditional interpretation is similar to the interpretation of nonlinguistic behaviour (similar in that both involve pragmatic interpretation).

            Stanley acknowledges that "each [context-sensitive] element brings with it rules governing what context can and cannot assign to it, of varying degrees of laxity" (Stanley 2000: 396;  emphasis mine). Presumably, when the degree of laxity of the linguistically encoded constraint is high (as it is for the majority of context-sensitive expressions), the content of the expression, hence the truth-conditions of the utterance containing it, can be determined only by appealing to pragmatic interpretation. This, in itself, is sufficient to make truth-conditional interpretation similar to that of kicks under the table: in both cases, as TCP emphasizes, pragmatic interpretation must be appealed to to make sense of the agent's behaviour, whether linguistic or nonlinguistic. (To be sure, truth-conditional interpretation also involves semantic interpretation, while kicks on the table require only pragmatic interpretation. So truth-conditional interpretation has a hybrid character and is not entirely similar to the interpretation of nonlinguistic acts — very far from it. But no one, and in particular no advocate of TCP, has ever denied that there was such a difference between the interpretation of utterances and the interpretation of nonlinguistic behaviour. On the contrary, advocates of TCP, e.g. Sperber and Wilson 1986, insist on the hybrid character of truth-conditional interpretation.)

            Contrary to what Stanley appears to believe, then, thesis (ii) is not necessary to reach TCP's conclusion that truth-conditional interpretation rests in part on pragmatic interpretation. Rather, it is that conclusion, established on the basis of (i), which makes (ii) plausible: if truth-conditional content is determined in part pragmatically, as (i) is sufficient to establish, why could it not be affected by pragmatically controlled (top-down) pragmatic processes as well as by linguistically controlled (bottom-up) pragmatic processes? Once we take what the speaker means into consideration in order to determine what she says, we can use it not only to assign values to free variables in the sentence's logical form, but also to make the resulting proposition more determinate in ways which may well go beyond what the logical form itself requires. Indeed, there is an indefinite number of examples in which the intuitive truth-conditions of an utterance do seem to involve unarticulated constituents. (1) below is a case in point:

 

(1) John and Claudia got married and they had many children

 

We naturally understand this as meaning (a) that John and Claudia got married to each other, (b) that they had many children together, and (c) that they had the children after getting married. Yet, arguably, none of these aspects of the intuitive truth-conditions of (1) is explicitly articulated in the sentence. They are, or seem to be, generated through free enrichment.

            Stanley's attack on TCP consists in showing, or attempting to show, that whenever the intuitive truth-conditions of an utterance seem to involve an unarticulated constituent, as in the above example, there actually is a free variable in logical form whose contextual instantiation generates the alleged unarticulated constituent. In other words, all alleged cases of free enrichment turn out to be cases of saturation; all alleged unarticulated constituents turn out to be articulated. That he purports to establish by means of a single argument: the argument from binding (section VII). If the argument actually has the power which Stanley ascribes it, thesis (ii) is disposed of, and the only obstacle that remains to full endorsement of Radical Literalism is (i).

 

VI. VERSIONS OF MINIMALISM

 

Since he argues that what is said is affected by extralinguistic context only in the course of assigning values to context-sensitive expressions in the sentence or to covert variables in logical form, Stanley's argument supports Minimalism. But Minimalism per se is not incompatible with TCP (section IV). There are several variants of Minimalism, only some of which are actually incompatible with TCP. Since Stanley's version is one of them, we must say something about the different variants and how they differ from each other.             Stipulative Minimalism (S-Minimalism) uses the Minimalist constraint as a criterion for demarcating 'what is said'. What is literally said is defined as satisfying Minimalism, that is, as being unaffected by top-down factors. Kent Bach ascribes to Grice a version of S-Minimalism. According to Grice's stipulation, Bach says, "what is said must correspond to 'the elements of [the sentence], their order, and their syntactic character' (1989, p. 87). So if any element of the content of an utterance... does not correspond to any element of the sentence being uttered, it is not part of what is said." (Bach 2000). Since that is a stipulation concerning the use of the phrase 'what is literally said', there is no way to disagree with such a view, except on terminological grounds. Even a Radical Contextualist cannot disagree with S-Minimalism, precisely because of its stipulative character.

            We can make S-Minimalism a little more empirical and a little less stipulative by augmenting it with an existence claim. The resulting view, S*-Minimalism, also uses the Minimalist constraint to define 'what is said', but it adds to the definition the following claim: that the notion so defined has a nonempty extension, i.e. that there is a level of meaning satisfying the Minimalist constraint. When S-Minimalism is thus strenghtened into S*-Minimalism, nonterminological disagreement becomes possible. Indeed Radical Contextualism denies the existence claim which distinguishes S*-Minimalism from S-Minimalism. According to Radical Contextualism, "no proposition could be expressed without some unarticulated constituent being contextually provided" (Recanati 1993: 260), hence the extension of the notion of 'what is said' defined according to the Minimalist stipulation is empty.

            Though it conflicts with Radical Contextualism, S*-Minimalism is still compatible with TCP; for the level of meaning it posits, which satisfies the Minimalist constraint by definition, need not be the same level of meaning as that which concerns TCP, namely the intuitive truth-conditions of the utterance (what is said_int); hence there need be no contradiction between TCP's nonminimalist characterisation of what is said_int and S*-Minimalism. (Indeed the Syncretic View incorporates both TCP and S*-Minimalism.)

            What, then, is the form of Minimalism with which TCP conflicts? As I pointed out above, the quarrel between TCP and Minimalism has been terminological to a certain extent. People using 'what is said' in accordance with the Minimalist stipulation have been annoyed when TCP theorists like myself started using the phrase 'what is said' in a different way. Instead of using the Minimalist constraint as a criterion for demarcating what is said, I explicitly put foward a different criterion: the 'availability' criterion, according to which what is said is the proposition determined by the truth-conditional intuitions of the participants in the talk-exchange themselves. But that terminological difference is not the whole story. Once 'what is said' has been demarcated using the availability criterion, it becomes an empirical question whether or not it satisfies the Minimalist constraint. The latter is no longer construed as a defining criterion, but as an empirical characterization. According to TCP, that empirical characterization is falsified by all the cases in which what is said (in the sense of TCP, that is, what is said_int) involves unarticulated constituents. Thus TCP-theorists claimed to have refuted Minimalism. What TCP-theorist were attempting to refute was neither S-Minimalism nor S*-Minimalism, however, but a third variant which I call I-Minimalism — Minimalism construed as an empirical characterization of what is said_int.

 

I-Minimalism

What is said_int is affected by the bottom-up process of saturation but not by top-down processes such as free enrichment.

 

            I-Minimalism is a non-stipulative version of Minimalism, for the Minimalist constraint is not used as a defining criterion for demarcating what is said. What is said is independently demarcated, using another criterion (the availability criterion), and it is claimed that what is said_int satisfies the Minimalist constraint, as a matter of empirical fact.^[8]

            Insofar as he attacks TCP and rejects (ii), what Stanley is defending is not S-Minimalism (which needs no defense, since it is vacuously true), nor even S*-Minimalism, but I-Minimalism. It is important to realize that Stanley thus agrees with TCP on the analysandum. What he is concerned with, like the TCP-theorist, are the intuitive truth-conditions of utterances.^[9] His defense of Minimalism is therefore strikingly unlike the usual defense. The usual defense of Minimalism again TCP consists in arguing that a decent semanticist should be concerned not with 'what is said' in the intuitive sense, but with something more abstract, which satisfies the Minimalist constraint but need not surface to consciousness. That is changing the subject, and we can't help feeling that the TCP-theorist and the Stipulative Minimalist are talking at cross-purposes. But Stanley and the TCP-theorist are clearly talking about the same thing: what is said_int. Their disagreement over Minimalism is a genuine empirical disagreement.

 

VII. THE ARGUMENT FROM BINDING

 

Stanley denies that there are unarticulated constituents. His argument is disarmingly simple. It is this:

 

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)

 

Stanley then uses data of the sort originally collected by Barbara Partee (1989) to show that, in each case in which an alleged unarticulated constituent has been postulated to account for the intuitive meaning of an utterance, one can intuitively 'bind' the alleged unarticulated constituent, i.e. make it vary according to the values introduced by some operator. For instance, the temporal location of the breafkast event, which was said to be an unarticulated constituent of the speaker's response in the example from section III, can be bound by a quantifier. We can say:

 

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

 

The temporal location of the breafkast event now systematically varies with the temporal values introduced by 'each time you offer me lunch'. It follows (according to the argument) that the alleged unarticulated constituent in the original example was not really unarticulated: it had to be the (contextual) value of a variable in the logical form of the sentence, since without a variable there could not be the sort of binding that occurs in (2).

            I will discuss this argument at length in what follows (Part 4). I will also offer an analysis of examples like (2) (section XVI). But first, I would like to clarify the very notion of an (alleged) unarticulated constituent. For many, if not most, of the examples which Stanley discusses are not, by my lights, even putative unarticulated constituents, even though they are often mentioned as such in the literature. They are pragmatically determined aspects of meaning, for sure, but their contextual provision is clearly triggered by something in the sentence.

 

 

PART 3

TESTING FOR UNARTICULATEDNESS

 

 

VIII. BOGUS UNARTICULATED CONSTITUENTS

 

It is standard in the literature to characterize an unarticulated constituent as follows. It is a constituent of the truth-conditional interpretation of an utterance which is not syntactically articulated in the sentence, yet cannot be disregarded without making the utterance semantically unevaluable. The contextual provision of such a constituent is therefore semantically mandatory even if it is not triggered by some constituent in the sentence. Without such a constituent, the sentence, though syntactically complete, would not express a complete proposition.

            The prototypical example of an utterance whose interpretation involves an unarticulated constituent is the sentence 'It is raining', used to talk about a particular place (Perry 1993: 206). 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. The statement 'It is raining' explicitly gives us the two-place relation (supplied by the verb) and the temporal argument (supplied by the present tense). But no constituent of the sentence stands for a place. Still the place has to be there for a complete proposition to be expressed — indeed we understand the statement as being about the place of utterance and we evaluate it as such.^[10] In contrast to the time of utterance, which is (indexically) articulated in the sentence, the place of utterance is an unarticulated constituent of the proposition which the utterance expresses.

            We find a similar notion in Bach's writings about 'completion'. According to Bach,

 

An (indicative) sentence is semantically underdeterminate if it fails to express a complete proposition — determine a definite truth-condition — even after ambiguity and vagueness are resolved and indexical references (including the time of the utterance) are fixed. (...) In these cases what the conventional meaning of the sentence determines is only a fragment of a proposition or what I call a propositional radical; a complete proposition would be expressed only if the sentence were elaborated somehow, so as to produce a completion of the proposition. (Bach 1994: 268-269)

 

For example, if someone says that 'John is too short', what John is too short for must be identified in order to assign a truth-value to the statement. (The utterance does not express the weak proposition that John is too short for something or other.) Yet that constituent of the proposition is not syntactically articulated, as it would be if the speaker had said 'He is too short to be a British policeman'. The sentence is not even elliptical, Bach insists: it is syntactically complete as it is. Yet it is semantically incomplete,  like Perry's rain example. Without the unarticulated constituent provided through 'completion', the sentence expresses only a propositional radical. Other examples mentioned by Bach include:

 

(argumental underdetermination)

Mary finished. (what?)

The cow jumped over. (what?)

Gentlemen prefer blondes. (to what?)

Mutual knowledge is not relevant. (to what?)

John is ready. (for what?)

 

(parametric underdetermination)

That lamp is short/cheap/old. (relative to what?)

That employee is good/talented/valuable. (in what respect?)

Even cowgirls sing the blues. (in addition to whom?)

 

Many other examples often discussed in the literature on context-sensitivity (e.g. implicitly relational words like 'local', 'enemy' or 'home') also require 'completion'.

            Whenever completion is required in order to get a full-fledged proposition, I think the constituent which has to be contextually provided is definitely not an unarticulated constituent; so the above characterization of unarticulated constituents is particularly wide off the mark, according to me. (I will defer discussion of 'It's raining' until section X.) What Bach describes as 'completion' is but a special case of saturation, hence not an instance of free enrichment at all.^[11] In completion, pace Bach, the contextual provision of the relevant propositional constituent is linguistically controlled: the sentence itself sets up a slot to be contextually filled.

            Take the adjective 'small', which is an instance of 'parametric underdetermination'. It expresses a complete content only when a comparison class has been provided. That is a conventional, linguistic property of that expression. The need for completion is an integral part of the meaning of the adjective, just as it is an integral part of the meaning of the genitive. 'John's car' means 'the car that bears R to John'; no definite content is expressed unless the free variable R has been given a value. Similarly, 'small' means 'small for an x'. No definite content is expressed by a sentence such as 'Sherman is small' unless a value is provided for the variable 'x'. The contextual element (the relevant comparison class) is articulated precisely because there is an expression in the sentence, namely the adjective 'small' itself, that triggers the search for a relevant comparison class.

            That is not to say that there is no difference between expressions like 'small' and, say, the genitive. In certain respects they behave quite differently. For example, the genitive is not 'bindable', that is, the value of the variable R must always be contextually identified and cannot vary according to the values introduced by some operator in the sentence. But many expressions in need of contextual 'completion' are bindable, as Partee (1989) and Mitchell (1986) first pointed out. Thus Stanley gives the following example (to show that the adjective 'small' carries with it a free variable):

 

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. But this difference between bindable and unbindable context-sensitive expressions is irrelevant to the articulatedness issue. We find exactly the same difference between indexical expressions (in the broad sense): some — e.g. 'I', 'here', 'now' — are rigid and unbindable, others — third-person pronouns, demonstratives — are flexible and bindable. In all cases, however, when a definite value is contextually provided for the indexical, it is obviously articulated (by the indexical itself). Similarly, I think there is no doubt the contextually provided element is articulated whenever it corresponds to a particular expression and is required for its interpretation.

            From what I have just said it follows that the lexical items in need of completion in Bach's examples are very similar to indexicals and can receive a similar treatment. Consider, for instance, 'John is too short'. Arguably, 'too' maps the property of being short onto a different property: the property of being too short, that is, the property of being short to a degree n such that any degree of shortness equal or superior to n prevents a certain condition c (e.g. being a British policeman) from obtaining. (This is very rough, I admit.) The content of 'too' can therefore be construed as a certain function f_too from properties to properties. But that content is not fixed once for all; it varies contextually. For, as we have seen, the condition c must be specified or identified;^[12] otherwise the construction is semantically incomplete. Different conditions c will determine different functions f_too as contents for the adverb 'too': too __ to be true, too __ to be a British policeman, etc. 'Too' can therefore be treated as an expression with an 'unstable character' (Kaplan 1989). The character of 'too' can be represented as a function from contexts in which a certain condition c is pragmatically singled out to the functions f_too which are the content of 'too' in those contexts. In this framework the contextually provided element is as articulated in the case of 'too' as in the case of standard indexicals.^[13] Indeed 'too' is an indexical adverb, like 'thus', in this framework. (On indexical adverbs, see Heal 1997.)

            If the contextual elements provided through 'completion' are articulated, as it seems to me clear that they are, why have able authors like Bach insisted that they were not? I think what they meant — which may well be right — was that the constituents in question are not syntactically articulated. In the relevant examples, it is not a syntactic component of the sentence which acts directly as a variable whose contextual value is the propositional constituent in question, as is the case for 'he' or 'she' (or for 'thus'). Rather, a variable is involved at the sub-atomic level. Only if we analyse the meaning of a given lexical item or construction do we find the variable in question. If that is right then the contextually provided constituent may be said to remain syntactically unarticulated even though it is semantically articulated through the lexical item or construction in need of completion.^[14]

            Whether or not such a distinction between different ways of being articulated makes sense, it is, in any case, irrelevant to the unarticulatedness issue. Whether something is articulated directly (syntactically) or indirectly (semantically or subsyntactically) does not change the fact that it is articulated. If it were not — if the linguistic expression itself, qua expression type, did not trigger the search for a contextual value for the variable — it would be possible to imagine contexts in which no such contextually provided element would be necessary to get a complete proposition. In other words, if the contextually provided constituent was really unarticulated, it would not be mandatory, in the relevant sense.

 

IX. UNARTICULATED CONSTITUENTS ARE NEVER MANDATORY

 

The mandatory vs. optional status of the putative unarticulated constituent is indeed the crucial issue. To believe in semantically mandatory yet truly unarticulated constituents one must hold the following thesis: that the contextual provision of some propositional constituent may be necessary for purely metaphysical reasons, without being made necessary by the conventions of the language. But that thesis is indefensible. Whenever the contextual provision of a certain propositional constituent is necessary, I will argue, it is necessary in virtue of the conventions of the language.

            Consider Perry's 'rain' example once again. The verb 'to rain' does not take a locational complement, except optionally in the form of an adjunct. Still, Perry tells us, the relation which the verb denotes is dyadic — it comprises two argument roles. That is a metaphysical fact about that relation. For that reason, even though, syntactically, the 'place' argument-role is optional and does not have to be filled, semantically it is foregrounded and requires completion. That is so in virtue of the simple fact that that relation is denoted. The situation is the same with other cases of 'argumental underdetermination'. For example, the thing noticed is an argument of the relation denoted by 'notice'. That argument may remain syntactically unarticulated ('I noticed'), but it has to be contextually provided since the argument-role which it fills is constitutive of the relation which the verb denotes.

            This line of reasoning presupposes that the same two-place relation is expressed or denoted whether or not, syntactically, the verb takes an overt complement. But how do we know it is the same relation? In some cases at least, a case can be made in favour of treating the objectless verb as denoting 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...x_n-1 iff there is something x_n such that Px₁...x_n

 

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).

            How do we know that suppressing the object of the verb in surface syntax does not amount semantically to the same result as applying 'Der' to the original two-place predicate? In some cases it seems that it does. Thus consider the verb 'to eat'. Among the argument-roles that are constitutive of the eating relation, there is the Eater and the Food. From that relation we can extract a property, by existentially quantifying the Food argument role. Let us define the property of eating¹ as the property one has in virtue of filling the Eater argument role in some instance of the dyadic 'eat' relation (eat²). It follows that one eats¹ iff there is something that one eats². Metaphysically, both the property and the relation exist (notwithstanding the fact that the former depends upon the latter). We know that the transitive verb 'eat' denotes the relation; but how do we know whether the intransitive verb 'eat' (as in 'John eats when he is nervous') denotes the same dyadic relation, or the derived property? We can reason as follows: If the dyadic relation were expressed, the Food argument-role would require completion (since there is no operator to bind the sub-atomic variable). But no such completion is required: 'x eats' is true iff there is something y that x eats². The fact that the relevant y does not have to be specified supports the conclusion that intransitive 'eat' denotes the property of eating¹. In contrast, intransitive 'notice', as in 'I noticed', behaves differently: '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. Similarly, 'x finished' is not true iff there is something or other which x finished. The object argument of 'finish' must be specified, even if it is unarticulated in surface syntax. This gives us reason to believe that the same two-place relation is expressed by 'finish' (or by 'notice') whether or not the complement is overtly articulated.^[15]

            To sum up, we cannot assume that the intransitive variant denotes the same two-place relation as the transitive verb from which it is derived, and conclude that the suppressed argument is semantically mandatory. Rather, we must look and see whether or not the argument is mandatory, and use the empirical fact that it is (or is not) to determine what the verb denotes (the relation or the property). The empirical fact in question is a brute linguistic datum. It is a conventional property of the English verb 'eat' that (i) it can be used intransitively (in contrast to 'devour'), and (ii) when so used it does not require completion of the suppressed argument-role. Similarly, 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 completion of the suppressed argument-role. Since completion, in this sort of case, is required in virtue of a linguistic convention governing the use of a particular lexical item, the propositional constituent contextually provided through completion is articulated, at the sub-atomic level. It is not a genuine unarticulated constituent.

            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 witout being linguistically required, since the intransitive verb only denotes the property and does not require completion. In such a case the eaten object would be a genuine unarticulated constituent resulting from free enrichment. But in the above cases of 'argumental underdetermination', the contextually provided constituent is subsyntactically articulated. Its contextual provision, therefore, is a bottom-up pragmatic process — a variety of saturation — rather than a top-down pragmatic process of free enrichment.

            I conclude that what characterizes genuine unarticulated constituents 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 the unarticulated constituent 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 constituent is mandatory in the relevant sense only if in every context such a constituent has to be provided (precisely because the need for completion is not a contextual matter, but a context-independent property of the expression-type). This, then, is the criterion we must use when testing for (genuine) unarticulatedness: Can we imagine a context in which the same words are used normally, and a truth-evaluable statement is made, yet no such constituent is provided? If we can imagine such a context, then the relevant constituent is indeed unarticulated; if we cannot, it is articulated, whether syntactically or subsyntactically.

 

X. 'IT IS RAINING'

 

Perry says that the contextual provision of a place is semantically mandatory for interpreting a weather statement like 'It's raining'. If that is right, then what I called the prototypical example of an unarticulated constituent is not really unarticulated but is the value of a sub-atomic variable. That, indeed, is the view offered by Ken Taylor in defense of Minimalism (Taylor 2000).

            Taylor states his general view ('Parametric Minimalism') as follows:

 

A sentence typically sets up a semantic scaffolding which constrains, without determining, its own contextual completion. The sentence does so by containing a variety of parameters the values of which must be contextually supplied in some more or less tightly constrained way. Sometimes the  to-be-contextually-evaluated parameter is explicitly expressed in the syntax of the sentence. This is the case with explicit indexicals, demonstratives and also with verb tenses. Sometimes, however, the to-be-contextually-evaluated parameter is “suppressed” or hidden. Saying just where such parameters hide is a difficult matter — one perhaps better left to linguists than to philosophers. But I venture the hypothesis that some unexpressed parameters hide in what we might call the subsyntactic basement of suppressed verbal argument structure.

 

The subsyntactic hypothesis applies, in particular, to 'It's raining':

 

The view which I favor supposes that the verb ‘to rain’ has  a lexically specified argument place which is q-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 marks rainings as a kind of change that places undergo. (...) Thus though:

 

(3) 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 (3) 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’.

 

I think such an analysis is unavoidable once we accept Perry's claim that, to evaluate (3), we need a place. Or at least, it is unavoidable if we understand that claim as follows: for any token u of the complete sentence 'It is raining', it is necessary, in order to evaluate u, to be given a place. If the necessity concerns all tokens, it is a linguistic property of the sentence-type, hence, presumably, it arises from the internal lexical structure of the verb 'rain'.

            But must we really accept Perry's claim, thus construed? Can we not imagine a context in which 'It is raining' would be evaluable even if no particular place were contextually singled out? I have no difficulty imagining such a context. I can imagine a 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 of the triggering detector is indicated by a light on a board in the Monitoring Room. After weeks of total draught, 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 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. There is no need to posit a lexically specified argument-role for a location in the sub-atomic structure of the verb 'rain'. 'Rain' is like 'dance' and other action verbs, contrary to what Taylor emphatically claims:

 

Despite the fact that one cannot dance without dancing somewhere or other, a sentence containing ‘to dance’ can be semantically complete, even if the place where dancing happens is not contextually provided. That a dancing must take place somewhere or other is a (mutually known) metaphysical fact about the universe —  a fact that supervenes on the nature of dancing and the structure of space-time. But that metaphysical fact is not explicitly reflected in the  subsyntactic structure of the lexicon. [...] Things are quite otherwise with the verb ‘to rain’. I take the verb itself, and its subsyntactic lexical structure, to be the source of the felt need for the contextual provision of a place or range of places where a raining happens. Facts about the subsyntactic lexical structure of the verb directly entail that nothing fully propositionally determinate has been expressed by an utterance of a sentence like (3) unless a place is contextually provided.

 

As our example shows, Taylor is empirically mistaken here. He is right about 'dance' but wrong about 'rain'. That raining must take place somewhere or other is a metaphysical fact, not a linguistic fact. That fact does not prevent an utterance like (3) from expressing a fully determinate proposition even if no place is contextually provided. 'Rain' is definitely not like 'notice' or 'finish'.

            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. In such cases, therefore, the place is a genuine unarticulated constituent. When we say 'It's raining' and mean: It's raining in Paris, the location is an unarticulated constituent of the statement, just as, when we say 'Look! He is eating' and mean: He is eating the dangerous mushroom, the mushroom is an unarticulated constituent. This is very different from cases of 'completion' where, as Taylor puts it, a sub-atomic variable "makes its presence felt by 'demanding' to be assigned a contextually supplied value".

 

XI. VARIADIC FUNCTIONS

 

According to Perry, Taylor and most unarticulated constituency theorists, the location argument-role is part and parcel of the Rain relation. As Crimmins puts it, "what we know about rain makes it obvious that this relation must have as arguments at least a time and place" (1992: 17). Since the relation comprises this argument-role, whenever a location is contextually provided it fills that role and thereby finds its way into the proposition expressed by the utterance.

            I have defended a more radical conception of unarticulated constituency. According to me, an unarticulated constituent is not even lexically articulated. Thus, I have argued, the Rain relation does not involve an argument-place for a location. In this framework an obvious difficulty arises, which does not arise on the Perry-Taylor view. If the relation itself does not come with an empty slot for a location, how will the contextually provided location manage to fill a role in that relation? How will the location find its way into the proposition? How will the Rain relation and the location of rain cohere into a single proposition, if the relation does not comprise an empty slot in which the location can fit? This reminds us of the well-known problem of the unity of the proposition.

            Something — but what? — must provide the glue for holding together the constituents of the proposition. That is (or was) the problem of the unity of the proposition. Even though it once had its detractors (e.g. Ramsey), Frege's solution has won the day. According to Frege, what provides the glue is the unsaturated nature of the relation, which comprises empty slots for the arguments. This solution has been so widely accepted that the problem is no longer considered a problem and is rarely mentioned in philosophy of language textbooks. It is not in my intention to change that situation — like almost everybody else, I am happy with Frege's solution. But there is at least an apparent tension between that solution and the theory I have sketched. Unarticulated arguments are characterized by their optionality, and that means that there isn't any 'empty slot' in need of completion waiting for them. This raises a major difficulty for our account: If there is no empty slot in the relation, how will the unarticulated constituents, once provided, cohere with it in the proposition? What will solve the problem of the unity of the proposition?

            When explicit, the location of rain is typically indicated by means of modifiers such as 'in Paris', 'here', or 'everywhere I go'. Such modifiers 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'. 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 arity. Quine's 'Der' operator contributes a variadic function which decreases the arity of the input relation Rⁿ by existentially quantifying the nth argument. Another subclass of variadic functions increase the arity of the input relation by adding a new argument-role. Adding an 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. 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.

            Variadic functions which increase the arity of the input relation through the addition of a circumstance to the set of its argument roles can be represented by means of an operator (or rather, a type of operator) 'Circ' which is the converse of Quine's Derelativisation operator 'Der'. When applied to an n-place predicate P, 'Circ' produces an n+1 place predicate ('Circ P') such that

 

(Der Circ P) x₁...x_n iff Px₁...x_n

 

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 example, the operator 'Circ_location' 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:

 

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

 

            A modifier such as 'in Paris' does not merely increase the arity of the input relation by adding an new argument-role, however; it also provides the extra argument needed to fill that argument-role. 'John eats in Paris' should therefore be represented as follows:

 

Circ_{location: 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.^[16]

            Let us now go back to the problem we started with. Just as the modifier is syntactically optional, the circumstance it introduces is semantically optional with respect to the input relation. The input relation takes n arguments, and the circumstance is not one of them. The circumstance is contributed from outside, as it were; it is not demanded by the relation itself, as it would be if it filled one of the argument-roles that are constitutive of the relation. Still, the problem of the unity of the proposition does not arise. Though optional with respect to the input relation, the circumstance is not left free-floating in the proposition, with no slot to occupy. That is so because, by applying to the input relation, the variadic function generates a new relation: the output relation, which counts an extra argument-role. With respect to that relation, the circumstance is not optional. It is a genuine argument of the output relation, even if, qua circumstance, it is additional (hence not an argument) with respect to the input relation. To put it in a nutshell: there are two relations — the input relation and the output relation — hence the optionality of the circumstance with respect to the input relation is consistent with the propositional integration of the circumstance qua argument filling a slot in the output relation.

            Let us apply this general solution to the rain example. As we have seen, the Rain relation does not involve an empty slot for a location. Thus we can say 'It is raining' without providing a location for the rain, whether linguistically or even contextually. That is the lesson of the weatherman example from section X. But if we do provide a location, either through the adjunction of a prepositional phrase or by purely contextual means, we thereby generate a new relation, in which there is an empty slot, an argument role which the location fills. Since there is an argument place for the location in the output relation (though not in the input relation), the location finds its way into the proposition and coheres with the other constituents. This is consistent with the optionality of the location with respect to the input relation — the relation to which, implicitly or explicitly, the variadic function applies.^[17]

 

 

PART 4

THE ARGUMENT FROM BINDING: A REFUTATION

 

 

XII. THE BINDING CRITERION

 

I have argued that, whenever a contextually provided constituent is unarticulated, we can imagine contexts in which the lack of such a constituent would not prevent the sentence from expressing a complete proposition. This gives us a criterion — the Optionality Criterion — for determining when a contextually provided constituent is unarticulated and when it is not.

 

Optionality Criterion

Whenever a contextually provided constituent is unarticulated, we can imagine another possible context of utterance in which the contextual provision of such a constituent would not be necessary for the utterance to express a complete proposition.

 

Using that criterion, I have established that the contextually provided location (when there is one) is a genuine unarticulated constituent of the statement that it is raining, precisely because it is optional. 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.

            This result conflicts with what we get when we apply another criterion: the Binding Criterion, which Stanley uses in his paper.

 

Binding Criterion

A contextually provided constituent in the interpretation of a sentence S is articulated whenever the argument-role it fills can be intuitively 'bound', that is, whenever what fills that role can be made to vary with the values introduced by some operator prefixed to S.

 

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, representing the argument-role to be filled, then the contextually provided constituent which fills it is articulated — it is the (contextual) value of the variable.

            The conflict arises because the location of rain, which is a genuine unarticulated constituent in virtue of the Optionality Criterion, turns out to be articulated in virtue of the Binding Criterion. We can say:

 

(4) Wherever I go, it rains

 

In such a statement the place where it rains can be undertood as varying with the places introduced by the quantifier 'wherever I go'. On the most natural interpretation 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 (4)), or be contextually given a value (as in (3)). Whatever location may be contextually provided for the rain is therefore not a genuine unarticulated constituent, but the contextual value of a free variable in logical form.

            The major problem I see with Stanley's argument against unarticulated constituents is that it works too well. It obliges us to treat as articulated not only contextual elements which can plausibly be regarded as values of variables in sub-atomic structure, 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 us start with an unproblematic case, in which the Binding Criterion works smoothly. We can, as Partee (1989) pointed out, bind the 'object' argument role of 'notice' even though it is unarticulated in surface syntax:

 

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

 

Applying the Binding Criterion, we conclude that the suppressed argument-role is still represented in logical form by a free variable. It follows that, when we say 'I noticed', the contextually provided object is articulated, appearances notwithstanding. This is OK, for there are reasons to believe that the object of 'notice' is indeed represented by a free variable in sub-atomic structure. But the Binding Criterion delivers the same verdict in indefinitely many cases in which the unarticulated nature of the constituent seems pretty well established. Let me give two striking examples.

            The first example is adapted from David Rumelhart (1979: 78):

 

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 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 unarticulated. 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. For we can say things like

 

However he did it, the policeman stopped the car

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 is unarticulated and results from free enrichment follows from the fact that intransitive 'eat', as Stanley himself accepts (p. 401, fn. 14), denotes the property of eating¹. No contextual specification of the thing eaten is required in virtue of the semantics of the verb. Still, intuitively, binding is possible. We can say:

 

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 articulatedness. The Binding Criterion, on which Stanley's argument rests, must be rejected.

 

XIII. EXISTENTIAL CLOSURE BY DEFAULT?

 

We have seen that, sometimes, a contextually provided constituent is optional, even though the argument-role it fills is bindable. Its being optional suggests that there is no free variable in logical form; for if there were one, completion would be required whenever the variable is left unbound. For example, there are contexts in which the simple sentence 'It is raining' (with no operator in front to bind the alleged variable) is interpreted as It's raining in some place or other. If there were a free variable in logical form, completion would be required and a definite location would have to be assigned to the variable in the course of truth-conditional interpretation. That fact argue against the Binding Criterion, according to which bindability entails articulatedness.

            In defense of the Binding Criterion, one could argue as follows. Who said that, whenever there is a free variable, completion was required? Following Partee and others (e.g. Nunberg 1992), we have noticed that there were two sorts of 'free variables' associated with indexicals and other context-sensitive expressions: those that are bindable and those that are not. The indexical 'here' acts as a free variable for a location (a location that has to be 'proximal', even if it need not be the place where the speaker is); but that variable is unbindable — its value can only be contextually provided. In contrast, 'there' also acts as a free variable for a location, but that variable is bindable. (I can say: 'Whenever I go and visit my father's new home, my brother is already there', on a bound reading in which my father's home keeps changing. I cannot say to my father: 'Whenever I come and visit you in your new home, my brother is already here', on a similarly 'bound' reading of 'here'.) However we account for this distinction, it is an empirical fact that there are these two sorts of variable. Now, why not accept that there is another distinction, among bindable variables, between those which require completion when unbound, and those which do not? On this view, there are free variables in logical form which can be bound (as the location variable can, in 'It is raining') but which do not require completion when left unbound.^[18] Such a free variable — let us call it an optional variable — can either be assigned a contextual value, or, if no value is contextually provided, undergo ECBD (existential closure by default). On this view it is the Optionality Criterion which must be given up, not the Binding Criterion. The fact that completion is optional in the alleged counterexamples to the Binding Criterion does not entail that the contextually provided constituent is unarticulated; all that fact shows is that, if articulated, the constituent in question is articulated by an optional variable, i.e. a variable susceptible to ECBD.

            On the view I have just sketched, all the cases I have classified so far as genuine unarticulated constituents would turn out to be constituents articulated by an optional variable. For example, intransitive 'eat' would not longer be said to denote the property of eating¹, but the relation of eating². That relation, the ECBD theorist would argue, involves an argument-role for the Food, represented by a free variable in logical form. But that variable belongs to the optional variety. 'He is eating' can thus be understood either with respect to a contextually provided object (e.g. the mushroom in the 'Look! He is eating' example) or via ECBD as saying that the individual referred to by 'he' is eating something or other. The Food argument-role can also be bound by an explicit quantifier, as we have seen ('Whenever his father cooks mushrooms, John eats'). Even the Rumelhart example could be handled by saying, in the manner of Taylor, that all verbs of action carry a 'manner' variable that is optional, that is, may or may not be contextually given a specific value when unbound.

            For that option to be worth pursuing, we must give it some initial plausibility by showing that there are clear cases of the phenomenon on the articulated side. In other words, we must find context-sensitive expressions which can not only be bound by an operator, but which can also be implicitly bound via ECBD. That is not such an easy thing to do. The third person pronoun 'he', which can be bound, cannot be bound implicitly via ECBD. There is no reading of 'He is bald' where that means that some male or other is bald. Nor can 'John is home' mean that John is at someone or other's home. (Yet 'home' can easily be bound by e.g. a quantifier: 'Everybody went home'.) Binding by default does not seem to occur with overt context-sensitive expressions. Why, then, should it occur with covert context-sensitive expressions? That is one of the questions which have to be answered by anyone willing to explore the ECBD option.

            As against the ECBD option, one may try to prove that the Binding Criterion cannot be salvaged in this or any other manner. To prove that, one might exhibit a method for indefinitely multiplying the sort of constituent to which the Binding Criterion says there corresponds a variable in logical form. Since the number of variables in the logical form of a sentence must be kept finite, the existence of such a method would refute the Binding Criterion.

            I will not say more about the ECBD option in this paper. Whatever attraction it has is due mainly to the support it receives from Stanley's argument. According to the argument, there cannot be binding without a bindable variable, hence the bindability of an argument-role shows that the contextually provided constituent filling that role is articulated by a free variable in logical form. If one finds Stanley's argument convincing, one is thereby led to accept the package Binding Criterion plus ECBD (since ECBD is the only way to account for the optionality of the relevant constituents consistently with the Binding Criterion). But I do not find Stanley's argument convincing — I think it rests on a fallacy, which I am now going to expose.

 

XIV. 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 articulated after all. The sentence 'it rains' really is the sentence 'It rains in 1', where the unpronounced location variable can either be bound or be contextually assigned a value. Fully spelled out, Stanley's argument against unarticulated constituents runs as follows:

 

1. Unarticulated constituent theorists say that in the simple statement 'It rains', the location of rain is unarticulated.

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 unarticulated constituent theorist 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 restricted 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 is said by some to involve an unarticulated location constituent).

 

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 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 arity 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 arity-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 clearly distinguished.

            At the beginning of his paper Stanley acknowledges the "effects of linguistic context on interpretation" and mentions type-shifting phenomena in this connection (Stanley 2000: 391 fn.). It is something similar that we have here. We start with a simple sentence, in which a certain relation R is said to hold of a sequence of n arguments; adjoining a prepositional phrase to that sentence modifies the relation by increasing the number of its argument-roles. Hence the fact that the relation expressed by the predicate in the complex sentence takes n+1 arguments cannot be used to infer that in the simple sentence also, the relation was n+1-ary; for the predicate has been modified in the transition from the simple to the complex sentence. It is this simple fact which Stanley overlooks, and which is responsible for the failure of his argument.

 

 

PART 5

'BOUND' UNARTICULATED CONSTITUENTS

 

 

XV. THE WEAK BINDING CRITERION

 

When a circumstance is explicitly provided, by means of an adverbial adjunct, a variadic function operates on the input relation expressed by the nuclear predicate, outputting a complex relation with higher arity. The same thing happens when a circumstance is contextually provided, except that the variadic function which applies to the input relation remains unarticulated. It is important to realize that (on the present account) what is unarticulated is not merely the argument which fills the new argument-role (say, Location) but the argument-role as well. The unarticulated constituent corresponds to what would be articulated if a prepositional phrase had been uttered. So the proper analysis of 'it's raining', when the location Paris is an unarticulated constituent of the interpretation, is

 

<In Paris> (it rains)

 

where the prepositional phrase 'in Paris' stands for a locative operator CIRC_{location: Paris}, and the angle brackets indicate that the operator in question remains unarticulated.

            Beside the articulated/unarticulated distinction, there is another important distinction between cases in which the contextually or explicitly provided circumstance is constant and cases in which it is variable and depends upon the values introduced by some quantifier in the sentence. On the articulated side, that is the difference between: 'In Paris it is raining' (constant location) and 'In every city I know it is raining' (variable location). The latter I represent as

 

(For every city x such that I know x) (in x (it rains))

 

Here, as we saw in the previous section, the quantified prepositional phrase (QPP) contributes both the variadic function 'in x' which creates an extra argument-role and the quantifier 'for every city x such that I know x' which binds it.

            In this particular example the variable location is articulated by the QPP. Could it be unarticulated? That is, is it possible for a circumstance to be both variable and unarticulated at the same time? Stanley claims that it is not. To say that a circumstance filling a certain argument-role varies with the values introduced by some quantifier in the sentence is to say that there is a free variable, representing the argument-role, which is bound by that quantifier. Now if there is a variable in logical form representing the argument-role, then the argument-role is articulated. So the reasoning goes. The natural conclusion is that only constant circumstances can be unarticulated.

            The piece of reasoning I have just reconstructed is different from the 'argument from binding' discussed (and refuted) in Part IV; for that argument was meant to show that there are no unarticulated circumstances at all, whether constant or variable. The argument from binding discussed in Part IV was offered in justification of the Binding Criterion, according to which bindability — not actual binding — entails unarticulatedness:

 

Binding Criterion

If an argument-role is bindable, that is, if what fills it can be made to vary by prefixing a quantifier to the sentence, then that argument-role is articulated even in the simple sentence (without the quantifier).

 

Stanley used that criterion in attempting to establish that, in examples like 'It rains', the contextually provided location is not a genuine unarticulated constituent, but results from contextually assigning a value to a free variable. (Indeed, the location can be made to vary by prefixing the quantifier 'everywhere I go' to the sentence.) That attempt fails, as we have seen; and it fails because the Binding Criterion 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 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. It follows that bindability does not entail articulatedness; only actual binding entails articulatedness. Hence the Binding Criterion must be rejected, and a weaker criterion adopted instead:

 

Weak Binding Criterion:

If an argument-role is actually bound by a quantifier in the sentence, that is, if what fills the role depends upon the values introduced by that quantifier, then (in that sentence) the argument-role is articulated by a variable which the quantifier binds.

 

That Criterion establishes that variable circumstances cannot be unarticulated. But it says nothing about constant circumstances — the sort of circumstance that features in standard examples of unarticulated constituent.

            Contrary to the Binding Criterion, the Weak Binding Criterion does not rest on a fallacy. Yet it, too, conflicts with the Optionality Criterion. This we can see by looking at some of Stanley's examples in which, he claims, binding actually occurs. Many of these examples are such that the circumstance they involve, though variable, has to be classified as unarticulated in virtue of the Optionality Criterion.

 

XVI. INDIRECT BINDING VIA UNARTICULATED FUNCTIONS

 

As Stanley points out, we can say

 

(5) Every time John lights a cigarette, it rains

 

meaning:

 

For every time t at which John lights a cigarette, it rains at t at the location in which John lights a cigarette at t

 

On this 'bound' reading the location of rain varies with the places in which the event of John's lighting of a cigarette occurs. To capture that reading, as well as the ordinary reading of 'It rains' (understood with respect to a contextually provided location), Stanley says we must postulate a variable in logical form, which can be either bound or free. With an unarticulated constituent analysis, he claims, there is no way to account for the bound reading. Is that true? Let us try.

            Instead of saying that there is a variable in the logical form of the sentence, which can either be bound (as in 5) or be assigned a contextual value (as in (3)), we can say that in the interpretation of both (5) and (3) there is an unarticulated constituent. In (3) the unarticulated constituent is a specific location provided by the context. In (5) the unarticulated constituent is a function from times (or from events) to locations — a function which determines a location only with respect to a time (or possibly an event) serving as argument to the function. The difference between the ordinary interpretation of 'it rains' in (3) and its interpretation in sentence (5) can be represented as follows:

 

Interpretation of 'it rains' in (3)

<here/in Paris> (it rains)

 

Interpretation of 'it rains' in (5):

<in location f(t) > (it rains)

 

            That the the variable location is unarticulated in (5) can be established by means of the Optionality Criterion. Just as the simple sentence 'It is raining' can be interpreted as meaning It is raining in some place or other, without any location being contextually provided, the complex sentence (5) can be understood as meaning Whenever John lights a cigarette, it rains in some place or other. Neither a specific location nor even a function from times or anything else to locations needs to be provided. When either is provided, it is for pragmatic reasons — in order to make sense of the utterance. Nothing in the sentence itself triggers the contextual provision of either a specific location or a function taking locations as values.

            It is surprising that Stanley did not think of this theoretical option, since he himself considers that the variable occurring in the logical form of the sentence 'It's raining' is not simply a locational variable, but a higher-level function variable f(x) taking a first-order variable as argument. In some contexts, Stanley says, 'x' will be given a location (say, Paris) as value and 'f' the identity function as value. Under those contextual assignments, 'It is raining' will express the proposition that it is raining in Paris. In a different context, 'x' will be an event, say today's concert, and 'f' a function from events to the places where the events occur. In such a context 'It rains' says that it rains at the place of the concert. Since he accepts that a function is contextually provided in such cases and assigned to the higher-level variable, why not consider the possibility that such a function can also be contextually provided without being triggered by a higher-level function variable in logical form? Why not indeed? Once this option is considered, it is pretty obvious that it is to be preferred to Stanley's own view, according to which the contextually provided function is articulated by a higher-level variable. For Stanley's view cannot account for the optional character of the contextual provision, while the alternative view can.

            Note that the function which bridges the gap between the quantifier and the indirectly bound argument-role (the 'bridging function', for short) need not be unarticulated; it can also be articulated:

 

(6) Every time I sing, it rains at the very place where I happen to sing.

 

Here there is a prepositional phrase which does not provide a specific location, but a variable location which is a function of the time (or event) of singing. This suggests that 'Everytime I sing, it rains' can itself be analysed along the following lines:

 

(For every t such that I sing at t) (at t (<in il: I sing in l at t> (it rains)))

 

The description within angle brackets provides a function from times to locations, through which the quantifier over times indirectly binds the location argument-role.

            In this framework examples like (5) involve not one, but two unarticulated constituents. First, the variadic function which introduces an extra argument-role for a location l is as unarticulated in this example as it is in 'It is raining' when understood with respect to a contextually provided constant location. True, the QPP 'Everytime John lights a cigarette' explicitly contributes (hence articulates) a variadic function, but it is a different one — one that introduces an argument place for a temporal circumstance. There are two variadic functions at work in (5), one locational and the other temporal ('at t'). The temporal one is articulated, the locational one unarticulated. Second, what fills the extra argument-role introduced by the unarticulated variadic function is a variable location determined by another unarticulated function, from times to locations.

            So far, so good. But there is a difficulty in store for the unarticulated constituent theorist (i.e. for myself). It is apparent in the formulas I have used to represent the unarticulated constituent. I have written '<in location f(t)>', where the angle brackets indicate the unarticulated nature of the constituent. Within the angle brackets we find the variable 't', which is bound by the quantifier outside the angle brackets. This is necessary to get the desired, bound reading. But how can a bound variable figure within an unarticulated constituent? A variable has got to be articulated. There is no sense in talking of 'unarticulated variables', for variables are linguistic expressions; they belong to the language, not to the reality to which the language corresponds. They are not like objects, functions, and the like, which are part of the world and to which semantic conventions map the expressions of the language. Worldly entities can be articulated or not, depending on whether or not they are contributed by linguistic expressions. But for linguistic expressions themselves, there is no possibility of unarticulatedness. This, then, is the real problem which 'bound' readings raise for the unarticulated constituent theorist: In order to capture them, it seems that we must posit a bound variable; but doing so is inconsistent with the presumptive unarticulatedness of the constituent.

 

XVII. 'WH+EVER'-PHRASES

 

The problem I have just mentioned arises in all the cases in which a variable circumstance seems to be unarticulated (i.e. when it passes the Optionality test). In section XVI we considered examples in which an unarticulated function bridges the gap between the quantifier and the unarticulated argument-role which it indirectly binds. In this section I will consider simpler examples in which no unarticulated bridging function is involved — examples such as:

 

(4) Wherever I go, it rains.

 

Here the phrase 'wherever I go' directly quantifies over locations, hence no unarticulated function is needed to bridge the gap between the quantifier and the extra argument-role for a location. Still, I want to maintain, that argument-role is unarticulated in (4). This can be established by means of the Optionality Criterion.

            I take it as obvious that, in an appropriate context, we could interpret sentence (4) as saying that

 

(7) Wherever I go, it rains in some place or other

 

I leave it as an exercise to the reader to imagine a context in which this interpretation of (4) would be appropriate. What matters is that this interpretation is not forbidden on purely linguistic grounds. In this respect there is a big difference between (4) and

 

(8) Everywhere I go, it rains

 

for that sentence cannot be interpreted as saying that everywhere I go, it rains in some place or other. That interpretation of sentence (8) is simply not available — indeed we cannot even make sense of it. In (8), the location of rain is linguistically determined — it has got to be the place where I go.

            At this point it must be noticed that (4) is actually ambiguous. Syntactically, (4) can be given two different analyses. On one analysis (4) is more or less equivalent to (8), and it does not tolerate the 'neutral' ('in some place or other') interpretation. That reading on which (4) is equivalent to (8) can be forced by suppressing 'ever':

 

(4a) Where I go, it rains

 

This can only be understood as meaning that it rains at the place where I go. On that reading (4) presumably results from a process of extraction with a trace left behind:

 

[Wherever I go]_i [it rains t_i]

 

The proper semantic analysis of (4), on that reading, is

 

For every l such that I go to l, it rains in l

 

But there is another reading of (4) — that which I want to focus on. On that reading the sentence-initial 'wherever I go' is base-generated in that position; it does not result from a process of extraction. There is no syntactic variable for a location in that case — no trace left by the movement of the wh-phrase (since no movement takes place). Syntactically, the sentence is to be analysed simply as:

 

[Wherever I go] [it rains]

 

Semantically, the proper analysis is

 

For all l, if I go to l, it rains

 

What characterizes this reading of (4) is that the location of rain is not linguistically encoded. As a result, 'wherever I go, it rains' (on that reading) can be understood as meaning: 'wherever I go, it rains in some place or other', in some appropriate context.

            Since, on this reading (henceforth the 'base-generated' reading), the sentence itself says nothing of the location of rain, it is always possible to add a prepositional phrase contributing that piece of information. Thus we can say:

 

(4b) Wherever I go, it rains here/in Paris/in all the alternative places I might have visited instead of that which I eventually chose

 

The contribution which the added locative phrase, e.g. 'here', makes in (4b) can also result from a process of free enrichment. Just as we can understand 'it's raining' to mean 'It's raining <here>', we can understand 'Wherever I go, it rains' (on the base-generated reading) to mean:

 

Wherever I go, it rains <here>

 

In such cases the contextually provided circumstance is unarticulated.

            A special case is that in which the base-generated reading of (4) is enriched with an unarticulated location which coincides with the location mentioned in the 'wherever'-clause:

 

Wherever I go, it rains <there/in the place where I go>

 

This is certainly the most natural — the most salient — interpretation for sentence (4) on the base-generating reading. The speaker who says 'wherever I go, it rains' is likely to mean that it rains in the place where he goes (whatever it is). Truth-conditionally, this is equivalent to the first reading of (4): that in which there is extraction and a trace left behind. We analysed the first reading as

 

For all l such that I go to l, it rains in l

 

What we have just seen is that the same truth-conditions can be determined by giving (4) the second, base-generated analysis and adding an unarticulated constituent:

 

For all l, if I go to l, it rains <there/where I go>

 

            A similar example, a variant of which I mentioned earlier, is

 

Whatever his father cooks, John eats

 

which is also syntactically ambigous. One one reading 'Whatever his father cooks' is base-generated in the sentence initial position, and 'eat' is intransitive:

 

[Whatever his father cooks] [John eats]

 

On the other reading the verb is transitive and there is extraction of the object out of the VP, with a trace left behind:

 

[Whatever his father cooks]_(i) [John eats t_(i)]

 

When there is no trace, an object filling the Food argument-role can still be contextually provided, but then it is unarticulated. Thus 'whatever his father cooks, John eats' (with intransitive 'eat') can be understood, inter alia, as meaning that he eats the things cooked by his father. That interpretation results from a process of free enrichment: an unarticulated variadic function maps the property of eating¹, expressed by intransitive 'eat', on the relation of eating² and the extra argument role thus provided is filled by the values introduced by the quantifier.

            In such cases the Optionality Criterion converges with syntactic analysis to suggest that there is an unarticulated constituent in the interpretation of the utterance. Yet, intuitively, binding occurs: the thing eaten covaries with the thing cooked, or the location of the rain covaries with the destination of the travel. To capture that covariation it seems that we must posit a bound variable within the unarticulated constituent. This raises a problem, as we have seen.

 

XVIII. TO BIND OR NOT TO BIND

 

In the two sorts of case I have discussed (those which involve bridging functions, and those which involve base-generated 'wh+ever'-phrases in sentence-initial position) there is, I think, enough evidence that some constituent in the interpretation of the utterance is unarticulated; yet the unarticulated constituent in question is variable rather than constant — it varies with the values introduced by a quantifier in the sentence. Stanley is right that there is a tension between the two sides of the phenomenon — variability and unarticulatedness — but I don't think the proper thing to do is to give up the claim that the relevant constituent is unarticulated. There are, I think, other ways to solve the problem of 'bound' unarticulated constituents.

            First, we may grant that the alleged unarticulated constituent, in our examples, involves a bound variable, and that bound variables cannot be 'unarticulated'. This does not necessarily prevent the constituent itself, or some sub-constituent of it, from being unarticulated.

            In the first type of example we discussed (section XVI), both the variadic function and the bridging function are unarticulated; only the values serving as arguments to the bridging fuction are articulated by the bound variable. In the other type of example (section XVII), the variadic function is, again, unarticulated; only the variable location filling the argument-role is articulated by the bound variable. So there are unarticulated as well as articulated elements in those cases. But they are not quite on a par, and we need not construe the articulated bit as an integral part of the alleged unarticulated constituent. In all the examples I have given, even if we suppress the unarticulated constituent by giving the neutral (e.g. 'in some place or other') interpretation, the bound variable is still there. So perhaps we can say that in those examples an unarticulated constituent takes advantage of an independently given bound variable. In the first type of example (e.g. (5)), a time variable is bound by the quantifier, and two unarticulated constituents jointly operate to take advantage of it: one (the variadic function) creates an unarticulated argument-role for a location, the other (the bridging function) provides a filler for that role by mapping the times in the domain of the quantifier to locations. In the second type of case (example (4), on the base-generated reading), a universal quantifier over locations has a conditional in its scope, according to the following schema:

 

("l) (g(l) —> p)

 

The antecedent of the conditional contains a variable bound by the quantifier. An unarticulated constituent (a variadic function) in the consequent takes advantage of that variable to create an extra argument role which can be filled by the values of the variable:

 

("l) (g(l) —> p <in l>)

 

            I am not saying that this is crystal-clear — it isn't and I wish I could understand the suggested interaction of unarticulated constituents and bound variables. But this is a proper object for empirical study, not a reason for despair.

            Another possible solution for the unarticulated constituent theorist confronted with the problematic cases consists in denying that a bound variable is really involved. There are different ways to support such denial, and Stanley mentions one of them: one may (i) ascribe semantic elements to bound variables as their worldly counterparts, and (ii) say that those semantic elements can be contextually supplied even if, syntactically, there is no bound variable which articulates them. Stanley credits Jeff King for emphasizing that option out to him. He rejects it on the following grounds:

 

An entity such as the denotation of a bound variable is a theoretical posit, part of the machinery of a particularly complex semantic theory. It is not something about which we have beliefs or intentions. They are therefore not supplied by pragmatic mechanisms. (Stanley 2000: 414)

 

Now I have no idea whether the suggestion that bound variables contribute special elements can be fruitfully elaborated, as some theorists believe; but I find the argument Stanley uses to dismiss that option unconvincing. The complexity and abstractness of the theorist's metalanguage, and the lack of mastery of that metalanguage on the part of ordinary users of the language, do not by themselves show that the theories couched in that metalanguage fail to capture the competence of ordinary users of the language, whether that competence is syntactic, semantic or pragmatic. The notion of a function is also a theoretical posit, part of the machinery of a complex theory. It is not something about which ordinary language users have beliefs or intentions. This does not prevent functions from being pragmatically supplied in the course of interpreting certain utterances, as Stanley himself acknowledges.

            Be that as it may, one can also support one's denial that bound variables occur in our examples by showing that they are strikingly similar to examples involving so-called 'unbound anaphora'. In 'Every villager owns a donkey and beats it', 'it' does not function as a bound variable, even though 'binding occurs' in the intuitive sense in which I have used that phrase so far: which donkey is beaten varies with the values introduced by the quantifier 'every villager'. True, there are accounts of such examples (e.g. Cooper's) which do posit a bound variable (though not one that directly corresponds to the pronoun). But which analysis of unbound anaphora is ultimately correct is still an open issue. I personally favour another approach, inspired from Evans's work but incorporating features from the 'pragmatic theory' criticized by him. If this approach is viable, there is a way of handling unbound anaphora without positing bound variables, which naturally extends to 'bound' unarticulated constituents.

            According to Evans, the above 'donkey' sentence says that, for every villager x, x owns a donkey and beats it, where 'it' refers to the donkey x owns. The pronoun refers, but it does so at the level of substitution instances: in every substitution instance of the universally quantified sentence, 'it' refers to what fills the y argument-role delineated by the predicate 'x owns y & y is a donkey', i.e. to what satisfies that predicate (when 'x' is assigned a particular villager). This opens the way for a theory incorporating the following features:

 

(i) Standard anaphoric uses (e.g. : 'I saw Paul the other day. He told me...') and E-type uses (as in the above example) are unified. There is a unique rule for all anaphoric uses of pronouns, in the broad sense: when anaphorically used, a pronoun refers back to a linguistic antecedent which is an argument-role articulated by a singular term position in the previous discourse. If what fills the position is indeed a singular term, the anaphoric pronoun inherits the reference of the singular term. If what fills the position is not a singular term but a quantifier or an indefinite description ('a donkey'), the pronoun refers to what, if anything, fills the argument-role, i.e. to what, if anything, satisfies the predicate containing the argument-place.

 

(ii) Anaphoric uses, in the broad sense just introduced, are considered as a subclass of referential uses, distinct from deictic uses. Just as, in deictic cases, a demonstrated object in the context of utterance can serve as index through which the ultimate referent is designated (as in deferred ostension, analysed à la Nunberg),^[19] anaphoric uses can be treated as referential cases in which the speaker demonstrates a linguistic index, namely some argument-role articulated in the previous discourse, and designates the ultimate referent (determined by the anaphoric rule) via that index.^[20]

 

This is not the place to elaborate that sort of account. But suppose one such account were available. Then we could say the following: The speaker who uses an E-type pronoun 'demonstrates' a certain argument-role, just as the speaker who uses a true demonstrative while pointing demonstrates something in the perceptual environment. And there is no reason why such demonstration could not remain linguistically unarticulated.

            On this view,

 

(9) Wherever I go, it rains there

 

is to be analysed as

 

(For every l) (I go to l --> it rains there)

 

where 'there' is an anaphoric pronoun (or 'pro-PP') linked to the location argument-role articulated in the clause 'I go to l': it demonstrates that argument-role and, in every substitution instance of the quantified sentence, refers (via the anaphoric rule, or a variant of it for pro-PPs) to whatever fills it. To be sure, the 'reference' thus achieved is not like ordinary reference. As Evans emphasizes,

 

It is the [anaphoric] rule which determines the reference of the pronoun in the relevant substitution instances; their reference cannot be determined by pragmatic factors, since pragmatic factors cannot determine the reference of a pronoun in a sentence [the substitution instance] whose interpretation we are considering independently of any context of utterance. (Evans 1980: 237)

 

However, once we split the relation of reference in two, according to Nunberg's analysis, we see that the speaker performs an act of demonstration aimed at an argument-role articulated in previous discourse. If the reference of the E-type pronoun at the level of substitution instances cannot be determined pragmatically but must be determined by a rule, as Evans insists, still the demonstratum (the index) is determined pragmatically, just as deictic reference is. Pragmatic factors determine both that the pronoun is used anaphorically (rather than deictically) and what the linguistic index is through which reference will ultimately be determined.

            What is thus achieved by partly pragmatic means can also be achieved in a purely pragmatic manner, through mechanisms of intention-recognition which do not require linguistic articulation. Instead of (9), the speaker can utter (4):

 

(4) Wherever I go, it rains

 

The speaker now tacitly demonstrates the location argument-role articulated in the first clause and, by making his intention to refer to whatever fills the role sufficiently manifest, manages to express the same content as if the pro-PP 'there' had been used in the second clause. By appealing to that sort of theory, we can perhaps (hope to be someday in a position to) account for 'bound' unarticulated constituents without having to posit bound variables in the analysis of the sentence.

 

XIX. FREE ENRICHMENT: THE 'SYNTACTIC' INTERPRETATION

 

Suppose nothing works and 'bound' unarticulated constituents turn out really to involve bound variables and to be articulated by those variables. What follows? Not much, I think. Though articulated in a certain sense, such constituents can still said to be unarticulated in another sense and to result from 'free enrichment'.

            So far free enrichment has been construed as a mechanism through which some constituent in the truth-conditional interpretation of an utterance is contextually supplied without being articulated in the sentence. This is the 'semantic' interpretation of free enrichment — that which I favour. But there is another, 'syntactic' interpretation. According to the syntactic conception, what free enrichment yields is not (some aspect of) the truth-conditional interpretation of an utterance, but rather a more elaborate representation which will eventually be given a truth-conditional interpretation.

            Let a 'representation' be a sequence of symbols in some linguistic/representational medium, and the 'interpretation' of a representation be some worldly entity or complex of entities to which the representation corresponds — which it represents. Then, on the semantic conception, free enrichment determines aspects of the interpretation of a given representation, aspects which are unarticulated in the sense that nothing in the representation corresponds to them. On the syntactic conception, free enrichment determines aspects of the representation which is interpreted: it contributes further symbols, further representational elements, which are unarticulated in the sense that nothing corresponds to them in the natural language sentence that has been uttered. The output of this process of free enrichment in the syntactic sense is a mental representation which articulates what the speaker means by his utterance, including those aspects of the speaker's message that are not articulated in the natural language sentence she uses.

            On the syntactic conception, free enrichment is still free: nothing in the natural language sentence triggers that process, which takes place as part of an attempt to make sense of the utterance. But what the process delivers is unarticulated only in the sense that nothing in the natural language sentence encodes that element. It is unarticulated in the sense of not being articulated in the natural language sentence. Still the element in question may be linguistic: it is a constituent in a (mental) representation, not a constituent in a state of affairs represented by a representation. Since the unarticulated constituent can be linguistic, it can be a bound variable. On the syntactic conception, therefore, the phrase 'bound unarticulated constituent' is no longer an oxymoron requiring the use of scare quotes. Bound variables can be truly unarticulated if free enrichment is construed syntactically.

            On the syntactic conception, the interpretation of an utterance proceeds in three steps. First, the linguistic module delivers a syntactico-semantic representation determined by sentence grammar irrespective of all pragmatic considerations. That is the Logical Form of the sentence. Second, the Logical Form is mapped to a more elaborate representation by various processes, including pragmatic processes such as free enrichment. The resulting representation is what bears truth-conditions. The assignment of truth-conditions to that representation is the third step in the process.

            The picture I have just sketched is very widespread (though not among semanticists), and it has a good deal of cognitive plausibility. Chomsky himself defends that picture. He writes:

 

I will understand LF to incorporate whatever features of sentence structure (1) enter directly into the semantic interpretation of sentences, and (2) are strictly determined by properties of sentence grammar. The extension of this concept remains to be determined. Assume further that there is a system of rules that associates logical form and the product of other cognitive faculties with another system of representation SR (read 'semantic representation'). Representations in SR, which may involve beliefs, expectations and so on, in addition to properties of LF determined by grammatical rule, should suffice to determine role in inference, conditions of appropriate use etc. (Some would argue that LF alone should suffice, but I leave that as an open empirical question.) (Chomsky 1976: 305-306; quoted in Evans 1980:  231)

 

It is, as Chomsky points out, reasonable to keep LF and SR distinct in principle, in order not to beg the question at issue, even if one wants to consider the possibility that LF-representations (the outputs of the linguistic 'module') and SR-representations (those which undergo truth-conditional interpretation) are actually identical.

            An overwhelming majority of TCP-theorists defend that picture too. That is the case, in particular, of those whom Stanley attacks in his paper: Sperber and Wilson, Carston, Stainton, and Bach. They explicitly say that pragmatic processes map the representations delivered by the linguistic module to the typically richer mental representations involved in linguistic communication.^[21] Stanley offers his arguments against unarticulated constituents in the course of criticizing the views of these researchers, but if one adopts the above picture, the fact that unarticulated constituents involve bound variables no longer raises a problem. Again, the variables in question will be found in the representations delivered by pragmatic processes as they apply to natural language structures; they will not be found in the natural language structures themselves, hence they will not be articulated in the relevant sense.

            In the second section of his paper, Stanley discusses cases of 'nonsentential assertion' such as the following. Suppose Bill walks into a room in which a woman in the corner is attracting an undue amount of attention. Turning quizzically to John, he arches his eyebrows and gestures toward the woman. John replies:

 

(10) A world famous topologist.

 

Stanley says this is a case of ellipsis, similar to Sarah's utterance of (11)

 

(11) Bill

 

in reply to the question: 'Who bought the bottle?'. What is special with (10), Stanley says, is only the fact that the linguistic antecedent is contextually provided (instead of being explicitly articulated). "Explicitly providing a linguistic antecedent by mentioning it is only the simplest way to provide it. There are other method of raising linguistic expressions to salience in a conversation without explicitly using them" (Stanley 2000: 404). In particular:

 

It is plausible to suppose that extra-linguistic context, such as Bill's gesture, and his quizzical glance at John, gave rise to the implicit question

 

Who is she?

 

John's utterance of (10) is then elliptical for "she is a world famous topologist" for the very same rezason that Sarah's utterance of (11) is elliptical for

 

Bill bought the bottle.

(Stanley 2000: 406)

 

In this sort of case Stanley appeals to implicit representations triggered by extralinguistic context to show that alleged unarticulated constituents are actually articulated (in the implicit representations in question). But the crucial question concerns the status of those implicit representations. According to the TCP-theorists whom Stanley is challenging, such representations are mental representations. To say that they are 'implicit' is to acknowledge that they are not overtly expressed in actual discourse. Hence the notion of ellipsis, strictly (i.e. syntactically) understood, does not apply; for syntactic ellipsis requires an explicit antecedent.^[22] Since the syntactic notion of ellipsis does not apply, (10) is best treated as a noun-phrase, not a complete sentence (Stainton forthcoming). Still, in the context of utterance it elicits a complete thought, precisely in virtue of those pragmatic processes of enrichment invoked by TCP theorists.

            On the syntactic conception, the representations which undergo truth-conditional interpretation are not directly the output of the linguistic module, but more elaborate, conceptual representations shaped by pragmatic processes and sensitive to "beliefs, expectations and so on" (Chomsky's words). At some point Stanley criticizes similar views (involving distinct representational layers) put forward by Partee and by Culicover and Jackendoff. According to them it is Discourse Representation Structures (Partee) or Conceptual Structures (Culicover and Jackendoff) that are given truth-conditional interpretations, rather than the syntactic structures of the natural language expressions which are mapped onto them by some preliminary process of interpretation. Criticizing those views, Stanley writes:

 

This picture of interpretation is prima facie difficult to accept. According to it, the interpretative process involves the production of an interpretively superfluous level of representation, namely the output of the syntactic mechanism. We would need a massive amount of empirical and methodological motivation to justify the added complexity such an interpretive process involves over straightforwardly applying a semantic interpretation to the output of our best syntactic theory. (Stanley 2000: 428)

 

This can be turned into a critique of the syntactic picture of free enrichment, so let me offer a brief response in closing this article.

            First, on the syntactic conception the output of the syntactic mechanism is not superfluous; it is the properly linguistic input to the comprehension process. Second, the empirical and methodological motivation for the multi-representational view can be found in the work of researchers in pragmatics and neighbouring areas of cognitive science. Third, we need not say that semantic (truth-conditional) interpretation applies to conceptual representations instead of applying to syntactic representations. If we are persuaded that LF-representations are directly interpretable — a claim which a Radical Contextualist would contest (section IV) — we can opt for an hybrid view according to which both LF-representations and SR-representations are subject to truth-conditional interpretation. That is indeed the view defended by Kent Bach and by stipulative minimalists. On this view, each utterance has two sets of truth-conditions. First, there are the truth-conditions of the sentence, obtained by submitting the LF-representation to semantic interpretation (in context). Second, there are the truth-conditions of the richer representation resulting from the interplay of linguistic and pragmatic factors. Unarticulated constituents affect only the second set of truth-conditions.

            Could we decide that only truth-conditional interpretation in the first sense (that which applies to LF-representations) matters to semantics, and relegate the other sort of truth-conditions to the realm of nonliteral meaning? That is what Bach does, but the price to pay is high; for the truth-conditions thus delivered by directly interpreting the syntactic structures of natural language sentences turn out to be very different from the intuitive truth-conditions of utterances. For example, the truth-conditions of sentence (4) (on the base-generated analysis) will be such that (4) will be true, provided it rains somewhere sometime. This is very far from the intuitive truth-conditions of (4) (in the mot salient interpretation, involving a bound unarticulated constituent). Bach is prepared to pay that price but Stanley himself is not, as I pointed out in section VI. "Accounting for our ordinary judgments about the truth-conditions of various sentences is the central aim of semantics", Stanley and Szabo say. Accordingly, they find it "worrisome" that on approaches such as Bach's ,"one has to abandon ordinary intuitions concerning the truth or falsity of most sentences" (Stanley and Szabo 2000: 240). But that is what happens if, applying semantic interpretation directly to LF-representations, we discount unarticulated constituents.

 

References

 

Austin, J.: 1975, How to Do Things with Words, 2nd ed., Clarendon Press, Oxford.

Bach, K.: 1994, 'Semantic Slack: What is Said and More', in S. Tsohatzidis (ed.), Foundations of Speech Act Theory,  Routledge, London, pp. 267-91.

Bach, K.: 2000, 'You Don't Say', forthcoming in Synthese.

Barwise, J.: 1989, The Situation in Logic, CSLI, Stanford.

Barwise, J. and J. Perry: 1983, Situations and Attitudes, MIT Press, Cambridge, Massachusetts.

Carston, R.: forthcoming, 'Explicature and Semantics'.

Chomsky, N.: 1976, 'Conditions on Rules of Grammar', Linguistic Analysis 2, 303-351.

Cooper, R.: 1979, 'The Interpretation of Pronouns', Syntax and Semantics 10, 61-92.

Crimmins, M.: 1992, Talk About Belief, MIT Press, Cambridge, Massachusetts.

Evans, G.: 1980, 'Pronouns', reprinted in his Collected Papers, Clarendon Press, Oxford, 1985, pp. 214-248.

Fillmore, C.: 1986, 'Pragmatically Controlled Zero Anaphora', BLS 12, 95-107.

Fodor, J.: 1983, The Modularity of Mind, MIT Press, Cambridge, Massachusetts.

Grice, P.: 1989, Studies in the Way of Words, Harvard University Press, Cambridge, Massachusetts.

Heal, J.: 1997, 'Indexical Predicates and Their Uses', Mind 106, 619-640.

Kaplan, D.: 1989, 'Demonstratives', in J. Almog, J. Perry and H. Wettstein (eds.), Themes from Kaplan, Oxford University Press, New York, pp. 481-563.

Larson, R. and G. Segal: 1995, Knowledge of Meaning, MIT Press, Cambridge, Massachusetts.

McConnell-Ginet, S.: 1982, 'Adverbs and Logical Form', Language 58, 144-184.

Mitchell, J.: 1986, The Formal Semantics of Point of View, Ph.D. dissertation, University of Massachusetts, Amherst.

Nunberg, G.: 1992, 'Two Kinds of Indexicality', in C. Barker and D. Dowty (eds.), Proceedings of SALT II, The Ohio State University, Columbus, Ohio, pp. 283-301.

Partee, B.: 1989, 'Binding Implicit Variables in Quantified Contexts', CLS 25, 342-365.

Perry, J.: 1993, The Problem of the Essential Indexical and Other Essays, Oxford University Press, New York.

Quine, W.v.O.: 1960, 'Variables Explained Away', in his Selected Logic Papers, Harvard University Press, Cambridge, Massachussets, 1995, pp. 227-235.

Recanati, F.: 1993, Direct Reference: from Language to Thought, Blackwell, Oxford.

Recanati, F.: 1995, 'The Alleged Priority of Literal Interpretation', Cognitive Science 19, 207-232.

Recanati, F.: 1999, 'Situations and the Structure of Content', in K. Murasugi and R. Stainton (eds.), Philosophy and Linguistics, Westview, Boulder, Colorado, pp. 113-165.

Recanati, F.: 2000, 'What Is Said', forthcoming in Synthese.

Rumelhart, D.: 1979, 'Some Problems with the Notion of Literal Meanings', in A.Ortony (ed.), Metaphor and Thought, 2nd ed., Cambridge University Press, Cambridge, 1993, pp. 71-82.

Sperber D. and D. Wilson: 1986, Relevance: communication and cognition, Basil Blackwell, Oxford.

Stainton, R.: forthcoming, 'In Defense of Non-Sentential Assertion'.

Stanley, J.: 2000, 'Context and Logical Form', Linguistics and Philosophy 23, 391-434.

Stanley, J. and Z. Szabo: 2000, 'On Quantifier Domain Restriction', Mind and Language 15, 219-261.

Taylor, K.: 2000, 'Sex, Breakfast, and Descriptus Interruptus', forthcoming in Synthese.