2.1 Grammar is grammar and pragmatics is pragmatics: Inferential pragmatics

Inferential pragmatics theories are all traceable to Grice (Reference Grice1989).³ The main protagonists within this pragmatics research paradigm are Griceans, neo-Griceans, and Relevance theoreticians. Their goals are: (a) to provide an inferential pragmatic theory to account for speaker-intended implicit interpretations; (b) based on this theory, to establish the division of labor between grammar, specifically semantics, and pragmatics, concentrating on difficult cases, often those termed generalized conversational implicature cases (GCIs); (c) to distinguish between different implicit interpretations: particularized conversational implicatures (PCIs), versus GCIs, implicatures versus inferences that form part of the truth conditions of the proposition expressed (the explicature), conventional implicatures versus conversational implicatures. Naturally, both codes and inferences are here discussed, the goal being a complementary distribution between the two.

2.2 (Some) pragmatics is grammar: Form/function pragmatics

Form/function pragmatics, as I propose to call the second research paradigm, has altogether different intellectual roots from inferential pragmatics. Whereas the inspiring figure in the first tradition is Paul Grice, form/function pragmatics emerged as a reaction to Chomsky's generative syntax. Hence, researchers within this paradigm focused on syntax initially, where, they argued, not all conditions (on transformations, in those days) could be grammatically specified, because they often pertained to information structure, a pragmatic concept. The thought behind most research within this paradigm is that grammar is at least partly geared towards communication. Hence, it is only natural for factors relevant to communication to play a role in it. Pioneering research was conducted by Susumu Kuno, Ellen Prince, Wallace Chafe, Sandra Thompson, and various generative semanticists (see Chafe, Reference Chafe and Li1976, Reference Chafe1994; Green, G. M., Reference Green, Cole and Morgan1975, Reference Green1976; Hooper and Thompson, Reference Hooper and Thompson1973; Kuno, Reference Kuno1972, Reference Kuno1987; Prince, Reference Prince1978, Reference Prince, Newmeyer and Robins1988; Thompson, S. A., Reference Thompson, Edmondson, Feagin and Mühlhäusler1990). In time, linguistic analyses were no longer restricted to syntactic structures, discourse markers too receiving prime attention (following Schiffrin, Reference Schiffrin1987). And some current approaches actually advocate the replacement of formal syntax by form/function theories (most notably, Goldberg, Reference Goldberg1995).

2.3 Grammar is (yesterday's) pragmatics: Historical and typological pragmatics

The pragmatists so far considered may infringe on the grammarian's territory, pushing some borders around, but they do not challenge deeply ingrained generative grammar assumptions. The functionalists discussed here do. Most importantly, many reject the innateness hypothesis, arguing that typological universals are better accounted for by reference to extralinguistic factors. Grammar is neither innate nor arbitrary. It evolves in real-time through discourse use. Since communicative needs are relatively similar across different speech communities, it's not surprising that similar grammatical constructions and semantic meanings evolve in a similar fashion from similar sources in unrelated languages. Now, not all extragrammatical forces are pragmatic in the sense being discussed, but we here focus specifically on the role of recurrent pragmatic inferences in shaping grammar.¹² Prominent practitioners within this research paradigm include Givón (Reference Givón1979 and onwards), Traugott (Reference Traugott, Lehmann and Malkiel1982 and onwards), Traugott and Heine (Reference Traugott and Heine1991), Traugott and Dasher (Reference Traugott and Dasher2002), Heine (Reference Heine1993), Bybee et al. (Reference Bybee, Perkins and Pagliuca1994), Comrie (Reference Comrie1994), Du Bois (Reference Du Bois1987), Haspelmath (Reference Haspelmath, Fischer, Norde and Perridon2004), and Croft (Reference Croft2000). Note that these researchers do not actually consider themselves pragmatists. Rather, they see themselves as historical linguists, as functionalists, and/or as typologists. Nonetheless, I insist that this research count as a relevant pragmatics research paradigm, because a crucial grammar/pragmatics interface is analyzed: the process by which recurrent pragmatic inferences turn into grammatical codes.¹³ The majority of the research in grammaticization (the creation of some grammatical form) and semanticization (the evolution of new meanings for old forms) is corpus-based, and so is some recent functional typological research (Haspelmath, Reference Haspelmath2008).¹⁴

2.4 Competition within and across paradigms

We have now briefly surveyed three paradigms of pragmatics research. Inferential pragmatics focuses on drawing the grammar/pragmatics divide at the synchronic level, distinguishing between predominantly truth-conditional codes and pragmatic inferences. Form/function pragmatics focuses on extralinguistic factors, which nonetheless play a grammatical role, because they manifest a coded association with specific linguistic forms. And typolo-gical/historical pragmatics focuses on explaining the diachronic relationship between pragmatic inferences and evolving grammatical codes. To give the flavor of the three research paradigms we briefly mentioned a few proposals within each paradigm, mainly ones pertaining to discourse reference. But what is the relationship between these theories? Are all of them needed? Indeed, some theories within and across paradigms stand in conflict. For lack of space, this section again focuses on reference.

2.4.2 Competition across paradigms

I here cite two cases where the grammar/pragmatics division of labor is under debate. Both form/function pragmatic theories and historical/typological accounts compete with inferential pragmatics as to what to relegate to pragmatics and what to grammar. What is viewed as currently pragmatically inferable under an inferential pragmatics approach may be taken as a form/function code (Hebrew cliticized pronouns). Similarly, a distributional pattern may be analyzed as grammaticized by historical/typological theories, although it seems straightforwardly derivable by inferential pragmatic theories (some reflexive pronouns).

Consider the following Hebrew discourse, where the same entity (the press_i), is already very highly accessible (the first mention here is the sixteenth reference to it):

Given the availability of cliticized pronouns for Hebrew hem ‘they’, it is the avoidance of the shorter forms in the second and fourth references that is puzzling on any inferential pragmatic theory. Minimizing processing cost should have prompted the speaker to use the reduced pronouns throughout according to Relevance theory, for the same interpretation is made available by the two forms (see especially Reboul, Reference Reboul, Connolly, Vismans, Butler and Gatward1997). Levinson actually predicts a disjoint reading here, because the speaker avoided the most minimal form. Form/functionalists Ariel and Reference Gundel, Hedberg and ZacharskiGundelet al., however, can account for the alternating uses of full and cliticized pronouns by noting the points at which the speaker switches from one to the other. Discourse connectives, such as ‘but’ and ‘another thing’ here, signal that a potential topic change may occur, thus reducing the accessibility of the current topic. Reference Gundel, Hedberg and ZacharskiGundelet al. can similarly explain the data as a change from ‘in focus’ to ‘activated’.¹⁵ Hence, the differential preference for full and for cliticized pronouns in different contexts. In this case, it looks like the form/function approach is superior to the inferential approach.

Next, what is the status of the binding conditions regarding the use of reflexives? Levinson (2000: Chapter 4), König and Siemund (Reference König, Siemund, Frajzyngier and Curl2000) and Ariel (Reference Ariel2008b: Chapter 6) are all in agreement that a pragmatic interpretative pattern lies behind these grammaticized conventions (see again sections 2.3.1, 2.3.2). As an inferential pragmatist, Levinson points out that reflexives in different languages don't necessarily share the same grammatical/pragmatic status. For example, in some languages co-arguments can, but need not always be reflexive-marked. Obviously, the pattern is only pragmatic in these languages. But the picture is more complicated than that. Even for a single language, sub-patterns may be either pragmatic or grammatical. For example, according to Ariel's (Reference Ariel2008b: 6.3) grammaticization analysis, some adjuncts obligatorily take reflexive forms (despite herself, with himself) when coreferential with a clause-mate antecedent, although other adjuncts only manifest a pragmatic preference for the extension of the Binding convention beyond co-arguments (except for him(self), picture of her(self), jokes about him(self)). Other sub-constructions obligatorily take a pronoun, even though an obligatorily coreferential co-argument is involved (He didn't have any spots on him(*self/*her)). Although this pattern is an exception to the general Binding principle, and although it is clearly pragmatically motivated (according to the pragmatic motivation here, unmarked coreference is not in need of special marking), a grammatical convention is nonetheless involved (note the unacceptable reflexive form above). In other words, a competition between pragmatic and grammatical analyses is at work not just for the general distributional pattern applicable to the language as a whole, but also at much lower-level generalizations within the same language.

Different pragmatic theories certainly compete for the best account, whether within (2.4.1) or across (2.4.2) research paradigms. While researchers have mainly engaged in intra-paradigm debates, it's time for cross-paradigmatic debates and collaborations to take the stage in pragmatic research.

2.5 Inferential, form/function, and historical/typological pragmatics too

Surprisingly perhaps, there is not much interaction between researchers subscribing to different research paradigms. The goal of this section is to prompt all researchers to open up to the idea that their research must take into consideration questions addressed by other research paradigms. The three paradigms, I argue, complement each other in crucial ways.

3.1 Saying and what is said

The verb “say” has a variety of everyday uses. We speak not only of speakers saying things but also of sentences, signs, and even clocks saying things. Even limited to acts by speakers, “say” has a range of common uses. On one end of that range, it denotes the act of uttering (a sentence, typically) and, on the other end, acts of stating or asserting (a proposition). Acts of the former sort are reported by direct quotation, of the form “S said ‘…’,” and acts of the latter sort by reports of the form “S stated/asserted that p,” where “p” denotes a proposition. Given that we have these other verbs and given that stating or asserting something entails meaning it (not that this in turn entails believing it), it makes sense to reserve the term “say” for the in-between act that is reported by indirect quotation, with sentences of the form “S said that p,” assuming that what is said is a proposition.

The notion of saying, along with the correlative notion of what is said, comes into the picture for a very simple reason: a speaker can say one thing while meaning something else. She could mean something instead of what she says, or she could mean something in addition to what she says. Indeed, a speaker can say something without meaning anything at all, as in recitation or translation. Acts of saying, in the sense in which we will be using the term, correspond to Austin's (Reference Austin1962) notion of locutionary act. Performing a locutionary act goes beyond merely producing certain sounds, even as belonging to a certain language. On the other hand, it must be distinguished from both the illocutionary act of doing something in saying something and the perlocutionary act of doing something by saying something. To perform a locutionary act is to utter a sentence “with a more or less definite sense and a more or less definite reference” (Austin Reference Austin1962: 93). To be sure, these different categories of speech are abstractions from the total speech act. It is not as though in uttering a sentence a speaker is performing a series of acts. Rather, in uttering, say, “I love turnips,” a speaker would be saying that she loves turnips, probably asserting that she loves turnips, and perhaps wanting and maybe even getting her audience to want to try some.

Grice's stipulated sense of “say” is not quite the same as Austin's. He writes, “I intend what someone has said to be closely related to the conventional meaning of the words (the sentence) he has uttered” (Grice Reference Grice, Cole and Morgan1975/1989: 25). Assuming that what is said must be a unique proposition, he required further that any semantic ambiguities be resolved and references be fixed. So far this sounds like Austin's notion of locutionary act, although, curiously, Grice did not connect his notion with his former teacher's (indeed, as we will see in the next section, Grice's analysis of speaker meaning could have benefited by taking into account Austin's distinction between illocutionary and perlocutionary acts). However, unlike Austin, Grice required that saying something entails meaning it. Otherwise, one merely “makes as if to say” it. This requirement seems odd (it conflicts with the first of our opening truisms), since if one can't say something without meaning it, one doesn't say anything unless one means it. There is a sense in which that is true, the sense in which “say” is synonymous with “state.” Indeed, in Grice's (Reference Grice1961) preliminary account of implicature, the preferred verb was “state,” not “say.” In my opinion, Grice's main reason for insisting on this stronger sense of “say” was that it supported his controversial view (proposed in Grice Reference Grice1968), not to be discussed here, that what expressions mean in a language ultimately comes down to what speakers mean in using them. In any case, surely there's a perfectly good sense in which one can say something without meaning it.

What is the rationale for adopting a locutionary notion of saying and the correlative notion of what is said? The point of tying what is said closely to the conventional meaning of the uttered sentence is to limit it to information carried by that sentence. We can think of what is said as, in effect, the interpreted logical form of the sentence. Grice's reason for requiring resolution of ambiguity is to further limit what is said to the sense of the sentence that is operative in the speaker's act of uttering it. Otherwise, whenever there is ambiguity (often!), multiple things would be said. Presumably it is the speaker's semantic intention that does the disambiguating. This intention determines what she takes her words to mean as she is using them, and is distinct from her communicative intention, which determines how she intends her audience to take her act of uttering those words.

As for fixing reference, in cases involving indexicals (including pronouns, certain temporal and locational adverbs, and tense), the point is more subtle. With them we need to distinguish their meaning from their reference and take into account the fact that it is the reference, not the meaning, that figures in what is said. The meaning helps pin down the reference but is not itself part of what is said. So, for example, if I utter the sentence “I love turnips,” I thereby say that I love turnips. I do not say that the current utterer of this sentence loves turnips. After all, what I said, that I love turnips, could be true (not that it is true) even if I hadn't uttered the sentence. The meaning rule, that “I” as used by a given speaker on a given occasion refers to that speaker, is not part of what I say, or would say, if I were to utter “I love turnips.” An analogous point applies to the use of the present tense. The general idea here was developed by Kaplan (Reference Kaplan, Almog, Perry and Wettstein1989a), who proposed that the character of an expression determines the content of the expression relative to a given context of use. The character is a meaning rule that provides for how this content, the expression's reference, is determined in the context. Obviously the rule for “I” is different from, for example, the rule for “you” and the one for “yesterday.”

There is an ongoing debate in philosophy regarding the range of expressions whose reference is literally determined, according to a meaning rule, as a function of their context of use. The primary question at issue is whether it is really the context of use, as opposed to the speaker's referential intention, that determines the reference. We will not pursue this issue here. Suffice it to say that there seems to be a basic difference between what determines the reference of terms like “she” and “that” as opposed to the reference of terms like “I” and “today.” Arguably, the difference is great enough to justify Strawson's (Reference Strawson1950) contention that speakers, not expressions, refer (see Bach Reference Bach, Lepore and Smith2006d for discussion of the question “What does it take to refer?”).

The above niceties aside, both ambiguity and indexicality are different ways in which linguistic meaning does not determine speaker meaning even if the speaker is being completely literal. The meaning of an ambiguous sentence underdetermines what a speaker could mean in using it literally, obviously because the sentence has too many meanings, i.e. more than one (a speaker can mean more than one thing by trading on an ambiguity, as with puns, but this is an exceptional case). The case of indexicality is different. When we use an expression like “I,” “tomorrow,” “she,” or “that,” the meaning of the expression merely constrains what we can mean in using it literally. Indexicality is like ambiguity insofar as in both cases linguistic meaning limits but does not fully determine what one can mean in speaking literally. They differ, however, in that with ambiguity there is too much linguistic meaning and with indexicality there is too little.

As we will see later on, following our discussions of meaning and implicature, ambiguity and indexicality are not the only ways in which linguistic meaning can underdetermine literal speaker meaning. It will emerge that finding a suitable notion of saying, together with the correlative notion of what is said, is not as straightforward as Grice supposed, and not just because he needlessly required that to say something entails meaning it. But what is it for a speaker to mean something?

3.2 Speaker meaning

Grice (Reference Grice1957) contrasted “natural” meaning with the sort of meaning (“non-natural,” he called it) involved in language and communication. Smoke means fire because it is naturally correlated with fire, but the word “smoke” means smoke by virtue of being conventionally correlated with smoke. Smoke means fire in the sense of indicating the presence of fire, and that is because it is correlated with fire. However, the word “smoke” is not correlated with the presence of smoke. It is a conventional means for talking about smoke, whether or not smoke is present. Its meaning is a matter of convention, since it could just as well have meant something else (and some other word could just as well have meant smoke). It means smoke because, and only because, speakers normally use it to mean that and expect others who use it to mean that as well.

So within the category of non-natural meaning there is both linguistic meaning, in this case what “smoke” means, and speaker meaning, here what a speaker means in using that word. Take an example. Suppose a speaker utters the sentence, “I smell smoke,” using the pronoun “I” to refer to herself, the verb “smell” (in the present tense) for olfactory sensing, and the noun “smoke” to refer to smoke, presumably some that is nearby. Given what these words mean and how, as syntactically determined, these combine to comprise the (linguistic) meaning of the entire sentence, the semantic content of the sentence, relative to that context, is that she (the speaker) smells smoke. This could well be what she means in uttering that sentence, but it might not be. For she could have been speaking figuratively, in which case she would have meant something else. She could have meant, regarding something her interlocutor had just said, that he was trying to divert her attention from what was at issue. On the other hand, she could have been speaking perfectly literally. Even then, it is one thing for the sentence to mean something and another for a speaker to mean that in uttering it.

3.3 Conversational implicature

Grice is best known, in both linguistics and philosophy, for his theory of conversational implicature. It was sketched in a section (III) of his 1961 paper, and developed in his William James Lectures at Harvard in 1967, which were subsequently published individually in disparate places and eventually collected as Part I of the posthumous Grice Reference Grice1989. The main ideas are laid out in “Logic and Conversation” (Grice Reference Grice, Cole and Morgan1975/1989), which, from what I have been able to ascertain from Google Scholar, is the most cited philosophy paper ever published. Grice's basic idea was not new, although his name for it was. What distinguished his work from previous work on “contextual” or “pragmatic” implication (see Hungerland Reference Hungerland1960) was his ingenious account of how it works (it also served as an antidote to the excesses of ordinary language philosophy, in ways chronicled in Chapman Reference Chapman2005). This account was essentially an extension of his theory of speaker meaning, but what made it original, as we will see, was the role of his “Cooperative Principle” and the various “maxims of conversation” that fall under it.

In Grice's view one can mean something either by saying it or by saying (or “making as if to say”) something else. What one implicates by saying something is generally not implied by what one says. That is why Grice used the verb “implicate” rather than “imply” and the neologism “implicature” rather than “implication.” For example, suppose you are asked about a dinner you had at an expensive restaurant, and you reply, “It didn't make me sick.” Your saying this implicates that it was not very good. However, what you said obviously does not imply this. After all, a dinner that does not make you sick can still be excellent. However, it is possible for what is implicated to be implied. If you are asked whether you have more than two children and you reply that you have three girls and a boy, what you say implies the very thing that you implicate, namely, that you have more than two children. That is because you also mean, albeit indirectly, that you have more than two children. There are many other things implied by what you say that you do not mean, hence do not implicate, for example that you have more than one child and that you have more than two girls (much of Davis's (Reference Davis1998) and other critiques of Grice assume that he did not require that implicatures be things speakers mean).

The mediocre meal example illustrates Grice's observation that conversational implicatures are cancelable – you could have added, “I don't mean to suggest that the meal wasn't great,” without taking back your assertion that it didn't make you sick. In fact, there are circumstances in which the implicature (that the dinner was not very good) would not have been in the offing in the first place. Suppose that you and your interlocutor had just learned that there had been an outbreak of food poisoning at the restaurant in question. In that case, your saying that the meal didn't make you sick would not implicate anything about its culinary quality.

How can a speaker implicate something that is not implied by what she says and still manage to convey it? She can do this by exploiting the fact that the hearer presumes her to be cooperative, in particular, to be speaking truthfully, informatively, relevantly, and otherwise appropriately. If taking the utterance at face value is incompatible with this presumption, the hearer, still relying on this presumption, must find some plausible candidate for what else the speaker could have meant. And, crucially, the speaker must intend him to do this. In the case of the speaker asked about dinner, the hearer must figure out what she meant, relying on the presumption that she intended it to be an accurate, informative, and appropriate answer to the question. In effect what the hearer does is, on the presumption that the speaker is being cooperative, to find a plausible explanation for why she said what she said.

3.4 Common misunderstandings about conversational implicature

There are two common misconceptions about the role of the maxims (or presumptions). First, they do not determine implicatures (even Grice (Reference Grice1989: 372) occasionally suggested that they do) but, rather, help explain how they get conveyed. They are considerations that speakers implicitly intend hearers to, and hearers do, take into account to figure out (“determine” in the sense of ascertain) what the speaker is implicating. Since that is a matter of speaker meaning, it is the speaker's communicative intention that determines (in the sense of constitute) what is implicated. Also, it should not be supposed, as it often is, that the maxims apply only to implicatures. This misconception is understandable insofar as the maxims play a key role in Grice's account of how implicatures get conveyed, but in fact they apply equally to completely literal utterances, where the speaker means just what she says. After all, the hearer still has to infer this. It is thus wrong to suppose that the maxims come into play only where linguistic meaning leaves off and speaker meaning and extralinguistic, contextual information take over (for more on context and what it does and doesn't determine, see Bach Reference Bach and Szabó2005).

Another misunderstanding concerns what, when a speaker says one thing but means something else, the hearer is to infer. Contrary to what philosophers and linguists seem commonly to suppose (perhaps because of the ambiguous phrase “infer what the speaker implicates”), the hearer does not have to infer the thing the speaker implicates. He merely has to infer that the speaker implicates (means) it. So, for example, if the speaker says and means that a certain new book has a beautiful cover and implicates that it is not worth reading, the hearer needs to infer that the speaker means that it is not worth reading. He does not need to infer that it is in fact not worth reading. The speaker might want him to believe that, but this is not necessary as far as communication is concerned. Indeed, he could even doubt that she believes it (he might think she resents the author's success).

A remarkably widespread misconception is that implicatures are inferences, or at least are determined (and not merely ascertained) by inferences rather than by the speaker. It is based on confusing what is implicated (by a speaker) with what is involved in figuring out what is implicated. Implicatures are things speakers mean, not what hearers, even rational ones, think they mean. Accordingly, if a speaker is to succeed in communicating something, the hearer must figure out that it is meant. That requires inference. Yet some of the most brilliant researchers, including Levinson (Reference Levinson2000), Geurts (Reference Geurts2010), and Reference Chierchia, Fox, Spector and MaienbornChierchiaet al. (forthcoming), write as if implicatures themselves are inferences.¹

A further misconception is that linguistic expressions can implicate things. Speakers do. To be sure, there are certain expressions that are characteristically used (by speakers) to implicate things (Davis (Reference Davis1998) regards this as a reason to say that sentences themselves implicate things, but he does so in the course of arguing that what is implicated is generally a matter of convention, not speaker intention). When this occurs we have what Grice calls generalized conversational implicatures (as opposed to particularized ones). These have been investigated in great depth by Levinson (Reference Levinson2000), who thinks they give rise to an intermediate level of meaning. In fact, they give rise to an intermediate kind of inference, but inferences are not meanings.

A related misunderstanding leads to Levinson's objection that Grice's approach cannot account for the (alleged) phenomenon of “pragmatic intrusion,” which he thinks is exemplified by so-called embedded implicatures. As Levinson puts it,

This objection is based on confusing the two sorts of determination mentioned above. The first two highlighted words are forms of “determine” in the sense of ascertain, but the last two, where Levinson draws his conclusion, are in the constitutive sense. In that sense, what is said neither determines nor is determined by what is implicated. This is a matter of the speaker's intention.

Levinson and many others misconstrue Grice's account as a psychological model of the hearer's inference, indeed one according to which the hearer must ascertain what the speaker says before figuring out what the speaker is implicating (see Bach Reference Bach2001: 24–5, and Saul Reference Saul2002b). But that is not how Grice intended his account. He required that “the presence of a conversational implicature be capable of being worked out” (1975/1989: 31), but he did not require that it must be.

This last misconception (for still more see Bach Reference Bach, Birner and Ward2006c) leads to the widespread misconception, evident from an extensive literature, that some implicatures are “embedded,” “pre-propositional” (Recanati Reference Recanati2003), or “pre-compositional” (Reference Chierchia, Fox, Spector and MaienbornChierchiaet al. forthcoming). This is thought to arise with utterances of sentences like these:

1. (12) It is better to get married and get pregnant than to get pregnant and get married.

2. (13) Bill thinks that there were four boys at the party.

Since the two infinitival clauses of (12) are semantically equivalent, a speaker is likely to implicate that what is better is getting married and then getting pregnant. With (13) the implicature is not that there were exactly four boys at the party but that Bill thinks that. In fact, such examples illustrate merely that the process of figuring out what is implicated does not require first ascertaining what is said. They do not show that the implicature is embedded in anything. Indeed, since speakers implicate, it does not even make sense to say that some implicatures are embedded. It is irrelevant that the hearer figures out what is implicated without having first to figure out what is said. Recanati supposes that Grice's account requires (conversational) implicatures to have a “global, post-propositional character,” on the grounds that for Grice “implicatures are generated via an inference whose input is the fact that the speaker has said that p” (Recanati Reference Recanati2003: 300). Recanati's point is that certain implicatures get “computed” before what is said is ascertained. However, this does not show that the implicature itself is somehow embedded. In the case of (13), one of Recanati's examples, the implicature is supposed to be that Bill believes that there were not more than four boys at the party in question. But the implicature is not literally embedded. What the speaker says is the proposition expressed by (13), and what the speaker implicates is this other proposition. How the hearer figures this out is another matter.

Not only that, it seems that the speaker does not really mean two things, the proposition that Bill believes that there were four boys at the party and the proposition that Bill believes that there were not more than four boys at the party in question. It seems, rather, that the speaker means but one thing, that Bill believes that there were exactly four boys at the party in question. This example illustrates that some apparent instances of implicature are really cases of something else.

3.5 Between saying and implicating

A speaker can say something and mean just that. The contrast between saying and implicating allows both for cases in which the speaker means what she says and something else as well and for ones in which the speaker says one thing and means something else instead. Grice counted both as kinds of implicature, although the latter might better be described as speaking figuratively (recall, though, that Grice described this as a case of merely “making as if to say” something, since for him saying entails meaning). Grice seems to have overlooked a phenomenon intermediate between saying and implicating, one that has been investigated by many others (Sperber and Wilson Reference Sperber and Wilson1986, Bach Reference Bach1994a, Carston Reference Carston2002, and Recanati Reference Recanati2004a). As they have observed, there are many sentence forms whose typical uses go beyond their meanings, even with references fixed and ambiguities resolved, but are not cases of implicating (or speaking figuratively). The reason is that what the speaker means, although distinct from what is said (strictly speaking), is too closely tied to what is said to be a case of merely implicating.

3.6 Summing up

Clarifying the distinction between saying, meaning, and implicating has required refining Grice's notions of each by way of introducing further distinctions. Borrowing from speech act theory, we invoked the distinction between locutionary and illocutionary acts to drop Grice's counter-intuitive requirement that to say something entails meaning it. In order to make sense of Grice's ingenious idea that speaker meaning involves a kind of self-referential yet audience-directed intention, we needed to distinguish the specifically communicative “effect” of understanding from other, perlocutionary effects on an audience. Next we clarified Grice's notion of conversational implicature, mainly by identifying various misconceptions about it. Many of these can be avoided by heeding the distinction between a speaker's meaning something and the hearer's figuring out that the speaker means it. Also, the conversational maxims or presumptions do not generate or even determine implicitures but, rather, provide considerations that the hearer may be intended to take into account to figure out what a speaker means/implicates. We then pointed out that the distinction between saying and implicating is not exhaustive and defended the use of the term “impliciture” for what falls between what is said (and meant) and what is implicated. Finally, by distinguishing the speaker's semantic intention from her communicative intention and both from what may be intuitive, evident to, or otherwise go on in the mind of the hearer, we defended the notion of a speaker's purely locutionary, or semantic, act of saying. The aim of all this has been to distinguish the linguistic from the extralinguistic information that speakers try to make available to their listeners, to identify the sorts of intentions they have in so doing, and to describe the means by which this information is made available to or is inferable by the hearer from the fact that the speaker did what she did.

We have not covered the many debates that Grice's notions have provoked, the radical alternatives to his approach, the range and variety of implicatures (never mind presuppositions) that philosophers and linguists have discussed, or how all this fits into speech act theory and pragmatics in general. Whereas we have focused on issues raised by Grice's most important and influential ideas, speaker meaning and conversational implicature, Neale (Reference Neale1992) presents a much fuller discussion of these and related ideas, Levinson (Reference Levinson2000) offers an in-depth study of generalized conversational implicature, Bach (Reference Bach1999a, Reference Bach2006a) and Potts (Reference Potts2005, Reference Potts2007a) address Grice's controversial notion of conventional implicature, Chapman (Reference Chapman2005) provides a full-length intellectual biography of Grice, Horn (Reference Horn2009b) gives a forty-year retrospective on implicature, Geurts (Reference Geurts2010) offers in-depth study of quantity implicatures, the most thoroughly studied kind, and the papers collected in Petrus Reference Petrus2010 present some of the most recent developments.

4.1 A “vulgar conflation”

In Bach's short manifesto unveiling his list of the top ten misconceptions about implicature (2006c: 23), #2 is the thesis “Implicatures are inferences.” For Bach, such a claim – whether explicit (as in the subtitle of Levinson Reference Levinson2000, “Generalized conversational implicatures as default inferences”) or implicit – is a “misdenomer” amounting to a “slight variation on the vulgar conflation of implying with inferring.” The distinction in each case is seen as a straightforward one: implying (or, more specifically, implicating) is something the producer or sender of a message (speaker or writer) does, while inferring pertains to the cognitive effort of the receiver. Bach submits the entry from the American Heritage Book of English Usage:¹

The distinction is vital for pragmatic theory because an interpreter may “recover” an implication (presupposition, implicature) that was not intended by the utterer, and a speaker may imply (presuppose, implicate) something that the interpreter does not grasp. In I. A. Richards's words (MWDEU 1994: 541), “An utterer may imply things which his hearers cannot reasonably infer from what he says”; in other cases, the expectation of inference may be reasonable but nevertheless unfulfilled, whether through inattention or a mismatch of shared beliefs. But what of the straightforward distinction between implying and inferring itself?

Most usage manuals endorse the distinction, but a closer look shows that it's not that simple. WDS (1942), for example, begins with this prescription, complete with pointy finger:

But then the rockier landscape of actual usage is surveyed (WDS 1942: 449–50):

So, in effect, “Don't use infer to mean imply – but if you do, make sure your subject is inanimate.” The secure footing appears to have become a bit slippery, presumably because a misstep with impersonal subjects yields no ambiguity.

In Merriam Webster's Dictionary of English Usage, the linguistically best informed usage compendium, the extensive entry for infer, imply (MWDEU 1994: 541–4) disentangles three uses of infer, eloquently illustrating that “Real life is not as simple as commentators would like it to be” (541). Besides the universally accepted ‘deduce, conclude’ sense for infer, MWDEU differentiates the impersonal-subject infer-for-imply illustrated in the Shakespeare, Milton, and Hardy examples above² from the “personal infer” attested in an Ellen Terry letter from 1896: “I should think you did miss my letters. I know it! but…you missed them in another way than you infer, you little minx!” The MWDEU suggests that this latter use, the specific target of twentieth-century prescriptive opprobrium, has always been largely restricted to informal spoken language.

Similarly, the OED's sense 4 for infer reads as follows:

But even if incorrect, not inexistent; cites range from the sixteenth century to the nineteenth (“Socrates argued that a statue inferred the existence of a sculptor”) and the twentieth (“I can't stand fellers who infer things about good clean-living Australian sheilahs”).

It should be noted that the direct target of objections like that of Bach (or Horn Reference Horn, Horn and Ward2004) is the nominal form (inference) rather than the verbal (infer), as in Levinson's implicatures as default inferences. But the speaker-oriented sense here boasts its own distinguished pedigree: the relevant OED sense 2 for inference is glossed as

– with no disparagement of the latter usage and with attestations back to a 1612 essay by Francis Bacon on judicial practice warning that “Judges must beware of hard constructions and strained inferences.”

What does not seem to be noted in any of the dictionaries and manuals is the asymmetry exhibited in these purportedly erroneous uses. While infer has been used for almost five centuries in the sense of ‘imply, convey’ (whether with impersonal or personal subjects), imply is always sender-, not receiver-oriented; it is never used for ‘infer’, to refer to the cognitive processes of the hearer or reader. Although this asymmetry may be seen as betokening the primary status of the speaker or producer of the message, we can find other cases in which a predicate exhibits an analogous ambiguity but where the direction of the meaning shift is less clear.³ Thus entendre exhibits a range of meanings in French from the speaker-oriented ‘mean, intend’ to the hearer-oriented ‘understand, hear’. This is especially significant for pragmatic theory because, as we shall see, conversational implicature first appeared in Mill's invocation of the sous-entendu, that which is literally under-meant or under-understood.

4.3 Scalar implicature and the maxim map

Conversational implicature differs from contextual implication, and non-demonstrative implication more generally, in being defined as a relation between a speaker (not a sentence!) and a proposition that typically arises from the exploitation of the maxims (Grice Reference Grice1989: 26ff.; cf. Horn Reference Horn, Horn and Ward2004); in the case of scalar implicature in particular, what is implicated depends on what isn't (but could have been) said. The crucial principle of informative strength or quantity – whether formulated à la Strawson Reference Strawson1952 channeling Grice (“One should not make the (logically) lesser, when one could truthfully (and with greater or equal clarity) make the greater claim”), à la Grice Reference Grice1961 (“One should not make a weaker statement rather than a stronger one unless there is a good reason for so doing”) or à la Grice [1967]1989 (“Make your contribution as informative as is required (for the current purposes of the exchange)”) – balances what the speaker can and does say with what she doesn't and hence presumably can't (or shouldn't) say. A speaker may opt for a weaker utterance from a belief that to utter its stronger counterpart might violate considerations of relevance, brevity, clarity, or politeness (note the codicils and parentheticals in each of the formulations above¹⁰), but especially – as Mill and Doyle foresaw – from a lack of certainty that the stronger counterpart holds.

This reasoning, exploiting Grice's first quantity maxim, is systematically exploited to yield upper-bounding scalar implicatures associated with relatively weak scalar operators, those configurable on a scale defined by unilateral entailment as in <all, most, many, some>. What is said in the use of a weak scalar value like those boldfaced in (2) is the lower bound (…at least…); what is implicated, in the absence of contextual or linguistic cancellation, is the upper bound (…at most …). What is communicated is the “two-sided reading

” that combines what is said with what is implicated. Thus in (2c), to quote Mill (Reference Mill1867: 512), “If we assert that a man who has acted in a particular way must be either a knave or a fool, we by no means assert…that he cannot be both” – but we do typically communicate this “exclusive” understanding of the disjunction.

The alternative view on which each scalar predication in (2) is lexically ambiguous between one-sided and two-sided readings contravenes Grice's (Reference Grice1989: 47) Modified Occam's Razor: “Senses are not to be multiplied beyond necessity.” Scalar implicature was introduced and formalized in work by Horn (Reference Horn1972, Reference Horn1989), Gazdar (Reference Gazdar1979), Hirschberg (Reference Hirschberg1991), and Levinson (Reference Levinson2000); cf. also Matsumoto (Reference Matsumoto1995), Katzir (Reference Katzir2007), and Geurts (Reference Geurts2010) for insightful discussions of certain problems arising in the implementation of the central notions involved and Bontly (Reference Bontly2005) for a defense of Modified Occam's Razor based on its role as a heuristic in acquisition.

The implicature-based approach to scalar predications has been vigorously challenged by relevance theorists (see Carston Reference Carston and Kempson1988, Reference Carston2002, Reference Carston2004a and work reviewed therein), who take such sentences to involve propositional ambiguity, with the pragmatically enriched two-sided meanings constituting not implicatures but explicatures, pragmatically derived components of propositional content.¹¹

Two major challenges to the Gricean picture of implicatures involve the number and status of the maxims and the relationship between implicature and propositional content. To begin with the former issue, Grice himself later acknowledged (1989: 371ff.) that the four macroprinciples (inspired by Kant) and nine actual maxims in his inventory are somewhat overlapping and non-coordinate. The number of maxims has been revised both upward (Leech Reference Leech1983) and downward. The dualistic program of Horn (Reference Horn1984b, Reference Horn1989, 2007a) follows Grice (Reference Grice1989: 371) in ascribing a privileged status to Quality, on the grounds that without the observation of Quality, or Lewis's (Reference Lewis1969) convention of truthfulness, any question of the observance of the other maxims fails to arise (though relevance theorists, beginning with Sperber and Wilson Reference Sperber and Wilson1986, offer a dissenting view). The remaining maxims are subsumed under two countervailing functional principles governing the economy of communication. On Horn's Manichaean model, implicatures may be generated by either the Q Principle (essentially ‘Say enough’, generalizing Grice's first sub-maxim of Quantity and collecting the first two ‘clarity’ sub-maxims of Manner) or the R Principle (‘Don't say too much’, subsuming Relation, the second Quantity sub-maxim, and brevity).

The hearer-oriented Q Principle is a lower-bounding guarantee of the sufficiency of informative content, exploited to generate upper-bounding (typically scalar) implicata. The speaker-oriented R Principle reflects Zipf's principle of least effort dictating minimization of form, exploited to induce strengthening implicata; it is responsible for euphemism, indirect speech acts, neg-raising, and meaning change (Horn Reference Horn and Burton-Roberts2007a). Opposition and equilibria between speaker's and hearer's communicative economies have been posited since Paul (Reference Paul1889: 351ff.) and Zipf (1949: 20ff.). According to the division of pragmatic labor (Horn Reference Horn and Schiffrin1984b), a relatively unmarked form – briefer and/or more lexicalized – will tend to become R-associated with a particular unmarked, stereotypical meaning, use, or situation, while its periphrastic or less lexicalized counterpart, typically more complex or prolix, will tend to be Q-restricted by implicature to those situations outside the stereotype, for which the unmarked expression could not have been used appropriately (as in kill vs cause to die, or mother vs father's wife). Formalizations of the division of pragmatic labor have been undertaken within bidirectional optimality theory and game-theoretic pragmatics; cf. e.g. Blutner Reference Blutner, Horn and Ward2004, van Rooij Reference van Rooij2009, and references cited therein.

Levinson's (Reference Levinson2000) framework posits an interaction of three heuristics: Q, I (for Informativeness, ≈ Horn's R), and M (Manner). Levinson's reconstruction of the division of pragmatic labor involves not Q but the M heuristic, given that some differs from all in informative content whereas kill differs from cause to die in complexity of production or processing. As Levinson acknowledges, however, the Q and M patterns are closely related, since each is negatively defined and linguistically motivated: S tacitly knows that H will infer from S's failure to use a more informative and/or briefer form that S was not in a position to have used that form. Unlike Q implicature, R/I-based implicature is not negative in character and is socially rather than linguistically motivated, typically yielding a culturally salient stereotype (cf. Huang Reference Huang2007 for a useful overview).

Relevance theorists (e.g. Sperber and Wilson Reference Sperber and Wilson1986, Carston Reference Carston2002) posit one pragmatic principle, that of Relevance, defined in non-Gricean terms. It may be argued, however, that the RT program is itself covertly Manichaean, given that Relevance itself is calibrated as a minimax of effort and effect. In the words of Carston (Reference Carston1990: 231), “Human cognitive activity is driven by the goal of maximizing relevance: that is…to derive as great a range of contextual effects as possible for the least expenditure of effort.”

4.4 What is meant and what is said: the -plicature family

We now return to the perennial dispute over the shape of the landscape of implied meaning. In recent years a partial consensus has formed as to semantic underspecification and pragmatic enrichment, one that transgresses the view inherited from Grice that the pragmatics can be simply “read off” the semantics. When we turn from the relatively straightforward cases of reference fixing and ambiguity resolution acknowledged by Grice himself to the more problematic phenomena of completion, saturation, and free enrichment (cf. Bach Reference Bach2001; Recanati Reference Recanati2001, Reference Recanati2002; Carston Reference Carston2002; and references therein, as well as the relevant chapters in Horn and Ward Reference Horn and Ward2004), it is clear we must grant what Bach (Reference Bach and Szabó2005) terms the “contextualist platitude”:

Thus, the speaker uttering the non-bracketed material in each example in (3) may well communicate the full sentences indicated, enriched by the bracketed addenda. As seen from the cancelability evidence in (4), however, this process, resulting in truth-conditionally relevant propositions not directly expressed, is pragmatic in character.

1. (3)

   1. a. I haven't had breakfast {today}.

   2. b. John and Mary are married {to each other}.

   3. c. They had a baby and they got married {in that order}.

   4. d. Dana is ready {for the exam}.

2. (4)

   1. a. John and Mary are married, but not to each other.

   2. b. They had a baby and got married, but not necessarily in that order.

Those enrichments constituting necessary conditions for the expression of truth-evaluable propositions involve what Recanati has called saturation and Bach completion. Recanati (Reference Recanati2002) distinguishes the bottom-up processes linguistically triggered by indexicals (I, today) and other expressions requiring saturation from those top-down modulation and free enrichment processes motivated on purely pragmatic grounds. At issue here are, for example, the underspecification of genitives (John's car – the one he owns? is driving? is following? is painting? is repairing?), unspecified comparison sets (Chris is tall – for an adult? for an adult American of the relevant sex?), and various expressions with apparent free variable slots: You are late (for what?), Robin is too short (for what?). For Recanati – contra Stanley Reference Stanley2000 and King and Stanley Reference King, Stanley and Szabó2005 – “unarticulated constituents” are real and cannot be reduced to independently motivated elements in abstract syntax or logical form.

Since Grice, the pragmatic landscape has exploded with aspects of meaning variously identified as conversational implicatures, conventional implicatures, presuppositions, implicitures, and explicatures. These are not simply diverse labels for given subclasses of implication but different ways of mapping the territory between the said and the meant. Situating “what is said” along this spectrum is itself controversial; what is said for Recanati (Reference Recanati2001), Ariel (Reference Ariel2008b), and the relevance theorists is enriched by pragmatically derived material (hence constituting an explicature). Levinson (Reference Levinson2000), on the other hand, responds to the apparent need to allow “pragmatic intrusion” into what is said by allowing conversational implicatures to have truth-conditional consequences for the propositions in question, contra Grice; in cases like (3c) or Deirdre Wilson's aperçu It's better to meet the love of your life and get married than to get married and meet the love of your life, an implicature (“P precedes Q”) can feed into (rather than just being read off) what is said. (See Carston Reference Carston2002 and Russell, Benjamin Reference Russell2006 for illuminating discussions of the complexity of conjunction buttressing.)

For orthodox Griceans, the pragmatically enriched proposition in such cases – what is communicated – is distinct from what is said. As we saw in §4.2, an “austere” conception of what is said (to borrow Jenny Saul's phrase; cf. Borg Reference Borg2004, Horn Reference Horn2009b), corresponding closely to the syntax of the sentence uttered and excluding pragmatically derived material, may have more to recommend it than first appears. Further, as Bach (Reference Bach2001) observes, once we abandon “OSOP” (the One Sentence, One Proposition assumption) we can recognize that a sentence may express not only more than one proposition but fewer than one. What is said in Dana is ready constitutes not a truth-evaluable proposition but a propositional radical. Completing such a radical within a given context to yield e.g. Dana is ready to write a dissertation yields not what is said (which is tightly constrained by the actual syntax) or an explicature (since there is nothing explicit about it), but rather an impliciture, a proposition implicit in what is said in a given context as opposed to a true implicature, a proposition read off what is said (or the way it is said). What Grice failed to recognize, argues Bach, is the non-exhaustive nature of the opposition between what is said and what is implicated.

“Scalar ‘implicatures’ are implicatures” is #9 in Bach's hit parade of misconceptions (2006c: 28–9): since a speaker uttering “Some of the boys went to the party” means not two separate things but just one, i.e. that some but not all of them went, this enriched proposition is an implic-i-ture (built up from what is said), not an implicature. But on the Gricean account (Horn Reference Horn1972, Reference Horn1989; Gazdar Reference Gazdar1979; Hirschberg Reference Hirschberg1991), the strong scalar implicature here is “Not all of the boys went to the party”; this combines with what is said (“Some…”) to yield what is communicated (“Some but not all…”). Thus the impliciture includes the scalar implicature rather than supplanting it.¹²

While Levinson (Reference Levinson2000) defines generalized conversational implicatures as default inferences, others argue that they are neither inferences – an implicature is an aspect of speaker's meaning, not hearer's interpretation¹³ – nor true defaults. This last point is especially worth stressing in the light of much recent work in experimental pragmatics (see e.g. Noveck and Posada Reference Noveck and Posada2003; Bott and Noveck Reference Bott and Noveck2004; Breheny et al. Reference Breheny, Katsos and Williams2006; Katsos Reference Katsos2008) suggesting that children and adults do not first automatically construct implicature-based enriched meanings for scalar predications and then, when the “default” interpretation is seen to be inconsistent with the local context, undo such meanings and revert to the minimal implicature-free meaning. To the extent that this empirical work on the processing of implicature recovery can be substantiated and extended, this is a very interesting result, but not (contrary to some claims) one that threatens the actual Gricean tradition, which predicts no automatic enrichment or default interpretation. This is clear from the passage distinguishing generalized and particularized implicature (Grice Reference Grice1989: 37, emphases added):

The classic contrast here dates back to Grice Reference Grice1961: §3 – the particularized implicature with the “Gricean letter of recommendation” for a philosophy job candidate (Jones has beautiful handwriting and his English is grammatical) vs the generalized implicature with logical disjunction (My wife is in Oxford or in London, implicating I don't know which). Crucially, an implicature may arise in an unmarked or default context without thereby constituting a default or automatic inference. (See Bezuidenhout Reference Bezuidenhout, Campbell, O’Rourke and Shier2002a; Jaszczolt Reference Jaszczolt2005; and Geurts Reference Geurts2009 for different views on defaults and their relation to implicature.)

Despite their substantial differences (from each other and from Grice) as to the role of implicature and the relation between what is implicated and what is said, the proponents of the approaches touched on above share Grice's commitment to situating implicature within a rationality-based pragmatics. On a competing view that has recently been elaborated by Chierchia (Reference Chierchia and Belletti2004) and his colleagues, scalar implicatures in particular are generated locally as part of the grammar and/or the conventional lexical semantics of weak scalar operators. Support for this variety of defaultism involves an appeal to cases in which the Gricean model appears to yield the wrong results, thus arguing for local computation of “embedded implicatures.” Others (e.g. Sauerland Reference Sauerland2004; Russell, Benjamin Reference Russell2006; Horn Reference Horn, Turner and von Heusinger2006) have challenged these conclusions and defended a global account of implicature along Gricean lines. In particular, Geurts (Reference Geurts2009, Reference Geurts2010) provides a broad survey of the landscape. Drawing a distinction between marked L[evinson]-type cases and unmarked C[herchia]-type cases of putative locality effects, Geurts (Reference Geurts2009) argues that unlike the latter type, the Levinsonian contrast-induced narrowings represent true problems for a classical Gricean (or neo-Gricean) theory of implicature but shows that these can be handled by allowing upper-bounding to enter into the reinterpretation of what scalar operators express, a reinterpretation that is itself pragmatic in nature. In his treatise on Q-implicatures, Geurts (Reference Geurts2010) argues that the conventionalist alternative to a Gricean approach is not only stipulative but also empirically flawed in predicting the full range of implicature-related results.

4.5 Conventional implicature from Frege to Grice (and beyond)

Alongside the successful conversational implicature model, Grice's category of conventional implicature – a non-cancelable but truth-conditionally transparent component of encoded content – plays the role of ugly stepsister. The coherence of this category has evoked much skepticism: Bach (Reference Bach1999a) consigns it to the dustbin of mythology, Carston (Reference Carston2002: 134) remarks that “there simply is no such thing as ‘conventional’ implicature in relevance theory (or, we would argue, in reality),” while Potts (Reference Potts2005, Reference Potts2007b) rehabilitates it in a different guise. But Grice's account of conventional content that does not affect the truth conditions of the asserted proposition has a rich lineage. Frege (Reference Frege1892, Reference Frege and Hermes1897, Reference Frege1918) delineates a class of meanings that, while of linguistic interest, do not “affect the thought”:

While recent scholarship largely follows Dummett (Reference Dummett1973) in dismissing Frege's positive proposals in this area as representing a confused and subjective notion of “tone,” this mischaracterizes Frege's actual account of the relevant phenomena. The two verbs in Geach's rendering highlighted above – hint and intimate – both translate Frege's andeuten, i.e. ‘conventionally implicate’; no subjectivity or confusion is involved.

For a range of constructions including discourse particles (but, even, Ger. ja, doch), subject-oriented adverbs, epithets, and other “loaded” words, a version of the approach proposed by Frege and Grice remains eminently plausible (Barker Reference Barker2003; Horn Reference Horn, Kecskes and Horn2007b, Reference Horn2008; Gutzmann Reference Gutzmann2008; Williamson Reference Williamson, Almog and Leonardi2009). Such an approach extends naturally to a range of other linguistic phenomena, including the familiar vs formal second person singular (“T/V”) pronouns of many modern European languages, evidential markers, and arguably the uniqueness/maximality condition on definite descriptions. In addition, certain syntactic constructions can be profitably analyzed along these lines such as the southern US English “personal dative,” a non-argument pronominal appearing in transitive clauses that obligatorily coindexes the subject as exemplified in I love me some squid, truth-conditionally equivalent to, but not fully synonymous with, ‘I love squid’ (Horn Reference Horn2008). In each case, we find aspects of conventional content that are not entailed and do not fall inside the scope of logical operators.

The category of conventional implicatures poses a complication for the distinction between what is said and what is meant. Such expressions present a recalcitrant residue for Grice (who was concerned with delineating what is said and what is conversationally, and hence calculably, implicated) as they did for Frege (who was concerned with the thought, i.e. with sense and potential reference); for both, detecting a conventional implicature facilitates the real work by clearing away the brush. But Grice did undertake to situate this relation within what we refer to (though he did not) as the semantics/pragmatics divide. His contributions in this area, if not always accepted, are widely recognized, as in this passage from Davidson (Reference Davidson1986: 161–2):

But how, exactly, does this sorting work? If descriptive content, reflecting what is said, is clearly semantic and if what is conversationally implicated (e.g. the “for all I know, not both p and q” upper-bounding implicatum associated with the utterance of the disjunction p or q or the negative effect of the Gricean letter of recommendation) is pragmatic (paceReference Chierchia, Fox, Spector and MaienbornChierchia 2004, among others), where is conventional implicature located? One standard view is that by falling outside what is said, the conventionally implicated must be pragmatic (see e.g. Gutzmann Reference Gutzmann2008: 59). One argument on this side is terminological; in Kaplan's words (1999: 20–21):

To be sure, conventional implicatures are implicatures. But then again, they are conventional; we are indeed dealing here, unlike in the maxim-based cases, with aspects of content.

Two decades after the William James lectures, Grice revisited these categories in his Retrospective Epilogue (1989: 359–65), where he sought to establish central and non-central modes of meaning through the criteria of formality (“whether or not the relevant signification is part of the conventional meaning of the signifying expression”) and dictiveness (“whether or not the relevant signification is part of what the signifying expression says”). Thus, when a speaker says “p; on the other hand, q” in the absence of any intended contrast of any kind between p and q, “one would be inclined to say that a condition conventionally signified by the presence of the phrase ‘on the other hand’ was in fact not realized and so that the speaker had done violence to the conventional meaning of, indeed had misused, the phrase ‘on the other hand’.” Crucially, however, “the nonrealization of this condition would also be regarded as insufficient to falsify the speaker's statement” (Grice Reference Grice1989: 361). Thus, formality without dictiveness yields conventional implicature. (As for dictiveness without formality, a plausible candidate is the pragmatically enriched content of relevance theorists, the truth-conditional pragmatics of Recanati Reference Recanati2001.)

In uttering a given sentence in a given context, the speaker may intentionally communicate more than one truth-evaluable proposition, but these communicated propositions do not necessarily have equal status; in They are poor but happy, the conjunctive content is truth-conditionally relevant while the contrastive content is not. Yet but and and are not synonyms. Conventional implicatures constitute part of encoded content but not part of truth-conditional content per se; their falsity does not project as falsity of the expression to which they contribute (cf. Barker Reference Barker2003). What they contribute is use-conditional meaning (Kaplan Reference Kaplan1999, Gutzmann Reference Gutzmann2008).

Besides detachability and non-cancelability (Grice Reference Grice1961, Reference Grice1989), additional diagnostics for conventional implicatures include their tendency to project out of embedded contexts, their immunity to certain kinds of objection, and their contextual variability or descriptive ineffability (Potts Reference Potts2007b, Horn Reference Horn2008). The last property is the difficulty of pinning down the precise contribution of but (contrast? unexpectedness?) or even (relative or absolute? unlikelihood or noteworthiness?) and in the intention prompting the use of a second person “familiar” (T[u]) or “formal” (V[ous]) pronoun (T can be affectionate, presumptuous, comradely, or condescending; V can be polite, aloof, diplomatic, or hostile). Ineffability has a plausible source: modulo the well-known problems associated with vagueness, it is plausible that the edges of truth-conditional meaning should for the most part remain discrete, while inconsistency in the mental representation of non-truth-conditionally relevant content is less pernicious. If you know that my use of vous rather than tu signals some aspect of formal respect, distancing, or lack of intimacy, my precise motives can be left underdetermined, but if you don't know whether I’m using a second person or third person pronoun, the indeterminacy is more costly.

For Frege and Grice, identifying the class of conventional implicature-licensing constructions – scalar particles, speaker-oriented sentence adverbs, epithets and slurs, prosodic features, evidential markers, “affected” pronominals, word order effects – serves to characterize them in terms of what they are not: they do not affect the thought or the truth-conditionally relevant meaning of a given expression, and at the same time they are not derivable from general principles of rationality. While they share the former property with conversational implicatures, they differ crucially from them in the latter respect.

4.6 Implication and speaker meaning

Whether conversationally or conventionally triggered, implicatures are generally understood to constitute a proper subset of speaker-meant implications. As we have seen, the assimilation of implicature to inference has been deplored as an instance of the imply/infer confusion. But just as the latter turns out to be more complex than meets the eye, so too the subsumption of conversational implicature within the category of speaker meaning is not entirely straightforward. Saul (Reference Saul2002a) has argued that the full range of Grice's remarks on the topic suggests that we need to allow for a category of audience-implicature along with the traditional utterer-implicature. Others (e.g. Bach Reference Bach2001, Reference Bach2006c), while acknowledging the role of the speaker's expectations about the inferences that the hearer can reasonably be expected to draw, defend the classical view that implicatures constitute part of speaker meaning.

In this respect, it is worth touching on evidence that imply in ordinary language use is not always definable in terms of speaker intention. A Google search on the string didn't mean to imply returns 929 valid hits (retrieved 16 June 2010) – “The President didn't mean to imply that AARP supports current health care legislation”, “Bolivian President Evo Morales Didn't Mean to Imply Straight Guys Go Gay Over Chicken”, etc. – with the sense of ‘x didn't intend in saying p that y should infer q’. So we can unintentionally imply propositions, whether or not we can unintentionally implicate them. This is reinforced by the 785 googla instantiating unintentionally imply/implied – “I’m sorry if I unintentionally implied here anything that touched your feelings”, “I realized that I might have unintentionally implied that no Christian would feel right about attending a pagan ceremony” – and the 705 hits for accidentally imply/implied, including “When you accidentally imply she's fat – send chocolates instead” and “Did you know you're supposed to say ‘best wishes’ to the bride instead of ‘congratulations’, lest you accidentally imply that the bride won her groom through trickery or deceit? Like she most likely did. But still.”¹⁴

Ideally, imply and infer would map respectively into what the speaker intends and what the hearer grasps. Most likely, that would make for a simpler pragmatics. But still.

5.1 Introduction

Speaker intentions made their way into contemporary pragmatics through three different but interrelated routes. One of them dates back at least to medieval philosophy and the inquiries into the logic of modal contexts, leading to the study of intentionality. Another begins with ordinary language philosophy of the mid-1950s and the attempts to define meaning through language use, which in turn led to the employment of the concept of speaker's intended effect of an act of communication. The third, and arguably the most influential route, was that of the attempts to rescue formal semantic analyses by employing a concept of meaning that would incorporate not only the truth-conditional content but also the intended implicated messages, forming the overall concept of communicated content. This introduction to intentionality and intentions is structured as follows. In section 5.2, we present the philosophical idea of intentionality and explain the relation between intentionality of mental states and linguistic intentionality conveyed through acts of communication. In section 5.3 we move to the role of intentions in communication, focusing on the second and third routes mentioned above, starting with Grice's and Searle's views, attempting also a typology, explanation and exemplification of various kinds of speaker intentions distinguished in the literature. Finally, we discuss the question as to where intentions are located, contrasting cognitive, interactional and discursive perspectives on this issue, and conclude in section 5.4 with a brief assessment of the advantages and weaknesses of utilising intentions and intentionality in pragmatic theory.

The first question to ask is why linguists would want to appeal to a concept as murky as intentions and award it the status of an explanandum in pragmatic theory. The main reason is that the possession of the concept of intentional action is crucial for understanding human behaviour. As Anscombe (Reference Anscombe1957: 83) says in her seminal book Intention,

The notions of intention and intentionality have since been deployed in a multitude of ways in explaining speaker meaning (Grice Reference Grice1957, Reference Grice1989), speech acts (Searle Reference Searle1969, Reference Searle1983), the development of language (Searle Reference Searle, Kecskes and Horn2007), social development (Tomasello et al. Reference Tomasello, Carpenter, Call, Behne and Moll2005), and the cognitive processes underlying action and meaning interpretation (Bara Reference Bara2010; Pacherie Reference Pacherie, Pockett, Banks and Gallagher2006, Reference Pacherie2008), to name just a few. However, this proliferation has generated challenges for the conceptual and theoretical status of intentionality and intention in pragmatics, as we will discuss.

5.2 Intentionality

Intentionality has a long and celebrated tradition in philosophy. Coming from the Latin term intendere, meaning aiming in a certain direction, directing thoughts to something, on the analogy to drawing a bow at a target, it has been used to name the property of minds of having content, aboutness, being about something (Duranti Reference Duranti1999; Harland Reference Harland1993; Lyons Reference Lyons1995; Jacob Reference Jacob2003; Nuyts Reference Nuyts, Verschueren, Östman, Blommaert and Bulcaen2000; Smith, D. W. Reference Smith2008; Jaszczolt Reference Jaszczolt1999; Woodfield Reference Woodfield and Asher1994). In other words, it means the ability of minds to represent objects, properties or states of affairs. In medieval philosophy, forms of beings consisted of so-called esse naturale, or natural objects, and esse intentionale, or concepts, mental images or thoughts. It is the latter that we are interested in here.

The very idea of intentionality dates back to ancient philosophy, as far back as the fifth century BCE and Parmenides of Elea. It was taken up by Aristotle and the Stoics (Caston Reference Caston2007), and was subsequently extensively used in medieval doctrines of knowledge and revived in nineteenth-century phenomenology by Brentano (Reference Brentano1874) and Husserl (Reference Husserl1900–1901), and later by Meinong, Twardowski, Heidegger, Sartre, Merleau-Ponty and many others. By phenomenology we mean the study of the way in which things (phenomena) are presented in consciousness, or generally the study of forms of conscious experience from the first-person point of view. Mental attitudes such as belief, desire or want are intentional in that they are about something, they have an object. For phenomenologists, things exist as physical objects, but they also have an intentional existence, so to speak, in acts of consciousness. Their intentional existence is revealed in our mental states or acts, as for example in (1) or (2).

1. (1) I am thinking about my holiday in Australia.

2. (2) I hope to meet Peter Carey one day.

In the case of hallucinations, the existence is only intentional.

Brentano and Husserl developed intricate arguments concerning the meaning of such mental existence, in particular the question as to whether there are intentional objects that are internal to the thinker's mind. This discussion, albeit fundamental in the study of phenomenology, can only be sketched here. For Brentano, objects of conscious mental acts had the status of mental entities. On this view, the act does not consist of a relation between its subject and an intersubjectively identifiable object but rather can be spelled out in a so-called adverbial theory: ‘Tom sees a horse’ amounts to a property of Tom's ‘seeing horsely’, so to speak (see Smith and Smith Reference Smith, Smith, Smith and Smith1995). However, the adverbial theory comes with a considerable limitation. It does not provide for the fact that our acts of consciousness are about things in the world, real world properties, or states of affairs. It was therefore subsequently replaced with the relational theory by Twardowski and Meinong. In the next logical step, Husserl rejected mental objects tout court. Current theories of intentionality adopt such a relational view.

Now, this is not to say that where the real object does not exist, ‘made-up’ substitutes are never needed in a theory of natural language meaning. To give one pertinent example, Clapp (Reference Clapp2009) discusses so-called negative existentials, i.e. sentences of the type in (3) and (4) (from Clapp (Reference Clapp2009: Reference Clapp1422)).

1. (3) The Loch Ness monster does not exist.

2. (4) Nessie does not exist.

As he puts it, the problem with negative existentials is that they presuppose existence and immediately deny it. We can, of course, stipulate that there is something real to which one refers in (3) and (4), à la Meinong, or say that there is no existential presupposition there, à la Russell. On the modern version of the classical phenomenology, mental objects are intentional, mind-independent entities; following McGinn (Reference McGinn2000), we can introduce the theoretical construct of representation-dependent, intentional objects. Then we replace the dichotomy ‘existent – non-existent’ with a new dichotomy ‘existent – intentional’ and thereby vindicate the latter as a theoretical construct.

Intentional content has been the subject of discussions between so-called neo-Fregeans and neo-Russellians (Recanati Reference Recanati1993; Siewert Reference Siewert2006). In the Anglo-American tradition in semantics and pragmatics we often think of the analytic tradition as separate from continental European psychologism with its emphasis on consciousness, mental states, first-person psychology and the like. However, it is worth remembering that the most ground-breaking achievements in phenomenology and in analytic philosophy took place at the same time, at the end of the nineteenth century and the beginning of the twentieth century, and that they are much more interrelated than they are usually taken to be. Frege's (Reference Frege1892) notion of sense, whose offshoots have been so influential in current semantic theory of the Anglo-American tradition, is a direct predecessor of Husserl's phenomenological theory of meaning and object. We briefly attend to this interrelation in what follows.

Husserl held various consecutive ideas of meaning, culminating in his view from Ideen (1913) according to which meaning is contained in the objective content of consciousness, called noema. An act of consciousness is directed at an object determined by noemata. Here Frege's concept of sense can be seen as a clear precursor for Husserl's idea of meaning as distinguished from objects, and hence subsequently also his notion of noema. However, their respective distinctions between meaning and object, and sense and reference, led them in very different directions. Whereas Husserl and other phenomenologists concentrated on experience, Frege and analytic philosophers concentrated on developing a theory of meaning making use of intersubjective generalisations over ‘meaning-giving mental acts’ such as Frege's sense.

But this departure from psychologism does not reflect a sudden split. It happened gradually through various reanalyses of intentionality and subsequently ascribing intentionality to linguistic acts. So, the question to ask is how the theory of meaning became dissociated from Husserlian meaning as noema, that is from meaning attached to mental acts, discussed above. First of all, for Husserl from his Ideen period (1913), it was language, rather than a particular mental act of a language user as it was claimed in earlier phenomenological accounts, that was a carrier of meaning. Already in Husserl's work of this period meaning acquires the status of an abstraction, noematic meaning, attached to relevant mental acts. There is one step from there to freeing theory of meaning from the constraints of psychology. And this step was taken by Frege and his battle against the ‘corrupting intrusion’ of psychology in logic and mathematics, and thereby subsequently in natural language meaning.

Frege's sense (Sinn), as contrasted with reference (Bedeutung), saved formal semantics from the problem of substitutivity of coreferential expressions. For example, since it is true that Hilary Mantel is the author of Wolf Hall, it should be possible to substitute ‘Hilary Mantel’ for ‘the author of Wolf Hall’ as in (6), preserving the truth of the sentence.

1. (5) The author of Wolf Hall is visiting Cambridge this spring.

2. (6) Hilary Mantel is visiting Cambridge this spring.

However, although Hilary Mantel is identical to the author of Wolf Hall, sentence (7) differs substantially from (8); (7) is informative while (8) is not.

1. (7) Hilary Mantel is the author of Wolf Hall.

2. (8) Hilary Mantel is Hilary Mantel.

These two ways of referring to the winner of the 2009 Man Booker Prize come with different ways of thinking, ways of understanding or, to use the celebrated phrase, different modes of presentation of the referent. The senses we grasp when we understand these two expressions are different. Now, when we embed (7) and (8) in the contexts expressing mental attitudes, using intentional verbs such as ‘believe’ or ‘doubt’, Frege's concept of sense acquires a new importance for semantic theory. In (9) and (10), using Frege's own explanation, the sense of ‘Hilary Mantel’ plays the role of the referent.

1. (9) Harry believes that Hilary Mantel is the author of Wolf Hall.

2. (10) Harry believes that Hilary Mantel is Hilary Mantel.

Analogously, in (11), the sense of the definite description ‘the author of Wolf Hall’ plays the role of the referent. The description refers not to its customary reference, but instead to the sense. Sense, however, is more than the speaker's way of thinking; it is an intersubjective way of thinking. It can function as the speaker's own mode of presentation of the referent, but equally it can function as someone else's way of thinking about this referent.

1. (11) Harry believes that the author of Wolf Hall is a man.

This objectivity of sense shifted intentionality from the domain of the individual and the mental to the domain of the external. For Kripke (Reference Kripke1972) and Putnam (Reference Putnam1975), proper names and natural kind words acquire reference through the causal link with the world. The topic of the externalist/internalist debate, albeit very important, is tangential to the present discussion of intentionality vis-à-vis intentions and will not be pursued here. Instead, we will now focus on one aspect of the debate, namely the role of intentionality in the theory of meaning.

Let us now return to the topic of the ‘antipsychologism’ of analytic phil-osophy. Frege can be credited with developing a new concept of logic. In his Begriffsschrift (Frege Reference Frege and Bynum1879), he developed an analysis of the logical form of sentences in terms of predicates and arguments, where the reference of a predicate is a function from objects to truth values. This development marked the end of the era of psychological logic that studied thought processes and subjective mental representations. According to Frege, the object of study of logic is not the agent who uses the rules of inference, but the rules of logical inference themselves, and thereby the languages of logic themselves. In Grundlagen der Arithmetik, Frege (Reference Frege1884) offers a new, depersonalised stance on truth, definitions, logic, mathematics and indirectly on natural language semantics. He says that ‘[t]here must be a sharp separation of the psycho-logical from the logical, the subjective from the objective’ (p. 90). The ways of thinking about objects must not be confused with the objects themselves. Instead of focusing on someone's ‘thinking that something is true or valid’, we now focus on truth and validity as such. To give further examples, in his Grundgesetze der Arithmetik (Frege Reference Frege1893: 202), he calls the effect of psychology on logic a ‘corrupting intrusion’ because it has to be emphasised that ‘being true is quite different from being held as true’. In Logic, he stresses that ‘Logic is concerned with the laws of truth, not with the laws of holding something to be true, not with the question of how people think, but with the question of how they must think if they are not to miss the truth’ (Frege Reference Frege and Hermes1897: 250).

This ban on psychologism has never been lifted since.¹ Frege's function/argument analysis, juxtaposed with Tarski's semantic definition of truth, led to the development of formal semantics of natural languages. Now, this rejection of psychology can easily be taken to suggest that the very idea of intentionality had to be poured out with the bath water as well. After all, what we have here is Frege's forceful rebuttal of Husserl's stance on logic as a study of mental processes.² As can be seen from the following section, however, intentionality need not necessarily pertain to mental states and proves to be a useful theoretical tool that can easily be dissociated from ‘psychologising’.

The follow-up question to ask is whether intentionality is a property of mental states only, or whether it can also be construed as a property of other objects. The answer gleaned from subsequent theorising is definitely positive. For the current purpose, however, we will not be concerned with the intentionality of human organs and limbs, or intentionality of information-processing systems discussed by Millikan and Dretske (see e.g. Lyons Reference Lyons1995; Jacob Reference Jacob2003). Neither will we be interested in Fodor's (e.g. Reference Fodor1975, Reference Fodor1981) discussion of intentionality as a feature of the brain. What we will focus on is the intentionality of linguistic expressions.

For Searle (Reference Searle1983, Reference Searle1992b), our beliefs and intentions have intrinsic, basic intentionality, while linguistic expressions have derived intentionality in the sense that the meaning of acts of speech can be analysed in terms of intentional states, such as belief or intention. In other words, Searle says that the mind ‘imposes’ intentionality, so to speak, on linguistic expressions in that the basic intention to represent is responsible for the derived intention to communicate. The intentionality of the mental state that underlies the act of communication bestows on that act so-called conditions of satisfaction. In brief, beliefs have intrinsic intentionality, while utterances have derived intentionality (Searle Reference Searle, Lepore and van Gulick1991: 84), or ‘I impose Intentionality on my utterances by intentionally conferring on them certain conditions of satisfaction which are the conditions of satisfaction of certain psychological states’ (Searle Reference Searle1983: 28).

What Searle proposes here is a so-called double level of intentionality. Mental states such as beliefs, wishes or hopes impose conditions of satisfaction on the expressions of these states. These conditions, in turn, play a major role in determining the meaning of the linguistic expressions. For example, a request may inherit conditions of satisfaction from the speaker's wish or desire. An important question arises at this point, namely what exactly does it mean to impose, or confer, conditions of satisfaction by one object on another? How can one transfer them from a wish, a mental state, to a request, a linguistic object? Harnish (Reference Harnish and Burkhardt1990: 189) points out that an intention to confer them does not suffice because there must be a constraint that one cannot intend and fail. Furthermore, there must be a restriction on types of objects from and to which intentionality can be transferred. As was proposed in Jaszczolt (Reference Jaszczolt1999: 106), we should divert our attention from the conferment as such and direct it instead to the fact that the same conditions of satisfaction pertain to the mental and to the linguistic. Once we pose the problem in this way, what remains is to take a stance on the relation between the mental and the linguistic, perhaps arguing as follows. Since the double level is supposed to characterise speech acts themselves rather than the relation between a speech act on the one hand and a mental state on the other, then we should think that speech acts have both basic intentionality, qua externalisations of mental states, and derived intentionality, qua linguistic objects. Since language is one of the vehicles of mental states, this seems to be a natural way of explaining the mysterious and rather unfortunately named ‘double level’ of intentionality.³ By this reasoning, intentionality is always an intrinsic rather than a ‘conferred’ property.

Next, just as intentionality is a property of linguistic acts, so having intentions is a property of their owners. The latter is the topic to which we now turn.

5.3 Intentions in communication

5.3.4 Locating intentions

According to the received view of meaning as the intended expression of thoughts, intentions pertain to the mental states of speakers. Recent work in social neuroscience has begun exploring the neural correlates of our ability to attribute mental states, including intentions, to others (Walter et al. Reference Walter, Adenzato, Ciaramidaro, Enrici, Pia and Bara2004: Reference Walter, Adenzato, Ciaramidaro, Enrici, Pia and Bara1854). This capacity to attribute mental states is termed theory of mind (ToM), with it being assumed that without ToM ‘other people's behaviour would be meaningless from a third-person perspective: behaviour would be observed, but the meaning of actions would not be understood’ (Ciaramidaro et al. Reference Ciaramidaro, Adenzato, Enrici, Erk, Pia, Bara and Walter2007: 3111). Experiments combining brain imaging with various kinds of language use prompts have found that a distributed neural system underlies the ToM mechanism, including the right and left temporo-parietal junctions (right TPJ and left TPJ), the precuneus, and the medial prefrontal cortex (MPFC) (Ciaramidaro et al. Reference Ciaramidaro, Adenzato, Enrici, Erk, Pia, Bara and Walter2007: 3105; Enrici et al. Reference Enrici, Adenzato, Cappa, Bara and Tettamanti2011; Walter et al. Reference Walter, Adenzato, Ciaramidaro, Enrici, Pia and Bara2004: Reference Walter, Adenzato, Ciaramidaro, Enrici, Pia and Bara1854). The detection of agency, for instance, has been tied to neural activity in the superior temporal sulcus (STS) (Frith and Frith Reference Frith and Frith2003), while the representation of our own and other people's mental states (including intentions) has been tied to activity in the medial prefrontal cortex (MPFC) (Walter et al. Reference Walter, Adenzato, Ciaramidaro, Enrici, Pia and Bara2004: Reference Walter, Adenzato, Ciaramidaro, Enrici, Pia and Bara1854).

Further work on intentions specifically has established that different parts of this distributed neural network are activated depending on the kind of prior intention involved. This lends some empirical support to the hitherto conceptual distinctions discussed in section 5.3.2, including that between private and social intentions, and within the category of social intentions, between communicative intentions and prospective social intentions, the latter being a form of higher-order intention (Ciaramidaro et al. Reference Ciaramidaro, Adenzato, Enrici, Erk, Pia, Bara and Walter2007: 3105; see also Saxe Reference Saxe2006; Walter et al. Reference Walter, Adenzato, Ciaramidaro, Enrici, Pia and Bara2004), and also between the representation of individual and we-intentions (Becchio and Bertone Reference Becchio and Bertone2004: 132).

Attempts to model the understanding of intentions (Bara Reference Bara2010; Pacherie Reference Pacherie, Pockett, Banks and Gallagher2006, Reference Pacherie2008) have also been given empirical support. Becchio et al. (Reference Becchio, Adenzato and Bara2006), for instance, have argued that the recognition, attribution and representation of action intentions can be traced to different forms of neural activity. They hypothesise that the recognition of the intentions of others is partly based on the same areas of the brain that are activated when one (intentionally) performs actions oneself. The neural basis of this has been argued to be ‘mirror neurons’ (i.e. neurons that fire during action execution and action observation) in the premotor cortex (Fogassi et al. Reference Fogassi, Ferrari, Gesierich, Rozzi, Chersi and Rizzolatti2005; Rizzolatti and Craighero 2004; Rizzolatti and Fabbri Destro 2007), although mirror neurons do not in themselves provide the basis for different forms of agentive understanding and shared intentionality (Pacherie and Dokic Reference Pacherie, Pockett, Banks and Gallagher2006).²⁰ The attribution of intention to a particular agent is traced in Becchio et al.'s (2006) experiments to another area of the brain, the inferior parietal lobe, while the representation of prior intentions, particularly those which are social, has been traced to the anterior paracingulate cortex (Walter et al. Reference Walter, Adenzato, Ciaramidaro, Enrici, Pia and Bara2004). Recent work has also shown that a common neural network is employed when subjects are comprehending communicative intentions, no matter whether the prompt is linguistic or gestural (Enrici et al. Reference Enrici, Adenzato, Cappa, Bara and Tettamanti2011). Although the area of study is in its relative infancy, then, social neuroscientists have begun tracing hypothesised mental states (i.e. intentions and inferences about intentions) to specific neural activity, lending support to the kinds of distinctions made in pragmatics between different kinds of intentions.

However, work in psycholinguistics on the comprehension and attribution of intentions suggests that while these may indeed have neural correlates, speakers consistently over-estimate their ability to project intended meanings onto addressees (Keysar Reference Keysar1994b, Reference Keysar2000; Keysar and Henly Reference Keysar and Henly2002). Moreover, hearers often do not routinely consider what the speaker knows (i.e. common ground) or other mental states in interpreting what has been said (Keysar Reference Keysar2007, Reference Keysar, Kecskes and Mey2008). This fundamental egocentrism in the early stages of processing language suggests that due caution should be given to interpreting results of experiments attesting to neural correlates of intentions. While no one would suggest that speakers do not at times have intentions motivating them to say things, or that hearers do not at times make conscious inferences about such intentions, the question is the extent to which such an explanation is sufficient to account for speaker meaning and communication (Haugh Reference Haugh, Kecskes and Mey2008c: 52; Reference Haugh2009: 93). The evidence from psycholinguistics suggests that while a (prior) intention-based account of speaker meaning may be necessary, it may not be sufficient.

Work from an interactional perspective also attests to the difficulty of locating intentions relative to meanings in discourse (rather than simply relative to utterances). Haugh (Reference Haugh2008c, Reference Haugh2009), for instance, argues that the intentions hypothesised to underlie implicatures are temporally, ontologically and epistemologically ambiguous when the analyst attempts to trace them in actual interactional data. In more closely tracing intentions in conversational interaction it becomes apparent that intentions can be characterised as being ‘emergent’, as both the speaker and the hearer jointly co-construct understandings of what is meant (Clark H. Reference Clark1997; Gibbs Reference Gibbs1999: 38, Reference Gibbs, Malle, Moses and Baldwin2001; Haugh Reference Haugh2007c: 95, Reference Haugh2008c, Reference Haugh2009; Kecskes Reference Kecskes2010a: 60–61). Kecskes (Reference Kecskes2010a), for instance, argues that John's initial intention to give Peter a chance to talk about his trip is not realised in the excerpt below.

1. (23)

   John: Want to talk about your trip?

   Peter: I don't know. If you have questions…

   John: OK, but you should tell me…

   Peter: Wait, you want to hear about Irene?

   John: Well, what about her?

   Peter: She is fine. She has…well…put on some weight, though

   (Kecskes Reference Kecskes2010a: 60).

Kecskes suggests that John's original intention is sidelined by Peter talking about Irene, perhaps because he thinks John might want to know about her (being his former girlfriend). He argues that ‘it was the conversational flow that led to this point, at which there appears a kind of emergent, co-constructed intention’ (Kecskes Reference Kecskes2010a: 61, original emphasis).

These ‘emergent’ intentions have interesting parallels with the notion of we-intention discussed in section 5.3.2. However, Haugh (Reference Haugh, Kecskes and Mey2008c) argues that the relatively static characterisation of we-intentions does not do full justice to the contingent and emergent nature of the inferential work underlying joint, cooperative activities, including conversational interaction. The key difference between the notions of we-intention and emergent intention is that the former assumes a monadic view of cognition, involving ‘the summative sequence of individual cognitive activities’ (Arundale and Good Reference Arundale, Good, Fetzer and Meierkord2002: 124), while the latter builds upon a dyadic view of cognition:

Each participant's cognitive processes in using language involve concurrent operations temporally extended both forward in time in anticipation or projection, and backwards in time in hindsight or retroactive assessing of what has already transpired. As participants interact, these concurrent cognitive activities become fully interdependent or dyadic. (Arundale and Good Reference Arundale, Good, Fetzer and Meierkord2002: 122)

Yet while there are differences in their underlying frameworks, the view in philosophical and cognitive pragmatics that we-intentions cannot be reduced to individual intentions due to the sharedness requirement being cognitively implausible (Becchio and Bertone Reference Becchio and Bertone2004: 127–8), and the complementary view in interactional pragmatics that models of individual intentions are ‘formally incapable of explaining the non-summative effects or emergent properties observable when individuals are engaged in interaction’ (Arundale Reference Arundale2008: 243), both strongly suggest that meaning cannot necessarily always be traced analytically to prior intentions of individual speakers (including their communicative intentions). Instead, the interpretation of meaning is ‘doubly-dynamic’ in the sense that it is ‘created in-between the interlocutors and the hearer may also be given freedom to create assumptions rather than recover them’ (Jaszczolt Reference Jaszczolt1999: 76; see also Hirsch Reference Hirsch2010). Gauker (Reference Gauker2001, Reference Gauker2003, Reference Gauker2008) has also argued that the role of situational inferences has been vastly underplayed in communicative intention-based models of meaning and communication.

However, such views should not be simplistically interpreted as constituting an anti-intentionalism stance as some have recently assumed (Åkerman 2009; Montminy Reference Montminy2010). Instead, there is growing evidence that we need to make clearer distinctions between speaker (intended) meaning, which pertains to the subjective processing domain at the utterance level, and ‘joint meaning’, which pertains to the interpersonal domain at the discourse level (Carassa and Colombetti Reference Carassa and Colombetti2009; Kasper Reference Kasper, Bardovi-Harlig, Felix-Brasdefer and Omar2006; Kecskes Reference Kecskes2008, Reference Kecskes2010a; Kriempardis Reference Kriempardis2009: 186). Different types of intentions arguably have different roles to play relative to these different types of meaning. Instead of remaining committed to a hardline intentionalist or anti-intentionalist stance, greater dialogue between those with different views on intention lies at the heart of advancing our understanding of meaning and communication (Haugh Reference Haugh2008a).

We have discussed thus far how intentions can be located both in the minds of individuals, as well as more diffusely in interaction, where they emerge ‘between the mind and the world’ (Jaszczolt Reference Jaszczolt1999: 117). There is, however, an additional level at which intentions can be productively located, namely, the social (or societal) level of analysis. The focus here is on deontological aspects of intention and intentionality. Philosophers have conceptualised this as commitment to undertake we-intended actions (Gilbert Reference Gilbert2009), for instance, or commitment to the truth conditions of utterances (Searle Reference Searle, Kecskes and Horn2007: 33–4). The notion of speaker accountability in ethnomethodological conversation analysis (Garfinkel Reference Garfinkel1967; Sacks Reference Sacks1992 [1964]: 4–5), where interlocutors hold themselves and others normatively accountable for the meanings that arise from what is said (Heritage Reference Heritage1984), can also be productively explored in regards to both speaker intentions and intentionality (Arundale Reference Arundale2008; Edwards Reference Edwards2006, Reference Edwards2008; Haugh Reference Haugh, Kecskes and Mey2008c, Reference Haugh2009). For instance, the ways in which speakers are held accountable to meanings through topicalising intentions, and how this intersects with interpretative and socio-cultural norms can be explored (Haugh Reference Haugh2008b, Reference Haugh, Kecskes and Mey2008c).

Holding speakers accountable for what they are understood by others to have implied can also enter into broader societal debates, as argued by Haugh (Reference Haugh2008b) in an analysis of the discursive dispute arising in the Australian media as to what was intended by comments in regards to the status of women made by a Muslim cleric.²¹ In the following excerpt, the cleric is being interviewed in the controversial wake of the publication of his sermon.

Martin (the interviewer) first asks in lines 1–2 why Hilali would say something offensive when he holds a position of responsibility in the Australian Muslim community (the grand mufti), to which Hilali responds by claiming he always speaks honestly, implying that he does not necessarily always say what others want him to say (line 3). Martin then invokes the folk view of meaning as residing in words (Bilmes 1986), in arguing that Hilali is responsible for how people understand him (lines 6–8), and that one cannot be absolved from this responsibility by claiming one was ‘misunderstood’ or ‘misinterpreted’ (lines 4–6). This stance, however, is implicitly rejected by Hilali in line 9, when he claims he stands behind the ‘correct’ interpretation of what he implied, in this case, what he intended by his comments. The deontological dimensions of speaker meaning were thus clearly topicalised in this particular interview.

In considering the question of where intentions can be located, then, it has become apparent that this concept is deployed in pragmatics for a number of different analytical purposes. In philosophical pragmatics it is used in accounting for how speakers mean things (communicative intentions) or undertake joint activities (we-intentions), although there are varying levels of commitment to the psychological reality of those intentions (Jaszczolt Reference Jaszczolt1999). In cognitive pragmatics there is a clearer commitment to the assumption that the recognition and attribution of (communicative) intentions underlies communication, but there is also consideration of a much more fine-grained range of different types of intentions (including the distinction between prior intentions versus intentions-in-action), with work attempting to correlate neural activity with such distinctions. Here, intentions tend to be conceptualised as being firmly located in the minds of speakers. In contrast, in interactional pragmatics the focus is on examining the relationship between speaker (intended) meaning and joint or interactionally achieved meanings, with the notion of ‘emergent intention’ (Kecskes Reference Kecskes2010a) or ‘emergent intentionality’ (Haugh Reference Haugh, Kecskes and Mey2008c, Reference Haugh2009) sometimes being deployed to account for the latter. Here, intentions and inferences are treated as contingent (and non-summative), arising in the course of interaction, and thus better traced with reference to dyadic views of cognition (Arundale and Good Reference Arundale, Good, Fetzer and Meierkord2002) rather than individual minds. Finally, in more discursive approaches to pragmatics, the analytical focus is on the normative work intention does when deployed in discourse or interaction, with a particular emphasis on how speaker commitment or accountability can be disputed. In these approaches, intentions can be found to be diffused across social networks, ranging from dyadic units through to larger social groups.

5.4 Concluding remarks: intentions as ‘creatures of darkness’ or a useful tool?

We are now in a position to address the methodological question as to how pragmatic theory, aspiring to high predictive power, can be founded on intentions and intentionality – theoretical notions which are inherently so imprecise and, moreover, possibly not directly empirically testable. In other words, this lack of testability may prove not to be a fleeting state of affairs but an inherent property of intentions. In answering this question one has to point out that the advantages seem to outweigh the shortcomings. We attempt to list here a few arguments in favour of an outlook that maintains the importance of intentions for theorising in pragmatics.

1. (1) If we ban intentions from pragmatics, we have to use a substitute theoretical tool such as default rules of inference, semanticised pragmatic relations between sentences as in dynamic approaches to meaning, or other similar solutions, e.g. constraints of optimality-theory pragmatics. None of these alternative tools has comparable predictive power as far as speaker meaning is concerned. Instead, we are forced to change the object of study of pragmatics from, so to speak, speaker meaning ‘whatever means the speaker may have used to convey it’ to speaker meaning ‘modelled on the fairly probable semantic patterns’.²²

2. (2) In the current state of experimenting in neuroscience, it seems very unlikely that intentions can remain creatures of darkness. Instead, they are being correlated with neural activity as discussed in the preceding section. Intentions in communication derive their theoretical status from intentionality of consciousness. The more we know about intentionality in the brain, the more will we know about intentions. The structure of the explanation is already there in the form of Gricean pragmatics; the scientific flesh is being provided as cognitive science progresses.

3. (3) Language is a vehicle of thought and pragmatic theory of its use in communication should derive from theories of thought. In order to theorise expression meaning (word/sentence meaning), the basic intentionality of thought needs to be taken into account. In this way, the extent to which expression meaning can be productively defined in terms of speaker meaning (intentions), as originally proposed by Grice, may be further explored.²³

4. (4) The notion of intention (and indeed intentionality) is already being productively deployed in many different ways in pragmatics. While this proliferation can at times create analytical confusion, it is also no doubt reflective of the metaphorical power of intentions and intentionality in advancing our understanding of how speakers mean things through the use of language.

We suggest, therefore, that while intentions may be difficult to pin down, it is clear that disciplines do not advance by avoiding slippery questions, particularly when they lie at their very foundations, as do the concepts of intention and intentionality in pragmatics. Ultimately, it is only through refining or even discarding certain views and developing alternatives that we will continue to advance in our theorisation and analysis of meaning and communication.

6.1 Introduction

Two-dimensional semantics in its most general characterization is a semantics that recognizes two kinds of content: narrow content and wide content.¹ Narrow contents can take the form of linguistic meanings (e.g., functions from context to content) or descriptive Fregean contents. Wide content is a set of possible worlds or a structured Russellian proposition consisting of properties and/or physical objects. On some approaches (for instance, Kaplan's and Stalnaker's), “narrow content” is not semantic content (it is not what is shared in communication) but belongs to what Stalnaker calls “the metasemantics.” On other approaches (e.g., Chalmers and Jackson), the narrow content is (part of) the semantic content.

Kaplan (Reference Kaplan, Almog, Perry and Wettstein1989a) originally introduced two-dimensional semantics in order to account for the semantics of context-sensitive expressions, primarily indexicals and demonstratives. Names, indexicals, and demonstratives have often been treated as having as their semantic content the individuals to which they refer. But indexicals and demonstratives do not acquire their content in the same way as names. Unlike names, as traditionally construed, their content varies across worlds, times, speakers, and locations. This difference between names and indexicals prompted Kaplan to introduce a further layer of meaning, viz. a function that takes an expression from worlds, times, speakers, or locations to its content. This function can then be either constant (names) or variable (indexicals and demonstratives).

Two-dimensional semantics has subsequently been employed by Robert Stalnaker (e.g. Reference Stalnaker and Cole1978), David Chalmers (e.g. Reference Chalmers1996, Reference Chalmers, Garcia-Carpintero and Macia2006), Frank Jackson (e.g., Chalmers and Jackson Reference Chalmers and Jackson2001), and others as a tool for explaining how sentences that have necessary contents can be cognitively informative and how sentences that are cognitively uninformative can be contingent, and David Chalmers and Frank Jackson have developed the approach further in debates about the role of conceptual analysis and reductive explanation.

Here I will introduce and assess the notion of context-sensitivity presented by the various two-dimensional frameworks, with a special focus on how it relates to the notion of cognitive significance and whether it includes an intuitively plausible range of expressions within its scope. I will argue that the two phenomena (viz. context-sensitivity and cognitive significance) are, to some extent, inseparable. I will conclude with a discussion of the prospects of using epistemic two-dimensional semantics to account for context-sensitive expressions in dynamic discourse.

6.2 Kaplan

6.3 Stalnaker

6.4 Chalmers

6.5 Dynamic two-dimensional semantics

As it stands, epistemic two-dimensional semantics does not allow for a treatment of dynamic discourse.¹⁰ Consider the following discourse fragment:

1. (4) John is now entering the room, and he is now taking off his hat and is therefore not now wearing a hat, but he is now putting the hat back on, and is therefore now again wearing a hat.

If the conjuncts in (4) are thought of as parallel-asserted, then (4) is a contradiction, not so if (4) is asserted at the right sort of pace in non-parallel fashion. Call a non-parallel assertion a ‘dynamical assertion’. So, a parallel assertion of (4) and a dynamical assertion of (4) have different intensions.

The fact that discourse fragments like (4) exist in natural language suggests a need for a dynamic two-dimensional semantics. Given length considerations I can only briefly sketch how such an approach might proceed. Consider:

1. (5) John₁ is now₁ spotting Susan₂.

Following Irene Heim (Reference Heim1982), let us introduce the notion of a filing system, that is, a system that keeps track of variables, names, and descriptive material introduced by the discourse. Here is an illustration.

Filing system F1:

1. x, y, t₁

2. Now t₁

3. John x

4. Susan y

5. Spot (x, y)

Additions to the discourse give rise to a new system:

1. (6) He₁ is now₂ walking over to her₂.

Filing system F2:

1. (7) And is now₃ starting a conversation with her₂

Filing system F3:

1. (8) She₂ is now₄ talking to a man₃.

Filing system F4:

1. (9) Now₅ he₁ is talking to the man₃ she₂ talked to just a moment ago₄

Filing system F5:

We can introduce a notion of truth of files as follows. Given an understanding of scenarios as centered possible worlds, we can take 1-models to be pairs of a domain D of actual individuals, some of which serve as a center and an interpretation function I.

The 1-interpretation function I maps the non-indexical variables (i.e., discourse referents) of file F to members of D, indexical variables of F to the center of D, and the descriptive material and names (i.e., predicate nominals) of F to properties or relations (or sets of n-tuples) on D. One set of assignments, i^@1, contains as an element the distinguished interpretations: the actual sets of assignments of extensions to the expressions (or the set of actual centered worlds). Let a 1-assignment A in a model M = <D, I> be a mapping of non-indexical variables onto elements of D, and indexical variables onto the centered elements of D. We can then say that assignment A verifies filing system F in M if there is an extension E of A such that the elements satisfy the descriptive material and names at the scenario. A 1-information state is a space of scenarios (interpretations) with the same filing system. A dynamic 1-intension is a sequence of 1-information states. We can then say that discourse fragments express dynamic 1-intensions which are sequences of spaces of scenarios that share a filing system. If we introduce a designated sequence of scenarios, we can then say that a dynamic 1-intension is true at a designated sequence of scenarios just in case each scenario in the designated sequence of scenarios is in the space of scenarios (or interpretations) with the same filing system as the designated scenario.

A different model is needed to model dynamic 2-intensions. A 2-model is a pair of an uncentered domain of actual individuals and an interpret-ation function. The 2-interpretation function I maps the indexical and non-indexical variables and the associated names (if any) of file F to members of D, and the descriptive material of file F to properties or relations (or sets of n-tuples) on D. One assignment, i^@2, is the distinguished interpretation, the actual assignment of extensions to the expressions (or the actual world). Let an assignment A in a model M = <D, I> be a mapping of variables and associated names (if any) onto elements of D. We can then say that assignment A verifies filing system F in M if there is an extension E of A such that the elements satisfy the descriptive material at the scenario.

A 2-information state is a space of worlds with the same filing system. A dynamic 2-intension is a sequence of 2-information states. A dynamic 2-intension is true at a distinguished sequence of worlds just in case each world in the designated sequence of worlds is in the space of worlds with the same filing system as the designated world.

We can give the following validation conditions for universal operators:

So, ‘It is metaphysically necessary that Φ’ is true iff at every 2-interpretation i, i assigns T to Φ, and ‘It is a priori that Φ’ is true iff at every 1-interpretation i, i assigns T to Φ. To illustrate consider (10) and (11).

1. (10) The brightest heavenly object is now₁ Hesperus₁, and the brightest heavenly object is now₂ Phosphorus₂, and Hesperus₁ is identical to Phosphorus₂.

2. (11) It is deeply epistemically possible that: the brightest heavenly object is now₁ Hesperus₁, and that the brightest heavenly object is now₂ Phosphorus₂, and that Hesperus₁ is not identical to Phosphorus₂.

(10) is true iff the dynamic 2-intension it expresses has the truth-value true at the designated sequence of worlds. So, (10) is true if the brightest heavenly object is Hesperus at the worlds that assign the utterance time to the variable associated with ‘now₁’, and the brightest heavenly object is Phosphorus at a designated world that assigns the utterance time to the variable associated with ‘now₂’, and Hesperus is identical to Phosphorus. (11) is true iff there is a sequence of scenarios at which the dynamic 1-intension of the embedded clause is true. This is so if the brightest heavenly object is the brightest object in the evening sky at the scenarios that assign the time in the center to the variable associated with ‘now₁’, and the brightest heavenly object is the brightest object in the morning sky at the scenarios that assign the time in the center to the variable associated with ‘now₂’, and the brightest object in the evening sky in scenarios in the first set is not identical to the brightest object in the morning sky in the scenarios in the second set.

6.6 Conclusion

As we have seen, in Kaplan's original framework, context-sensitive expressions are expressions with a variable character – a linguistic meaning that determines different contents relative to different contextual parameters. This approach to context-sensitivity has much to recommend it but it fails to capture a plausible connection between context-sensitivity, broadly understood, and cognitive informativeness. The two phenomena are not unrelated. Cognitively informative necessities arguably contain what David Chalmers calls “twin-earthable expressions,” linguistic strings that have twin tokens with a different content. Twin-earthable expressions are broadly context-sensitive. When I use ‘water’, the term picks out H₂O. When my twin on Twin-Earth uses ‘water’, it picks out XYZ. On the assumption that we speak the same language, an assumption which is consistent with the definition of Twin-Earthability, ‘water’ has a variable character and hence is broadly context-sensitive. But if ‘water’ is context-sensitive, we have a partial explanation of the cognitive significance of sentences such as ‘water is H₂O’. The sentence, relative to my context, expresses a proposition of the form ‘a = a’, yet owing to my cognitive limitations I may not know what content is determined by the character of ‘water’ at my location. Being told that the character of ‘water’ determines H₂O is informative for me. Thus, if we keep fixed the language spoken, there is a tight connection between twin-earthability and cognitive informativeness.

I have assessed three types of two-dimensional semantic frameworks in terms of how well they account for the connection between cognitive significance and the broader notion of context-sensitivity. One is a diagonalized character approach that takes names and kind terms to have a constant character but allows for the possibility that a diagonal proposition that functions much like a variable character has some explanatory power with respect to the phenomenon of cognitive informativeness. A second approach is to widen the range of expressions that have a variable character to include various noun phrases and descriptions. A third approach is to treat expressions as having epistemic meaning in addition to semantic meaning. I argued that while all three approaches go some way toward explaining the connections between context-sensitivity and cognitive informativeness, the non-epistemic approaches seem to fall short of offering a fully adequate account of cognitively informative necessities and informative contingencies. I concluded by pondering a new problem for the epistemic approaches posed by indexicals in dynamic discourses and sketched a solution.

7.1 The modularity issue

According to a view which used to be standard, and which, for reasons that will soon emerge, I call the modular view, knowledge of a language, and especially semantic knowledge or semantic competence, enables language users to ascribe truth conditions to arbitrary sentences of that language. To be sure, when a sentence is context-sensitive (as most sentences are), it only carries truth conditions ‘with respect to context’; so knowledge of the context is required in addition to knowledge of the language. But (according to the view in question) the context at issue involves only limited aspects of the situation of utterance: who speaks, when, where, to whom and so forth. Given a context thus understood, the rules of the language – e.g. the rule that ‘I’ refers to the speaker – suffice to determine the truth-conditional contribution of context-sensitive expressions. There is no need to appeal, in addition, to pragmatic competence. That is the gist of the modular view.

By ‘pragmatic competence’, I mean the ability to understand what the speaker means by his or her utterance. As Grice emphasised, 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 recognised by the hearer. Pragmatic competence is needed to determine what the speaker means on the basis of what she says; but what the speaker says is supposed to be autonomously determined by the semantics (with respect to context), irrespective of the speaker's beliefs and intentions. So the modular story goes.

On this conception, semantics and pragmatics are insulated from each other. Pragmatics takes as input the output of semantics, but they do not mix, and in particular, pragmatic processes do not interfere with the process of semantic composition which outputs the truth conditions.¹ This makes sense if one construes semantic competence and pragmatic competence as belonging to two distinct ‘modules’ (Borg Reference Borg2004: chapter 2) – hence my name for the view.² Semantic competence belongs to the language faculty, Borg says; it is an aspect of our ‘knowledge of language’. Pragmatic competence has more to do with the so-called ‘theory of mind’, that faculty in virtue of which human subjects are able to explain other people's behaviour by ascribing intentions to them and reading their mind.

The modular picture I have just described has started to lose grip in recent years. Nearly everybody nowadays acknowledges the fact that the reference of indexicals and, more generally, the semantic value of context-sensitive expressions cannot be determined without appealing to full-fledged pragmatic factors (e.g. speaker's intentions). The semantic value of a context-sensitive expression varies from occurrence to occurrence, yet it often varies not as a function of some objective feature of the context but as a function of what the speaker means. Pragmatic competence, therefore, is required not only to determine what the speaker means on the basis of what she says, but also to determine what is said in the first place. That means that we have to give up the modular view, and accept that pragmatics and semantics do mix in fixing truth-conditional content.

Of course, if one wants to maintain a semantics pure of pragmatic intrusion, one can, but then one has to construe the goal of semantics differently than it is on the standard conception. Pure semantics will no longer deliver truth conditions, but it will deliver, say, conditional truth conditions, or schemata, or characters, or propositional radicals, or whatever. To get full-blown truth-conditional content, pragmatics will be needed. This non-modular approach to truth-conditional content is one of the key ingredients of contemporary Contextualism.

7.2 The ‘extent of context-sensitivity’ issue

According to most contemporary theorists, context-sensitivity is pervasive in natural language. In addition to the obvious indexicals, many expressions turn out to be context-sensitive in one way or another. This covers two types of case. There are expressions which display hidden indexicality in that, when properly analysed, they turn out to behave very much like standard indexicals, or to contain hidden constituents which do; and there are expressions which display other forms of context-sensitivity. In any case, we don't know in advance whether a given expression is, or is not, context-sensitive: it is an empirical question, to be resolved through linguistic analysis. So we must reject a certain presumption which was still prevalent twenty years ago and which we may call the ‘literalist presumption’.

To introduce the literalist presumption, let us start from the following (uncontroversial) premiss:

1. There is a ‘basic set’ of expressions whose content is known to depend upon the context in a systematic manner: the indexicals.

The presumption which the pervasiveness of context-sensitivity leads us to reject can now be stated as follows:

1. Literalist presumption: expressions not in the basic set are (by default) assumed to be context-insensitive.

The literalist presumption is implicitly at work in a number of fallacious arguments using Grice's ‘Modified Occam's Razor’, or an equivalent principle of parsimony, to demonstrate that a semantic analysis in terms of conversational implicature is preferable to an account in terms of truth-conditional content proper (Recanati Reference Recanati and Tsohatzidis1994, Reference Recanati2004a: 155–8). Classic examples involve the use of Modified Occam's Razor against Strawson's (1952) view of the contextually varying truth-conditional contributions of ‘and’, or against Donnellan's view of the contextually varying truth-conditional contributions of definite descriptions. In each case, the possibility that the relevant expression (which seems to carry different contents in different contexts) might be context-sensitive even though it does not belong to the basic set is ignored, in virtue of the literalist presumption, and the argument proceeds as if the only options available to account for the data were lexical ambiguity on the one hand and conversational implicature on the other hand (with Modified Occam's Razor being used to rule out the former option).

Rejection of the literalist presumption is another key ingredient in contemporary Contextualism. It corresponds to a stance I dubbed ‘Methodological Contextualism’ (Recanati Reference Recanati and Tsohatzidis1994). According to Methodological Contextualism, we don't know in advance which expressions are context-sensitive and which aren't. For all we know, every expression might be context-sensitive. Here the universal quantifier takes scope over the epistemic modal, so what generalises is the possibility of context-sensitivity. For every expression e – including ‘and’ or definite descriptions – it may be that e is context-sensitive and contributes different contents in different contexts (even though e is not ambiguous). As a result, we need to draw a general distinction between the linguistic meaning of an expression and its contribution to propositional content, while allowing for special cases in which they will be identical, instead of doing the opposite (i.e. equating conventional meaning and propositional contribution, while allowing for exceptions – the expressions in the ‘basic set’).

Methodological Contextualism goes together with the view that context-sensitivity is a pervasive phenomenon in natural language. Not everybody accepts this view, however. The so-called ‘semantic minimalists’ (e.g. Cappelen and Lepore Reference Cappelen and Lepore2005a) believe that context-sensitivity is a very limited phenomenon, corresponding roughly to expressions in the basic set, so they don't feel compelled to reject the literalist presumption.

7.3 Does context-sensitivity generalise?

The two contextualist ingredients I have described so far correspond to views that are widely shared among contemporary theorists. Those who still believe in modularity and are faithful to the literalist presumption are a rather small minority (see Chapter 25 for more on their view). But there is another minority, at the other end of the spectrum: the radical contextualists. What characterises their view is the generalisation of context-sensitivity.³

There are several possible arguments for the generalisation of context-sensitivity. First, one should distinguish the claim that context-sensitivity generalises at the sentential level from the much stronger claim that it generalises at the constituent or the lexical level. One generalises context-sensitivity at the sentential level if one holds that the truth conditions of a sentence always depend upon the context (so that there are no ‘eternal sentences’). Here is an example of an argument for that conclusion:

1. (a) A successful sentence (i.e. a sentence that succeeds in expressing a proposition) expresses either a singular proposition or a general proposition.

2. (b) If a sentence expresses a singular proposition, it does so in a context-sensitive manner because (i) a sentence expresses a singular proposition only if it contains a successful referring expression (i.e. an expression which succeeds in referring), and (ii) reference is inherently context-sensitive. (That is so because an expression-token only refers to some object in virtue of contextual relations between the token and the object.⁴)

3. (c) If a sentence expresses a general proposition, it does so in a context-sensitive manner because (i) a sentence expresses a general proposition only if it contains a (successful) quantificational expression, and (ii) quantification is inherently context-sensitive. (That is so because the domain of quantification depends upon the context. An expression succeeds in quantifying only if the context supplies an appropriate domain of quantification.)

4. (d) Conclusion: Whenever a sentence expresses a proposition, it does so in a context-sensitive manner.

The argument I have just presented relies on many controversial assumptions regarding, inter alia, the nature of reference, the semantics/meta-semantics distinction, the semantic analysis of plurals, proper names and definite descriptions, quantifier domain restriction, and so on and so forth. It is not my intention to go into these thorny issues here, in order to evaluate the argument. Rather, I will focus on arguments for the generalisation of context-sensitivity at the level of sentential constituents, or at the lexical level, rather than at the sentential level. As I already suggested, the claim that context-sensitivity generalises at the constituent or lexical level is much stronger than the claim that it generalises at the sentential level. What we are now considering is the possibility that every expression (not just every sentence) might be context-sensitive. This is a very radical form of Contextualism indeed.

We have already encountered the claim that ‘every expression might be context-sensitive’ in the context of Methodological Contextualism, but then the universal quantifier ‘every expression’ took scope over the epistemic modal. What Methodological Contextualism meant to generalise to every expression was only the possibility of its being context-sensitive. In the context of Radical Contextualism, the claim that ‘every expression might be context-sensitive’ is understood differently: the modal now scopes over the universal quantifier. The possibility that is being considered is the possibility that, for every expression e, e is context-sensitive. Here what tentatively generalises is (actual) context-sensitivity, not the possibility of context-sensitivity.

In what follows I will present two types of argument for the generalisation of context-sensitivity at the lexical or constituent level. One such argument involves the phenomenon of pragmatic modulation. It is of special importance and will be discussed in sections 7.5 and 7.6. Other arguments, to which I now turn, are based on considerations from lexical semantics.

7.4 Arguments from lexical semantics

What is the meaning of a word? Let us assume that utterances express ‘propositions’ or ‘thoughts’, and that these propositions/thoughts are made out of, or can be analysed into, certain building blocks or constituents, to be called senses. The standard assumption regarding word meaning is that the conventions of the language associate expressions with senses. What an expression contributes, when it is used (together with other expressions) in making a complete utterance, is supposed to be the sense which it independently possesses in virtue of the conventions of the language. Indexicals (and context-sensitive expressions more generally) are considered an exception: their sense is not to be equated with their linguistic meaning, but depends upon the context. Now Radical Contextualism rejects the very idea that the conventions of the language associate expressions with full-fledged senses. It generalises the distinction between the lexical meaning of an expression and its sense (or ‘content’ or ‘propositional contribution’), which is said to depend upon the context.

Putnam's ideas about the lexical semantics of nouns can be seen as a forerunner of Radical Contextualism. Putnam criticises the Fregean idea that the meaning of a noun is a (context-independent) sense which determines the noun's reference (i.e. its extension). In many cases, Putnam points out, we start with the reference: the noun is associated with contextually given exemplars known to fall into the extension of the noun, or, on a more refined picture, with a stereotype used to identify the exemplars to be found in one's local environment. What determines the extension of the noun is not the lexically given stereotype, however, but a certain relation R of similarity to the local exemplars. So the extension-determining sense of an expression is context-dependent on two counts: it depends upon the contextually given exemplars (e.g. whether the transparent, thirst-quenching liquid around us is H₂O or XYZ), and it depends upon the relation R, which itself may vary according to the interests of the conversational participants (Putnam Reference Putnam1975: 238–9). The sense thus determined in context is utterly different from the lexical meaning of the noun, which Putnam describes as a ‘vector’ consisting of, inter alia, a ‘semantic marker’ (e.g. liquid) and the exemplar-fixing stereotype.

Putnam's story is meant to apply to the nouns that are used to talk about things in the environment, and which we learn by getting acquainted with the things they are used to talk about. Not merely natural-kind terms like ‘water’, but also, he says, nouns for artefacts like ‘pencil’. It could presumably be extended to other categories of lexical items, such as verbs and adjectives. On a Putnamian semantics, context-sensitivity generalises to all the words which have a ‘referential’ dimension and directly connect up with aspects of the world around us. That is arguably the core of the language.

Another class of expressions which might claim to be ‘the core of the language’, though for totally different reasons, are the most frequently used expressions – light verbs (e.g. get, have, take), prepositions and the like. Such expressions exhibit a high degree of polysemy, and this raises a problem for lexical semantics. What do such words mean? A number of scholars believe their meaning is schematic and has to be fleshed out on any particular use. This suggests that, perhaps, their conventional meaning is not a full-fledged sense. Can one argue that they are ambiguous between a number of distinct senses? That is not obvious because it does not seem that there is a discrete list of such senses available but, rather, a continuum of possible senses to which one can creatively add in an open-ended manner. That is not to say that the meaning of such an expression reduces to an abstract schema: the expression is undoubtedly also associated in memory with conventional ways of using it in collocations with (more or less) determinate senses. All this – the abstract schema or schemata, the collocations, the senses – arguably goes into the linguistic meaning of the expression, which starts looking rather messy. On such a view, the meaning of an expression does not have the right ‘format’ to be what the expression contributes to propositional content.⁵ In other words, linguistic meanings are not senses (though they may involve senses, inter alia).

The two views of lexical semantics I have just mentioned may well be wrong, of course. When it comes to lexical semantics, nearly everything is up for grabs. That, however, is precisely the point. As theorists, we have an idea what senses are, i.e. what words contribute when we speak. We know, more or less, how to model that. But we know very little about what words themselves mean and what relation there is between word meaning and contributed sense. In view of the limits of our knowledge, it is reasonable to give up the simplifying assumption that linguistic meanings are senses, in order at least to start making serious enquiries in that area.

If we give up that simplifying assumption, we are left with the idea that lexical meaning plays some role in determining the sense which is an expression's contribution to the thought expressed. This idea can be expressed by saying that the sense of an expression is a function of the lexical meaning of that expression and some factor x, where ‘x’ is whatever, in addition to lexical meaning, is needed to determine sense.⁶ If, as seems very likely, ‘x’ includes the context in which the expression is tokened (and in particular the most important among contextual factors: what one is talking about), then we get a radical form of Contextualism that ‘generalises indexicality’.

7.5 Pragmatic modulation

The arguments from lexical semantics I have just mentioned support Radical Contextualism because they cast doubt on the idea that an expression's contribution to truth-conditional content (its sense) is fixed by the rules of the language independent of context. Most expressions are treated as indexical-like in that their semantic content depends upon the context somehow. But this does not (yet) mean that the semantic contribution of every expression is context-dependent. The next argument for Radical Contextualism leads to that stronger conclusion, however. I call it the argument from pragmatic modulation.

The argument from pragmatic modulation is independent of the arguments from lexical semantics, and it can be put forward even if we grant the assumption which the arguments from lexical semantics lead us to doubt: that the conventions of the language directly associate expressions with full-fledged senses.⁷ Let us, indeed, grant that assumption. One can still deny that the senses which – on this view – are the meanings of expressions are also what these expressions contribute when they are used (together with other expressions) in making a complete utterance. Because of pragmatic modulation, an expression may, but need not, contribute its sense – i.e. the sense it independently possesses in virtue of the conventions of the language (assuming for the sake of argument that it possesses such a sense); it may also contribute an indefinite number of other senses resulting from modulation operations (e.g. free enrichment, predicate transfer, sense-extension etc.) applied to the proprietary sense.

For example, a sentence like (1) has several readings.

1. (1) There is a lion in the middle of the piazza.

On one reading ‘lion’ is given a non-literal interpretation and means something like ‘statue of a lion’. On that reading (1) may be true even if, literally, there is no lion in the middle of the piazza (but only a statue of a lion). Or consider (2).

1. (2) The ATM swallowed my credit card.

This can be given a literal reading, if we imagine a context à la Putnam in which ATMs turn out to be living organisms. But the sentence can also and typically will be interpreted non-literally. In an ordinary context, ‘swallow’ will be given an extended sense, corresponding to what ATMs sometimes do with credit cards (something which, indeed, resembles swallowing). The sentence may be true, on such a reading, even though no real swallowing takes place. In a less ordinary context in which there is a person disguised as an ATM, the predicate ‘ATM’ in the description will be given a non-literal reading: through ‘predicate transfer’ (Nunberg Reference Nunberg1995), it will acquire the meaning ‘person disguised as an ATM’ (just as ‘lion’ acquires the meaning ‘statue of a lion’ in the previous example). The sentence may well be true, on that interpretation, even though no real ATM swallows anything, provided the person disguised as an ATM does swallow the credit card.⁸

Accepting pragmatic modulation (here, the process mapping the literal meaning of ‘lion’ or ‘ATM’ to the relevant representational reading, or the meaning of ‘swallow’ to its extended reading) as a possible determinant of truth-conditional content leads to a radical form of Contextualism, because modulation itself is context-sensitive: whether or not modulation comes into play, and if it does, which modulation operation takes place, is a matter of context. It follows that what an expression actually contributes to the thought expressed by the utterance in which it occurs is always a matter of context.

Of course, not everybody accepts that pragmatic modulation affects the truth-conditional content of an utterance. On the currently dominant picture, pragmatics comes into play in the determination of truth-conditional content but does so only when the semantic rules of the language prescribe it (as when an indexical demands a contextual value). On this view the only truth-conditional role of pragmatics corresponds to what I have called ‘saturation’ (in contrast to ‘modulation’). Saturation is a pragmatic process of contextual value-assignment that is triggered (and made obligatory) by something in the sentence itself, namely the linguistic expression to which a value is contextually assigned. 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. The expression itself acts as a variable in need of contextual instantiation. So pragmatic competence comes into play, but it does so under the guidance of the linguistic material: the pragmatic process of saturation is a ‘bottom-up’ process in the sense that it is signal-driven, not context-driven. In contrast, pragmatic modulation is a ‘top-down’ or context-driven process, i.e. a pragmatic process which is not triggered by an expression in the sentence but takes place for purely pragmatic reasons – in order to make sense of what the speaker is saying. In other words, it is a ‘free’ pragmatic process – free because it is not mandated by the linguistic material but responds to wholly pragmatic considerations. That is clearly the case for the pragmatic processes through which an expression is given a non-literal interpretation: we interpret an expression non-literally in order to make sense of the speech act, not because this is dictated by the linguistic materials in virtue of the rules of the language.

The dominant view is that the only pragmatic process that can affect truth-conditional content is saturation. No ‘top-down’ or free pragmatic process can affect truth-conditions – such a process can only affect what the speaker means (but not what she says). As Stanley puts it, ‘all truth-conditional context-dependence results from fixing the values of contextually sensitive elements in the real structure of natural language sentences’ (Stanley Reference Stanley2000: 392). Or, as King and Stanley put it, there can only be ‘weak’ pragmatic effects on truth-conditional content. They define a weak pragmatic effect as follows:

Radical Contextualism rejects the view that only weak pragmatic effects can affect what is said. It holds that truth-conditional content may be affected not only by saturation (as when an indexical is assigned a contextual value) but also by free pragmatic processes of modulation. In (1) the non-literal reading of ‘lion’ arguably results from a pragmatic operation that is not dictated by the lexical item lion in virtue of its standing meaning. There is no slot to be filled, no free variable or context-sensitive element whose value is to be fixed, or anything of the sort. Moreover, nothing (except the desire to make sense of the speaker, in a context in which the literal reading is unlikely) prevents the sentence from being interpreted literally. The pragmatic effect here looks like a strong pragmatic effect, yet it affects truth-conditional content. Since the semantic content of any expression can be pragmatically modulated in this way, what an expression actually contributes depends upon the context.

7.6 Why resist Radical Contextualism?

I have presented five distinct issues relevant to the overall debate between Contextualism and Literalism. On each issue there are two sides: the contextualist side and the literalist side.

1. Modularity. Is pragmatic competence involved in the determination of truth-conditional content? Contextualism: Yes. Literalism: No.

2. Extent of context-sensitivity. Is context-sensitivity pervasive in natural language? Contextualism: Yes. Literalism: No.

3. Generalisation of context-sensitivity (1). Is it true that there are no eternal sentences, i.e. no sentence which expresses a proposition independent of context? Contextualism: Yes. Literalism: No.

4. Generalisation of context-sensitivity (2). Are all/most expressions like indexicals in that their lexical meaning does not add up to a full-fledged sense? Contextualism: Yes. Literalism: No.

5. Generalisation of context-sensitivity (3). Does pragmatic modulation affect truth-conditional content? Contextualism: Yes. Literalism: No.

On the first two issues (the modularity issue, and the extent of context-sensitivity issue) Contextualism wins in the sense that it is the dominant position. The role of a speaker's intentions and pragmatic competence in fixing contextual values for indexicals etc. is widely acknowledged, as is the pervasiveness of context-sensitivity in natural language. With respect to the ‘generalisation of context-sensivity’ issues, however, it is the other way round: Literalism wins, sociologically speaking. There is widespread resistance to the most radical forms of Contextualism – those which generalise context-sensitivity. Why?

In this section, I will discuss some of the reasons one might have for resisting the generalisation of context-sensitivity.⁹ I will argue that they do not carry much weight. This will bring a sixth issue to the fore – the systematicity issue, to which the final section of this chapter will be devoted.

Let us start with the arguments from lexical semantics. Why not simply accept them? I submit that the main source of resistance to the idea that a word's lexical meaning does not add up to a full-fledged sense is the following. Most semanticists worry more about compositional semantics than about lexical semantics, so they make their lives simpler by uncritically accepting the simplifying assumption (inherited from our elders) that lexical meanings are senses. In this way we don't have to care about how senses are generated – we take them simply as given. Admittedly, this is an acceptable idealisation at a certain stage in the development of semantics; but as soon as one gets interested in the foundations of lexical semantics, one should start by lifting the simplifying assumption in order at least to consider the issue with an open mind.

The resistance to pragmatic modulation is a more serious matter. Here it seems that there are substantive reasons to be suspicious of Radical Contextualism. The first reason is this. If free pragmatic processes are allowed to affect semantic content, semantic content leaps out of control – it is no longer determined by the rules of the language but varies freely, à la Humpty Dumpty. But then, how can we account for the success of communication? Communication (content sharing) becomes a miracle since there is nothing to ensure that communicators and their addressees will converge on the same content. Now communication is possible (it takes place all the time), and there is no miracle. It follows that we should give up the view that free pragmatic processes play a role in the determination of semantic content (Cappelen and Lepore Reference Cappelen and Lepore2005a: chapter 8).

This argument fails, I believe, because the problem it raises is a problem for everybody, as soon as one gives up the modular view. Whenever the semantic value of a linguistic expression must be pragmatically inferred, the question arises, what guarantees that the hearer will be able to latch on to the exact same semantic value as the speaker? Whether the pragmatic process at stake is saturation or modulation is irrelevant as far as this issue is concerned, so the argument fails as an argument specifically intended to cast doubt on pragmatic modulation.¹⁰

Another argument against pragmatic modulation as a possible determinant of semantic content can be put as follows. ‘What is said’, the truth-conditional content of an utterance, is what is literally said, and that – by definition – has to be determined by the conventions of the language. Pragmatics can enter the picture, provided its role is to assign a contextual value to a lexical item in a bottom-up manner, i.e. in accord with (and under the guidance of) the conventional meaning of that context-sensitive item. In contrast, strong pragmatic effects achieved in order to make sense of the speech act without being linguistically mandated take us into the realm of speaker's meaning, away from literal meaning.

Insofar as this argument is based upon a certain understanding of the phrase ‘what is said’ (or ‘what is literally said’), it is not substantive, but verbal. There is no doubt that one can define ‘what is said’ in such a way that only weak pragmatic effects can affect what is said. But what the advocate of pragmatic modulation means by ‘what is said’ corresponds to the intuitive truth-conditional content of the utterance.¹¹ According to the contextualist side in the debate, the intuitive truth conditions of an utterance of (1) or (2) are affected by free pragmatic processes. Assuming this is true, this does not prevent us from defining another notion of what is said, conforming to literalist standards. Let ‘what is said_min’ be the proposition expressed by an utterance when strong pragmatic effects have been discounted (the so-called ‘minimal proposition’), and let ‘what is said_int’ correspond to the intuitive truth conditions of the utterance. According to the contextualist view under discussion, what is said_int may be affected by top-down pragmatic processes. This is compatible with the claim that only weak pragmatic effects can affect what is said_min. So that claim in no way counters the idea that pragmatic modulation affects the (intuitive) truth-conditional content of utterances.

According to yet another argument, if we accept the view that pragmatic modulation affects semantic content, we blur the semantics/pragmatics distinction to the point where there no longer is any difference between what is said and what is meant; so – assuming this distinction is essential – we should reject the view that pragmatic modulation affects truth-conditional content. I find this argument unconvincing, because acknowledging the effects of pragmatic modulation on truth-conditional content (what is said_int) in no way prevents one from distinguishing what is said_int from other things that are conveyed by an utterance without belonging to its intuitive truth-conditional content, e.g. the particularised conversational implicatures (Grice Reference Grice1989) or the effects achieved through ‘staging’ (Clark H. Reference Clark1996). In other words, we can distinguish between ‘primary’ pragmatic processes, such as modulation, and ‘secondary’ pragmatic processes that do not contribute to what is said_int (Recanati Reference Recanati and Davis1989, Reference Recanati2004a). So we do not lose the distinction between what is said and what is meant.

Conclusion: the three arguments against pragmatic modulation I have extracted from the literature and presented in this section are no good.¹² However, the most important argument against pragmatic modulation is, by far, the systematicity argument, which I have not yet introduced. That argument, which I take to be the main source of resistance to Radical Contextualism, deserves separate discussion since it raises a new issue relevant to the overall debate.

7.7 The systematicity issue

Many theorists argue as follows. If Radical Contextualism is true, the project of constructing a systematic truth-conditional semantics for natural language is doomed to failure. We should therefore reject Radical Contextualism, since it leads to scepticism. Or rather, we should reject that ingredient which is incompatible with the project of constructing a systematic truth-conditional semantics. That is not the generalisation of indexicality prompted by the lexical semantics considerations; for, if indexicality is compatible with formal semantics, generalised indexicality should be compatible with it, too. The problem, rather, comes from the acceptance of pragmatic modulation as a determinant of semantic content. That is what we should reject.

Of course, one has to say why the acceptance of pragmatic modulation as a determinant of semantic content is incompatible with the project of building a systematic semantics. Here is a first sketch of an argument for that conclusion.

I grant the objector that modulation is unsystematic. Still, I think it is easy to make room for it within a systematic semantics.¹³ In general, nothing prevents unsystematic factors from being handled systematically, by being assigned their proper place in the theory. In the case at hand, we can define a function mod taking as argument an expression e and the context c in which it occurs: the value of mod is the particular modulation function that is contextually salient/relevant/appropriate for the interpretation of that expression in that context. If no modulation is contextually appropriate and the expression receives its literal interpretation, the value of mod will be the identity function. In this framework, we can distinguish between the literal sense of a simple expression e, namely its semantic interpretation I(e), and the modulated sense M(e)_c carried by an occurrence of e in context c. The modulated sense of an expression e (in context c) results from applying the contextually appropriate modulation function mod (e, c) to its semantic interpretation I(e):

So far, this is very standard: in distinguishing I(e) from M(e)_c we are just appealing to the traditional semantics/pragmatics distinction. What is not standard is the claim that the semantic interpretation of a complex expression (e.g. a sentence) is a function of the modulated senses of its parts and the way they are put together (Recanati Reference Recanati2010: chapter 1). This is what examples like (1) and (2) suggest if we take at face value the effects of modulation on truth-conditional content which they seem to display. On the resulting view the semantic process of composition and the pragmatic process of sense modulation are intertwined. For simple expressions, their semantic interpretation is their literal sense, but for complex expressions pragmatic modulation is allowed to enter into the determination of semantic content. This is non-standard, for sure, but there is nothing unsystematic about this view.

The systematicity objection can be understood differently, however. What is not systematic enough, according to the objection, is not so much the radical contextualist's theory of utterance interpretation, but utterance interpretation itself (what the theory is about) as construed by the radical contextualist. We have seen that, in the contextualist framework with its ‘free’ pragmatic processes, interpretation (content recovery) is no longer driven by the linguistic material. In introducing modulation (in contrast to saturation), I said that in saturation ‘pragmatic competence comes into play, but does so under the guidance of the linguistic material’, whereas modulation ‘is not triggered by an expression in the sentence but takes place for purely pragmatic reasons – in order to make sense of what the speaker is saying’. This suggests that, for a radical contextualist, utterance interpretation is pragmatic through and through and does not significantly differ from ‘the kind [of interpretation] involved in interpreting kicks under the table and taps on the shoulder’ (Stanley Reference Stanley2000: 396): it is not the systematic affair which formal semanticists have claimed it to be. In this way, we reach the conclusion that Radical Contextualism is incompatible with the programme of formal semantics.

Thus understood, I think the objection is confused. Even though free pragmatic processes, i.e. pragmatic processes that are not mandated by the standing meaning of any expression in the sentence, are allowed to enter into the determination of truth-conditional content, still, in the framework I have sketched, they come into the picture as part of the compositional machinery. Semantic interpretation remains grammar-driven even if, in the course of semantic interpretation, pragmatics is appealed to not only to assign contextual values to indexicals and free variables but also to freely modulate the senses of the constituents in a top-down manner. Semantic interpretation is still a matter of determining the sense of the whole as a function of the (possibly modulated) senses of the parts and the way they are put together.

If what I have just said is right, the systematicity issue is orthogonal to the other issues I discussed as relevant to the Contextualism/Literalism debate. The systematicity issue can be formulated thus:

1. Systematicity. Is semantic interpretation a matter of holistic guesswork (like the interpretation of kicks under the table), rather than an algorithmic, grammar-driven process as formal semanticists have claimed? Contextualism: Yes. Literalism: No.

On that issue I am happy to part company with the most radical contextualists – the ‘sceptics’ who would go for the holistic guesswork answer (assuming they exist, which I doubt). Like Stanley and the formal semanticists, I maintain that semantic interpretation is grammar-driven. But this issue is orthogonal to the others! So, without contradicting what I have just said, I can still hold that a good deal of holistic guesswork comes into play in semantic interpretation, e.g. in order to fix the values of context-sensitive elements or to pick the right modulation functions. Both in the case of saturation and in the case of modulation – the two types of contextual process that play a part in the determination of truth-conditional content – one has to rely heavily on pragmatic competence. Accepting this point means that one endorses a quite radical form of Contextualism; but this is compatible with maintaining that semantic interpretation is grammar-driven and proceeds recursively. I conclude that there is no reason to take one's adherence to the project of building a systematic semantics to prevent one from being a contextualist with respect to any of the issues talked about earlier.

Of course, a radical contextualist will be prone to set limits to systemat-icity in semantics; but that is not the same thing as getting rid of systematicity altogether. As I said, when it comes to fixing the values of context-sensitive elements or to picking the right modulation functions, pragmatic competence takes over: formal semantics has nothing to say regarding the pragmatic mechanisms in play, and the hand-waving word ‘salience’ which semanticists like to use is only a placeholder. But these limitations put on semantics are nothing but the price to pay for giving up the modular view. We have to admit, once and for all, that (intuitive) truth-conditional content or ‘what is said’ is not something purely semantic – something that can be retrieved simply by exercising one's semantic competence. Pragmatic competence massively comes into the picture. To reach that conclusion, however, there is no need to consider anything fancier than deictic pronouns. In particular, there is no need to go into the ‘generalisation of context-sensitivity’ issues which have loomed large in our discussion of Radical Contextualism.

What I have just said shows that there is a continuum of positions with respect to systematicity. The more literalist one is, the stronger the form of systematicity one will be in a position to claim for semantics. This means that even a moderate contextualist – someone who (merely) gives up the modular view and accepts the pervasiveness of context-sensitivity – will have to set limits to systematic semantics. These limitations I think most language theorists currently accept. Radical contextualists who generalise context-sensitivity simply go a bit farther in the same direction, but, I insist, there is only a difference of degree between them and the moderate contextualists. (In this respect I agree with Cappelen and Lepore Reference Cappelen and Lepore2005a.) Radical Contextualism, therefore, is not a revolutionary position which threatens the programme of formal semantics, as many theorists have claimed. If revolution there is, it antedates Radical Contextualism and coincides with the advent of the moderate form of Contextualism which almost everybody embraces.

8.1 Introduction

Consider the cartoon in Figure 8.1,¹ which is an optimal innovation: it includes a novel stimulus intended to further activate coded, salient meanings, so that both the novel and the salient, though different, may interact and affect pleasurability (Giora, Reference Giora2003; Giora et al., Reference Giora, Fein, Kronrod, Elnatan, Shuval and Zur2004).

Put differently, this cartoon will be optimally innovative only to those familiar with the coded meanings it echoes (see e.g., the poster in Figure 8.2). The initiated audiences are thus invited to invoke a wide range of coded, salient meanings which might allow an insight into the ironic message of the cartoon. However, those not in the know will only have access to what is literally spelled out, superimposed on the pictorial background (in addition, of course, to the figurative title and the symbol of the crown). Indeed, if taken at face value, the cartoon will encourage people to get terrified and stop doing what they are doing on account of the nasty weather (referred to in the background). But the poster in Figure 8.2, if retrievable from memory, will bring to mind an altogether different meaning which this cartoon must be weighed against: the British mindset on the eve of World War II – the “British restraint and stiff upper lip.”²

Figure 8.1. Spirit of the Blitz 2009

Figure 8.2. Keep Calm and Carry On

Additionally, one of Lance Corporal Jones's catchphrases in Dad's Army – the British sitcom about the Home Guard in the Second World War (Perry and Croft, Reference Perry and Croft1968–Reference Perry and Ludlow1977) – might also spring to mind, echoing, via another irony, the mindset derided here:

1. (1) Don't panic!³

The title of the cartoon – Spirit of the Blitz 2009 – allows another ironic turn of the screw, reminding us of the spirit of the blitz exercised by the people of Britain, who, during the 1940 bombing of London, exhibited stoical courage and endurance.⁴

All these meanings, if coded and available, will be invoked automatically as a direct response to the stimulus in Figure 8.1, despite their apparent irrelevance to the immediate context at hand – the ferocious weather in Britain during December 2009. When weighed against contextual information – the blitz-like storm which forced many in Britain to assume the spirit of the blitz rather than the spirit of Christmas – these meanings strongly deride the panicky spirit of the Brits on Christmas Eve 2009.

The way this cartoon can be interpreted illustrates the need to take various factors into consideration when examining the end-product of utterance interpretation. Indeed, the psychology of utterance processing takes into account a number of factors that shape utterance interpretation, such as (i) salient/coded meanings, (ii) contextual information, and (iii) their unfolding interaction (or lack of it).

Debates within psycholinguistics can, thus, be viewed as divisible into two main approaches (for a review, see Giora, Reference Giora2003: chapters 1–3). At one end of the spectrum are context-based models which assume that a strong context reigns supreme in that it governs early processes and facilitates contextually compatible meanings only. Consequently, the output of the interpretation processes must be seamless, involving no contextually incompatible meanings and interpretations (the connectionist model, e.g., Bates, Reference Bates, Bizzi, Calissano and Volterra1999; Bates and MacWhinney, Reference Bates, MacWhinney, MacWhinney and Bates1989; MacWhinney, Reference MacWhinney and MacWhinney1987; Small et al., 1988; the constraint-based model, e.g., McRae et al., Reference McRae, Spivey-Knowlton and Tanenhaus1998; Pexman et al., Reference Pexman, Ferretti and Katz2000; the direct access view, e.g., Gibbs, Reference Gibbs1979, Reference Gibbs1994; Keysar, Reference Keysar1994a; Ortony et al., Reference Ortony, Schallert, Reynolds and Antos1978).

At the other end are lexicon-based models which hold that coded meanings of stimuli are speedy responses, activated automatically, regardless of contextual information. As a result, initially accessed meanings may fail to meet context fit. Consequently, they may induce incompatible interpret-ations which will either feature in final outputs alongside the appropriate ones or will be subjected to revisitation or suppression processes (the modular view, e.g., Fodor, Reference Fodor1983; Swinney, Reference Swinney1979; the standard pragmatic model, e.g., Grice, Reference Grice, Cole and Morgan1975; Searle, Reference Searle1979a; the graded salience hypothesis, e.g., Giora, Reference Giora1997, Reference Giora2003; Peleg et al., 2004, Reference Peleg and Eviatar2008). The context-based and the lexicon-based models, then, have different predictions, especially with regard to initial processes, which, in turn, affect later interpretation processes.

Specifically, according to the context-based view, specific and supportive contextual information penetrates lexical access and selects the contextually appropriate meaning only, the consequence of which is contextually compatible interpretations only (section 8.1.1). In contrast, the lexicon-based view predicts that coded meanings cannot be blocked and, therefore, at times, end-product interpretations will also involve contextually incompatible meanings and interpretations (section 8.1.2). To tease apart these two approaches, we need to look at research into utterance interpretation processes.

8.2 Salient meanings and salience-based interpretations are not necessarily literal

According to the graded salience hypothesis (Giora, Reference Giora2003, 2006; Giora and Fein, Reference Giora and Fein1999a), neither salient meanings nor salience-based interpretations need to be literal. Similarly, non-salient interpretation need not be figurative.

8.2.3 Non-salient interpretations are not necessarily non-literal

Consider the example in Figure 8.1, which is literal: its final interpretation includes a number of salient meanings to which it adds a novel twist. Consider further Know Hope (in Figure 8.3)⁵ which is a literal, non-salient use, harping on the salient, literal “no hope.”

Figure 8.3. Know Hope

In Giora et al. (Reference Giora, Fein, Kronrod, Elnatan, Shuval and Zur2004) we studied such (Hebrew) innovations, termed “optimal innovations.” To be optimally innovative, a stimulus should evoke a novel – less or non-salient – response to a familiar stimulus, alongside a salient one from which it differs (both quantitatively and qualitatively), so that both can interact, regardless of non-literality. Admittedly many optimal innovations may be non-literal (novel metaphors or unfamiliar ironies). However, quite a few are literal (as can be deduced from Figures 8.1 and 8.3). Consider the following examples: Body and sole – the name of a shoe shop – which evokes the salient body and soul, both of which are literal but only the first is non-salient; Curl up and dye (the name of a hair salon) which is non-salient and literal and which evokes the salient but metaphorical Curl up and die (see Giora et al., Reference Giora, Fein, Kronrod, Elnatan, Shuval and Zur2004).

Given that optimal innovation involves processing a salient meaning on top of the novel interpretation, it is no wonder that optimal innovations take longer to process than their salient meanings, regardless of metaphoricality (as shown by Giora et al., Reference Giora, Fein, Kronrod, Elnatan, Shuval and Zur2004). Indeed, in Giora et al. (forthcoming: Experiment 2), both healthy individuals and individuals diagnosed with Asperger‘s syndrome took longer to process and more frequently erred on (Hebrew) non-salient interpretations compared to salient ones, regardless of metaphoricality. Thus, literal optimal innovations such as a Tverian horse (meaning ‘a horse from Tiberias’), reminiscent of a salient, metaphorical collocation – a Trojan horse, were slower to induce correct meaningfulness judgments compared to familiar literal collocations, whose meanings are salient. Non-salient interpretations then are not necessarily non-literal.

8.3 Opting for the literal interpretation is not necessarily a default strategy

To further test the claim that opting for a non-literal rather than a literal interpretation may be a default strategy, independent of explicit contextual information (including information about the speaker and the addressee), one needs to neutralize factors affecting non-literality such as degree of salience (recall that familiar items may have a lexicalized non-literal meaning; see section 8.2.1), semantic anomaly (known to trigger metaphoricality; see e.g., Beardsley, Reference Beardsley1958), and contextual information (since breach of pragmatic maxims or contextual misfit may invite a non-literal interpretation; see e.g., Grice, Reference Grice, Cole and Morgan1975). It thus follows that for this claim to be experimentally substantiated, testing it should involve novel items susceptible to a literal interpretation and presented outside a specific context.

Indeed, in Giora et al. (2010), materials were affirmative statements of the form X is Y (such as This is Memorial Day; I am your doctor) and their negative versions (This is not Memorial Day; I am not your doctor), all of which could, potentially, be assigned a literal interpretation in that they were all equally novel, semantically intact, and presented in isolation (see Figures 8.4 and 8.5).

Figure 8.4. This is Memorial Day

Figure 8.5. This is not Memorial Day

Participants were asked to rate the interpretation of these targets on a seven-point scale ranging between two specific (either literal or non-literal) interpretations presented randomly at the scale's ends.

Results show that the negative statements were rated as significantly more metaphorical than their affirmative counterparts, supporting the claim that opting for the literal interpretation need not be a default strategy.

Indeed, when the negative statements were embedded in contexts equally supportive of either their metaphorical or their literal interpretation, reading times were faster for items embedded in metaphorically than in literally biasing contexts (Giora et al., 2011a), further supporting the claim that a non-literal interpretation may be a default interpretation.

Additional evidence supportive of this claim comes from research into irony interpretation. Giora et al. (2005b) show that negating affirmative overstatements results in assigning these statements an ironic interpretation, even though these statements are amenable to literal interpretation as well. Specifically, findings demonstrate that negative overstatements (He is not exceptionally bright) come across as ironic even outside a specific context. When presented in isolation, they are rated as more ironic than other alternatives such as affirmative overstatements (He is exceptionally bright) or negated non-overstatements (He is not bright).

Taken together, findings from various negative statements indicate that it is not the case that, when available, literal interpretation is a default. Instead, our findings demonstrate that even non-conventional utterances, susceptible to literal interpretation, are often perceived as non-literal, even when no supportive information of that interpretation is made manifest.

8.4 Context effects – later interpretation processes

Although the various approaches outlined above differ in their predictions with regard to the effects of a “strong context” on early lexical processes (section 8.1), there seems to be an agreement that contextual information should affect later interpretation processes. But what these effects should look like is still a matter of debate. To test the various predictions of the approaches with regard to later processes, I will focus here on late effects of a “strong context” (such that anticipates an ironic utterance) on irony interpretation and on the effects of coherence on later processes of negated information.

8.5 Coda

Is our mind “efficient” enough to engage in processing only contextually appropriate interpretations given that contextual information is strongly supportive of that interpretation, as argued by the contextualist school (e.g., Gibbs, Reference Gibbs1994; Vu et al., Reference Vu, Kellas, Metcalf and Herman2000; see section 8.1.1)? Luckily, it is not. In fact, there is enough evidence now to allow the conclusion that processing utterances, even inside highly biasing contexts, may involve entertaining meanings and interpretations solely on account of their meaning salience and consequently their salience-based interpretation, regardless of contextual fit, as argued by the graded salience hypothesis (see section 8.1.2). Additionally, the activation of such incompatible meanings and interpretations in utterance processing is not unconditionally aborted by suppression processes (as assumed by Fodor, Reference Fodor1983 or Grice, Reference Grice, Cole and Morgan1975). Rather, such meanings and interpretations are retained because they are deemed functional in shaping up final interpretations (the suppression/retention hypothesis, Giora, Reference Giora2003). Such a functional view of suppression and retention allows for the poetics of linguistic and non-linguistic stimuli such as optimal innovations (whether literal or non-literal; section 8.1), for discourse resonance, and humor (section 8.4).

It has also become clear that, contra Grice (Reference Grice, Cole and Morgan1975), non-literal interpretations may be a default interpretation even when innovative, free of semantic anomaly, and context-less (section 8.2.3). Indeed, when affirmative statements of the form X is Y are negated (This is not Memorial Day), they are assigned a non-literal interpretation even when presented in isolation. Similarly, negative overstatements such as He is not exceptionally bright are assigned an ironic interpretation even outside a specific context.

The review of the literature introduced in this chapter reveals that the psychology of utterance processing is a multi-faceted phenomenon; its products may, at times, be surprisingly creative and even amusing.

9.1 Introduction

Most of the time, when we speak, we do more than express propositions; we suggest, promise, offer, accept, order, threaten, assert – we perform speech (or illocutionary) acts. The history of the research on this topic – initiated by Austin (Reference Austin1975) – is well-documented, and many textbooks, handbooks and encyclopaedias contain excellent surveys, thus treating speech acts as a major topic (e.g. Levinson Reference Levinson1983: chapter 5; Jaszczolt Reference Jaszczolt2002: chapter 14; Sadock Reference Sadock, Horn and Ward2004). However, the main contemporary pragmatic theories of utterance interpret-ation devote little space, if any at all, to the way utterances are interpreted as speech acts, that is to the way they are assigned an illocutionary force (see, for instance, Sperber and Wilson Reference Sperber and Wilson1986; Levinson Reference Levinson2000; Carston Reference Carston2002; Recanati Reference Recanati2004a; Jaszczolt Reference Jaszczolt2005). One might think that speech acts went out of fashion simply because the topic had been exhausted by the considerable number of publications spanning from Austin's work in the late fifties to the late eighties – when other topics, such as the pragmatic determinants of literal meaning, came to the fore.

Yet, contemporary literature is rife with confusions stemming from the lack of careful consideration of the role of illocutionary force attribution in utterance interpretation. In particular, two crucial mistakes must be avoided. The first consists in conceiving of illocutionary forces as determined by sentence meaning; the second equates utterance content and speech act content. Interestingly enough, both confusions can be traced back to the founding fathers of modern pragmatics: the former to Searle, the latter to Grice. I will start this chapter by considering these two problematic legacies in turn. Next, we will see how avoiding the confusion between sentence meanings, utterance meanings and illocutionary contents helps to better grasp the major issues related to the analysis of illocutionary force attribution.

9.2 Searle: illocutionary forces as intrinsic to sentence meanings

According to Searle (Reference Searle1969, Reference Searle and Gunderson1975a, b; Searle and Vanderveken Reference Searle and Vanderveken1985), the meaning of a sentence corresponds to the speech act any literal utterance of this sentence constitutes. In Searle's conception, the study of linguistic meaning amounts to the study of speech acts. It follows that in order to determine the literal illocutionary force of an utterance, it suffices to know its linguistic meaning. Searle's view rests on the assumption (explicitly stated by his Expressibility Principle; cf. Searle Reference Searle1969: 20–21) that any illocutionary act type IA can be matched with a certain sentence type s in such a way that IA corresponds to the literal meaning of s. Detailed criticisms of such a ‘literalist’ view – which takes literal utterance meaning to be determined by sentence meaning – can be found in Recanati (Reference Recanati1987: 219–24), Carston (Reference Carston2002: 30–42, 64–70) and Kissine (2011). The important point for present purposes is that Searle's conception implies that illocutionary forces are located at the level of conventional sentence meaning. In his view, if a sentence is used literally, its illocutionary force is directly derivable from its linguistic meaning.

Sentences with imperative mood might seem to provide the strongest case for such a direct derivation of illocutionary force from sentence meaning. This is the reason why I will use imperative sentences to build up my case against incorporating illocutionary forces within sentence meaning. In languages that have a morphological imperative mood – and we will see in a moment that this qualification is important – directive speech acts such as requests, orders, commands etc. are prototypically realised by uttering grammatically imperative sentences. Of course, this does not entail that the linguistic, conventional meaning of the imperative mood is to be analysed in terms of a directive illocutionary force. Yet, I am not flogging a dead horse here, for such an idea underlies many recent accounts of imperative mood: see, for instance, Han (Reference Han2000), Barker (Reference Barker2004), Portner (Reference Portner2007) or (Benjamin) Russell (Reference Russell2007). Without getting into the details, all such theories presuppose that the imperative mood, at the level of sentence meaning, encodes the notion that the speaker (S) is prompting the addressee (A) to bring about the truth of the propositional content (for recent and critical surveys, see Schwager Reference Schwager2006; Iatridou Reference Iatridou2009).

As is well known, many imperative sentences may be used to perform non-directive speech acts (e.g. Wilson and Sperber Reference Wilson, Sperber, Dancy, Moravcsik and Taylor1988). The most obvious examples include permissions (1), advice (2), good wishes (3), and threats (4):

1. (1) A: May I have this piece of cake?

   B: Yes, take it.

2. (2) Always cut your fingernails round and your toenails square.

   (from Hamblin Reference Hamblin1987: 11)

3. (3) Have a nice journey.

4. (4) Hit me, and I'll hit you back.

Virtually everyone agrees that in (sincerely) performing a directive speech act F_D(p),where F_D stands for the directive force and p for the propositional content, S expresses her intention/desire that A bring about the truth of p with F_D(p) as a reason (e.g. Searle Reference Searle and Gunderson1975a; Bach and Harnish Reference Bach and Harnish1979; Searle and Vanderveken Reference Searle and Vanderveken1985; Alston Reference Alston2000). Clearly, the examples in (1)–(4) can be felicitously uttered even though it is mutually manifest that S does not intend A to bring about the truth of the propositional content; for instance, because S would prefer H not to (1), because S does not care whether H does so or not (2), because H has no active control on the truth of the propositional content (3) or because H's bringing about the truth of the propositional content is undesirable for S (4).¹

Those scholars who analyse the meaning encoded by the imperative mood as including some reference to the directive illocutionary force have two options available in order to explain away examples like (1)–(4).² The first consists in elaborating a semantic account of the imperative mood that is flexible enough to predict non-directive uses of imperative sentences (e.g. Wilson and Sperber Reference Wilson, Sperber, Dancy, Moravcsik and Taylor1988; Clark B. Reference Clark1993; Allan Reference Allan2006a; Schwager Reference Schwager2006). This amounts to rejecting the equivalence between the directive illocutionary force and the imperative mood – which is precisely a theoretical recommendation I wish to make in this chapter. The second option is to maintain that the imperative mood encodes the illocutionary directive force, and to claim that (1)–(4) are either indirect or non-literal speech acts.

Let us start by considering the possibility that (1)–(4) are indirect speech acts. Traditionally, a speech act is said to be indirect whenever its performance by means of the utterance u requires u to constitute another, direct speech act. (We will qualify this definition later on, but it is perfectly suited for the needs of the present discussion.) For instance, while (5) is literally a question, it will also constitute a request in many contexts.

1. (5) Don't you think you've had enough beer already?

Likewise, (6) is, literally, an assertion; but it can be interpreted as a directive speech act.

1. (6) You've had enough beer already.

The important point is that while (5) and (6) often constitute indirect directive speech acts, they still remain a question and an assertion, respectively. In other words, if an illocutionary act IA₂ is performed by way of performing another illocutionary IA₁, such that both IA₁ and IA₂ correspond to the same utterance u, the interpretation of u as IA₁ remains available.

In sum, the claim under examination holds that (1)–(4) are, qua direct speech acts, directives (since the imperative mood encodes the directive illocutionary force), but that, qua indirect speech acts, they are interpreted as permission, advice, a good wish and a threat, respectively. However, it is impossible to interpret (1–4) as directive speech acts, except in very specific contexts. And whenever one sets up such a context, it becomes impossible to interpret the examples at hand as permission, advice, a good wish and a threat, respectively. Imagine, for instance, that S and A are involved in some kind of sadomasochistic game, and that S utters (4). To be sure, the imperative in (4) then receives the directive illocutionary force; but in such a context, the example cannot be read as a threat anymore.

It thus turns out that, if the imperative mood is to be associated with the directive illocutionary force, (1)–(4) cannot mean what would be literally said. So we are led to the idea that, while (1)–(4) are directive speech acts when taken literally, they are not interpreted as such when S is speaking non-literally. In other words, non-directive imperatives should be treated in the same way as sarcastic or ironic declaratives. By uttering (7), S may mean that the party is boring; similarly, by uttering (1)–(4) non-literally, S will give advice, give permission, express a good wish or make a threat.

1. (7) This party is great.

The first problem here is that irony may be, and often is, missed. Hearing (7) you can fail to discern the sarcasm, and come to believe that I really love the party. Yet, it seems totally implausible to suppose that, when hearing (1)–(4), we may miss the alleged non-literal meaning and interpret these utterances as plain directive speech acts.

Perhaps such an argument does not settle the issue, for it may be argued that the difficulty of accessing the (alleged) literal meaning of (1)–(4) – that is, to interpret them as directive speech acts – stems from the fact that the (alleged) non‑literal readings – viz. permission, advice, good wish and threat – are so conspicuous that one can hardly miss them. In order to bring home the point that in (1)–(4) the imperative mood is used literally, albeit non-directively, let us consider in more detail the idea that the threat in (4) means the opposite of what is literally said (cf. Dominicy and Franken Reference Dominicy, Franken, Vanderveken and Kubo2002). A first possibility would consist in assuming that only the imperative conjunct is non-literal; (4) would then amount to something like (8).

1. (8) Don't hit me, and I'll hit you back.

This is clearly not what S means by (4). One can envisage, as a second possibility, that both conjuncts are used non-literally such that (4) amounts to something like (9).

1. (9) Don't hit me and I won't hit you back.

Conjunctions of the form !p and q (where ! stands for the imperative mood) entail the corresponding conditionals if p, q; thus (4) entails (10).

1. (10) If you hit me, I will hit you back.

However, what (9) entails is (11), not (10):

1. (11) If you don't hit me, I won't hit you back.

To be sure, (11) can be ‘perfected’, that is pragmatically enriched, into a biconditional (e.g. Geis and Zwicky Reference Geis and Zwicky1971; Horn Reference Horn2000) that would ground the entailment relation between (9) and (10). However, ‘conditional perfection’ is a pragmatic, hence defeasible interpretive process, so that (9) – which is supposed to be what is meant by (4) – is compatible with the falsity of (10). The examples in (12) are typical instances of cancelled pragmatic enrichment.

1. (12)

   1. a. Don't hit me and I won't hit you. But/actually, even if you hit me, I won't hit you back.

   2. b. If you don't hit me, I won't hit you back. But/actually, even if you hit me, I won't hit you back.

By contrast, the threat in (4) proves infelicitous if the conditional in (10) turns out to be false, and the examples in (13) are sheer contradictions.

1. (13)

   1. a. Hit me, and I'll hit you back. ^?But/ ^?actually, even if you hit me, I won't hit you back.

   2. b. If you hit me, I'll hit you back. ^?But/ ^?actually, even if you hit me, I won't hit you back.

It can be concluded from the foregoing that non-directive uses of the imperative mood are as literal and direct as the directive ones. This I take to be a clear case against building the illocutionary force within sentence meaning (for the same point, see Wilson and Sperber Reference Wilson, Sperber, Dancy, Moravcsik and Taylor1988).

Scholars eager to pair sentence meanings with illocutionary forces sometimes invoke typological data (e.g. Han Reference Han2000: 164). The rationale is along the following lines: ‘If natural languages bother to devote a specific form to directive speech acts, then the directive illocutionary force is not a matter of pragmatic processing, but part and parcel of sentence meanings.’ Sadock and Zwicky (Reference Sadock, Zwicky and Shopen1985) claim that every language has a specific imperative sentence-type associated with directive speech acts. Yet, the morphosyntactic system of many languages lacks – totally or partially – distinctively imperative linguistic forms (this is the case for 122 languages out of the 552 analysed in van der Auwera and Lejeune Reference Auwera, Lejeune and Haspelmath2005). In languages with defective or empty imperative paradigms, various compensatory strategies can be found: aorist (e.g. Georgian), subjunctive mood (e.g. French or Armenian), optative mood (e.g. Eskimo), irrealis mood (e.g. Javanese), indicative mood (e.g. Hebrew) and, perhaps more surprisingly, passive forms (Maori) (see, for instance, Xrakovski Reference Xrakovski2001; Allan Reference Allan2006a; König and Siemund Reference König, Siemund and Shopen2007). If the directive illocutionary force were encoded within the linguistic meaning, how are we to explain these typological facts? Should we accept that, in languages that lack genuine imperative mood, some sentence-types are linguistically ambiguous?

First, such an ‘ambiguity’ thesis violates Grice's Modified Occam's Razor (which recommends avoiding the multiplication of linguistic meanings beyond necessity). Second, the ambiguity thesis proves hard to maintain across languages. To see this clearly, contrast the use of future in Nunggubuyu and in French. Nunggubuyu lacks any morphosyntactic imperative; the same construction is used to express future time reference and to perform directive speech acts (Heath Reference Heath1984, Reference Heath1986). Out of context, it is impossible to decide whether (14a) should be translated as (14b) or (14c).

1. (14)

   1. a. Ba‐buraː‐ (Verstraete Reference Verstraete2005; after Heath Reference Heath1984: 343)

      2sg(class‐A)³‐sit-nonpast

   2. b. ‘You will sit down.’

   3. c. ‘Sit down.’

In French, an authoritative way to order things is to use the future constructions. However, French has also a morphological second person imperative; accordingly, a literal translation of (15a) can only be (15b) – not (15c).

1. (15)

   1. a. Tu partiras demain.

      You.sg leave‐ind.fut.simple.2sg tomorrow

   2. b. ‘You will leave tomorrow.’

   3. c. ‘Leave tomorrow.’

As far as I can see, there is no principled ground for the following joint claims: (a) that the Nunggubuyu construction ‘Class-A prefix + non-past’ is ambiguous between two illocutionary forces, each such sentence being either a direct and literal assertion about the future or a direct and literal directive speech act; and (b) that in French requests performed with the future indicative are indirect.

But if future tense constructions are said to be ambiguous between the assertive and the directive forces across languages – thus in French too – why should we refrain from extending this rationale to other morphosyntactic forms? As I have just mentioned, there exists a wide range of linguistic strategies for compensating for defective imperative paradigms. Following the thread, one would have to assume linguistic ambiguity for any such form that happens to be prototypically used to perform directive speech acts in some language. Take Lingala, which has an imperative form for the singular second person only; in directive speech acts with second person plural, subjunctive mood is used instead (van der Auwera and Lejeune Reference Auwera, Lejeune and Haspelmath2005). On the one hand, no sensible semantic theory of mood would consider it plausible that, in Lingala, the ambiguity between a directive and some other illocutionary force might characterise the plural second person subjunctive form only, and not extend to the singular second person subjunctive form. On the other hand, data on the distribution of moods militates against the ad hoc hypothesis of a cross-linguistic ambiguity of the subjunctive mood. For instance, French has both second person imperative and subjunctive forms; but the subjunctive proves unacceptable whenever imperative forms are available. To borrow an example from Schlenker (2005), one can advise the Queen to be prudent using either a third person subjunctive (16) or a second person imperative (17); by contrast, the second person subjunctive in (18) is deviant.

In (16), the Queen is addressed as a third person; second person addressed directive speech acts are unacceptable with the subjunctive.

Moreover, the acceptability of subjunctive in French is not linked to the presence or the absence of the directive force. For instance, in French equivalents of (4), the surface form of the first verb is unambiguously imperative (20a) (cf. note 2), and the first clause has clearly no directive force. In such environments, second person subjunctives are unacceptable (20b); however, in the third person – for which there is no morphological imperative form – the subjunctive is fine (20c).

At this point, I see no justification for arguing that a certain sentence-type – e.g. future in Nunggubuyu or subjunctive in Lingala – is ambiguous between several illocutionary forces, rather than admitting that illocutionary forces belong to the level of utterances rather than to that of sentences.⁵

9.3 Grice's heritage: illocutionary forces and utterances

In the previous section, we saw that illocutionary forces do not belong to the level of sentence meaning. From this, however, one should not conclude that whenever a sentence is uttered and acquires propositional meaning, a speech act has been performed eo ipso. Yet, a customary – but misguided – shortcut is precisely to recast the opposition between sentences and utterances in terms of the contrast between sentence meanings and speech acts. This opposition plays a major role in a much-discussed issue of contemporary philosophy of language: that of contextual contributions to the propositional content. To take a well-worn example, (21) will mean different things in different contexts – depending on what John is ready for.

1. (21) John is ready.

Phenomena of this kind have been taken to show that literal content cannot be determined by linguistic meaning alone. Yet, opposing this interpretation, so-called ‘semantic minimalists’ argue (a) that the contents expressed by (21) on different occasions of utterance are speech act contents; (b) that these speech act contents do not correspond to the literal, semantic content of the sentence uttered. In other words, while different utterances of the same sentence correspond to the performance of different speech acts, with possibly different contents, speech act contents are to be distinguished from the semantic content proper, which – except a restricted set of indexicals – is entirely determined by sentence structure and remains constant across contexts of uses (Cappelen and Lepore Reference Cappelen and Lepore2005; Soames Reference Soames2002). Implicit in this view is the assumption that any utterance constitutes a speech act, or, in other words, that for a sentence to be uttered amounts to its acquiring an illocutionary force.

Such an assumption can be traced back to Grice's notion of meaning_nn. Saying, for Grice, is intending to provoke some cognitive response, such that the reason for the cognitive response is the recognition of this very intention (or, at least, cases of saying can always be reconstructed in this way). The nature of the response S intends to provoke by her utterance in the mind of A determines the ‘central’ speech act the instance of saying corresponds to: if it is the belief that p, S's saying will be an assertive speech act; if it is the intention to bring about the truth of p, S's saying will be a directive speech act (cf. Grice Reference Grice1968, Reference Grice2001: 50–55).

Now take a sarcastic utterance like (7) above: S says that the party is great, but, clearly, she doesn't mean it. To the best of my knowledge, no theory of irony would claim that in such a case S asserts that the party is great. Furthermore, a sarcastic utterance that p is clearly not accompanied by an overt intention (to communicate that p) of the kind that, according to Grice, characterises saying. This entails, in Grice's view, that the sarcastic S does not say anything, but just acts as if she was saying something – a very counter-intuitive consequence (see Neale Reference Neale1992; Carston Reference Carston2002: 114–16; Kissine Reference Kissine2009).

The existence of cases where S says that p without asserting that p constitutes a strong argument against equating utterances and illocutionary acts. Austin (Reference Austin1975) explicitly distinguished between sentence meaning (phatic act), contextual meaning (locutionary act) and illocutionary act (for a detailed discussion, see Kissine Reference Kissine2008b, Reference Kissine2009). In other words, even though sentence meaning and utterance meaning are not to be confused, illocutionary force attribution constitutes yet another level of interpretation.

In a series of papers, Kent Bach (e.g. 1994a, 2005) argues that what is said by an utterance does not coincide with the content of the illocutionary act performed by this utterance; what is said corresponds to the locutionary act performed. However, unlike, for instance, Recanati (Reference Recanati2004a and this volume) or Carston (Reference Carston2002 and this volume), Bach defines what is said as the output of the semantic interpretation of the syntactic structure; pragmatic contributions to what is said are limited to determining the reference of a restricted set of indexicals. Moreover, in Bach's view, what is said does not always correspond to a full-blown proposition. Take (21), for instance. As pointed out above, it is impossible to decide what John is ready for in the absence of any contextual information. But such extra-linguistic information does not contribute, according to Bach, to what is said. Illocutionary content is the place where such pragmatic influences come into play. What is said by (21), for Bach, is the sub-propositional radical [John is ready___], whose empty slot has to be completed at the illocutionary level.

One of the reasons Bach – rightly – invokes for distinguishing between locutionary and illocutionary acts is that sometimes we do not mean what we say; irony is a case in point. Accordingly, in such cases, the utterance reduces, at the literal level, to a locutionary act that corresponds, in Bach's view, to the semantic interpretation of the syntactic structure and to the indexical resolution, both of which remain blind to the pragmatic, wide context.

Bach's conception does not conform to Austin's original definition of the term locutionary act (cf. Kissine Reference Kissine2008b, Reference Kissine2009). More importantly, while the existence of cases where we do not mean what we say justifies the rejection of Grice's equation of saying with performing a speech act, it also shows that ‘forceless’ saying – the locutionary act – is already endowed with propositional, pragmatically determined meaning. Imagine that John is notoriously bad at preparing his talks on time. Imagine that, a few hours before John's talk, S sarcastically says:

1. (22) Of course, John is ready.

Whatever theory of irony one favours, it is clear that in saying (22) S does not mean what she says. But in order to determine what she does mean – something like ‘Of course, John is not ready for his talk’ – one has to know what S does not mean. And at this stage one cannot avoid determining what John is ready for in (22). In other words, S's locutionary act – what she says in (22) – is not, pace Bach, a sub-propositional radical: it is a full-fledged, contextually determined proposition.

There is another argument to support this claim. The verb say, as used in indirect reports of the ‘S said that p’ kind, is ambiguous, according to Bach (Reference Bach and Szabó2005), between locutionary and illocutionary meanings. Therefore, Bach must accept that in reporting the ironic utterance in (22) by (23) one transmits S's locutionary act (since her illocutionary act has a different content):

1. (23) S said that John was ready.

Some may feel inclined to say that the report in (23) is false in a context where the audience does not have access to the non-literality of S's original utterance. However, no sense could be made of (24) if (23) were not true:

1. (24) S said that John was ready, but she didn't mean that/it.

Even though the truth of (23) does not depend on the possibility for the audience of recovering the non-literality of the original utterance, it remains impossible to maintain that the embedded clause in (24) is sub-propositional. The main reason is, of course, that the reference of the demonstrative that or of it in (24) must be a full-fledged proposition.

9.4 Locutionary acts

In section 9.2, we saw that illocutionary forces do not belong to sentence meaning. In the previous section, we saw that some utterances do not have any (direct and literal) illocutionary force, even though they are endowed with propositional and context-dependent meaning. However, until now the discussion was limited to grammatically declarative sentences. A chief reason for introducing the notion of a locutionary level – intermediate between sentence meaning and illocutionary level – is, as we have seen, the existence of cases where what is said differs from what is meant. Parallel cases of non-literality can be found with imperative sentences. In the following example, S does not literally request, order, allow or wish A to ruin her carpet.

1. (25) [A spills his glass of wine over the carpet, and clumsily attempts to wipe it off. S says:] Go on! Ruin my carpet!

Exactly as S does not literally assert anything when she sarcastically says that the party is great, the speaker of (25) does not perform any literal directive speech act. But if so, what is the literal content of (25)?

In section 9.2, I have argued that the meaning of the imperative mood should not be analysed in illocutionary terms. Without getting into details, imperative clauses may just be, at the literal level, expressions of propositional contents under a certain attitude or with a certain mode of presentation. Let me just quote two possible accounts, among many. According to the first, the imperative mood presents the utterance content as desirable and potential (Wilson and Sperber Reference Wilson, Sperber, Dancy, Moravcsik and Taylor1988; Clark B. Reference Clark1993; for critical discussions, see Dominicy and Franken Reference Dominicy, Franken, Vanderveken and Kubo2002; Schwager Reference Schwager2006). According to the second, the imperative mood functions like a necessity operator; roughly, the ‘attitude’ bearing on the propositional content would be derived from the context-dependent base (i.e. from the domain of quantification) of the modal (Schwager Reference Schwager2006). Independently of the account favoured, locutionary acts must be thought of not only as having propositional content, but also as endowed with a certain mode of presentation of this content (Kissine Reference Kissine2009).

Assessing the semantic accounts of sentence types – e.g. imperative, indicative, subjunctive, interrogative – falls far beyond the scope of this chapter. The important point is that the semantics of sentence types predicts which locutionary-act types the utterances of the sentence can constitute; the nature of the locutionary act performed constrains, in turn, the range of those direct illocutionary acts the locutionary act can constitute.

9.5 Forceless meaning and indirect speech acts

Conceiving of illocutionary forces as optional properties of utterances allows a fresh perspective on indirect speech acts. Classically, a speech act is said to be indirect whenever its uptake (i.e. A's understanding the utterance as being this speech act) is tied to the uptake of another speech act (Searle Reference Searle, Cole and Morgan1975b; Bach and Harnish Reference Bach and Harnish1979: 70; also Recanati Reference Recanati1987). For instance, if S utters (26) as an answer to A's offer to go to the movies, A can infer, by using various conversational, cooperation-based principles, that in addition to stating that she is tired, by (26) S rejects A's offer.

1. (26) I'm very tired.

Some indirect speech acts are highly conventionalised. For instance, although (27) has an interrogative syntactic structure, it constitutes, in English, an extremely conventionalised means to request things.

1. (27) Can you pass me the salt?

After Morgan (Reference Morgan and Cole1978), it is customary to think of such cases as ‘short-circuited’ implicatures. In a nutshell, Morgan's idea is that the link between (27) taken as a literal question about A's ability to pass the salt and (27) taken as a request to pass the salt can be reconstructed in Gricean terms; however, such an inference generally does not take place. This is so because the link between the Can you___? construction and the directive interpretation is highly conventionalised and largely automatic. (Of course, conventionalisation is not arbitrary. In particular, the link between the literal meaning and the indirect force must be easy to grasp. Not every question of the form Can you___? will easily receive a directive interpretation. For instance, that S is able to pass the salt is a preparatory condition which must be fulfilled in order for a request to pass the salt to be successful. According to Searle (Reference Searle, Cole and Morgan1975b), this is why a conversationally irrelevant question about a preparatory condition will be readily interpreted as the performance of the corresponding request.)

A view rival to the ‘short-circuited implicature’ account is that of Sadock (Reference Sadock1974), according to whom each sentence-type is associated with an illocutionary force at the syntactic level. Under such an analysis, (27) is linguistically ambiguous between the interrogative and imperative meanings. As we have seen in section 9.2, the doctrine of a linguistic coupling between sentence types and illocutionary forces is problematic for independent reasons. There is one point of Sadock's analysis that is worth considering here, though. According to him (1974: 97–109), certain grammatical properties distinguish questions used as requests, and conventionalised forms whose meaning is (allegedly) ambiguous between a question and a directive speech act. Take (28) as an instance of the former case, (29) of the latter.

1. (28) When will you close the door?

2. (29) Will you close the door?

Sadock assumes that only grammatically imperative sentences can be followed by please or by an indefinite vocative.

1. (30) Close the door, please.

2. (31) Close the door, someone.

The unacceptability of (32)–(33) shows, according to Sadock, that (28) is an indirect request performed by means of uttering an unambiguously interrogative sentence.

1. (32) ^?When will you please close the door?

2. (33) ^?When will you close the door, someone?

By contrast, the acceptability of (34)–(35) (allegedly) reveals that (29) is linguistically ambiguous between being a grammatically imperative sentence – hence a direct request to close the door – and being a grammatically interrogative sentence – hence a direct question about S's ability to close the door.

1. (34) Will you please close the door?

2. (35) Will you close the door, someone?

In reaction to Sadock's argument, Bach and Harnish (Reference Bach and Harnish1979: 200–202) point out that please is also acceptable in (36).

1. (36) Can you reach the salt, please?

If (27) is, grammatically, a request to bring about the truth of the propositional content, one should expect (36) to be a request to reach the salt. This is counter-intuitive: (36) is a request to pass the salt, not to reach the salt. Likewise, as pointed out by Bach and Harnish, the following example is clearly not ambiguous between an indicative and an imperative underlying structure, despite the acceptability of please.

1. (37) I'd like some salt, please.

Following Morgan's lead, Bach and Harnish claim that certain questions of the form Can you___? or Will you___? are standardised means to perform requests. In order to explain away Sadock's grammatical constraints, they argue that (29) is ungrammatical, although pragmatically acceptable.

Both Sadock's and Bach and Harnish's accounts presuppose that the grammatical acceptability of please is linked to the imperative mood. However, the acceptability of please or of the indefinite vocative does not depend on the utterance's mood, but on whether or not the utterance's (primary) illocutionary force is directive. The following examples are acceptable only if S believes (or pretends to believe) A to have control over his having a nice journey back or getting well soon, that is only if the utterances at hand constitute (pretend) directive speech acts.

1. (38) ^?Have a nice journey back, please.⁶

2. (39) ^?Get well soon, please.

Even under such a reading, indefinite vocatives prove unacceptable:

1. (40) ^?Have a nice journey back, someone.

2. (41) ^?Get well soon, someone.

When the imperative clauses cannot receive a directive illocutionary force, as in (4) and (42), please and the indefinite vocative are clearly ruled out.

1. (4) Hit me and I'll hit you back.

   [repeated]

2. (42) Be tall and people will be respectful.

3. (43) ^?Hit me, please, and I'll hit you back.

4. (44) ^?Be tall, please, and people will be respectful.

5. (45) ^?Hit me, someone, and I'll hit you back.

6. (46) ^?Be tall, someone, and people will be respectful.

It thus seems that the acceptability of please does not depend on grammatical factors – on the sentence mood – but on pragmatic ones – the utterance's illocutionary force. What do we make of the fact that the adjunction of please and of someone is constrained by the presence of the directive illocutionary force and not of the imperative mood? According to Bach and Harnish's standardisation thesis, when an indirect speech act is conventionalised, hearers automatically derive the secondary indirect force without going through the derivation of the primary, literal speech act. That is, (29) is directly understood as a request to close the door, whereas (28) is interpreted as a question, and it takes supplementary pragmatic reasoning for A to understand it as a request to close the door.

So why not say that (29) has the force of a request only, whereas (28) has the primary force of a question, and, indirectly, constitutes a request? We have seen above that some utterances constitute a locutionary act with a certain content p but do not correspond to any illocutionary act with the content p. The same applies here. In (29), the content of the directive illocutionary act – of the request to close the door – differs from that of the locutionary act which the utterance of the interrogative sentence corresponds to.⁷ Since the only illocutionary force of (29) is a directive one, it is also, trivially, the primary force. As expected, constructions whose acceptability depends on the primary force being directive are allowed to take please. By contrast, the literal force of (28) is that of a question; since the directive force is not primary in this case, please is not allowed.

To repeat, (29) is syntactically and semantically an interrogative; however, it is not interpreted – nor intended by S to be interpreted – as a question. Such a rationale presupposes that literal and serious utterances of interrogative sentences are not necessarily associated with the act of requesting information, exactly in the same way as not all utterances of imperative sentences constitute directive speech acts. Interrogatives that are neither requests for information nor expressions of ignorance include the following: rhetorical questions (47), exam questions (48), guess questions (49), and surprise questions (50).

1. (47) [Peter, who had made a New Year resolution to give up smoking, lights up. Mary says:] What was your New Year's resolution?

   (from Wilson and Sperber Reference Wilson, Sperber, Dancy, Moravcsik and Taylor1988)

2. (48) Where did Napoleon die?

3. (49) Which hand is the marble in?

4. (50) A: The President has resigned.

   B: Good heavens. Has he?

   (from Wilson and Sperber Reference Wilson, Sperber, Dancy, Moravcsik and Taylor1988)

For a careful and illuminating account of the relationship between the interrogative mood and the speech act of questions the reader is referred to Fiengo (Reference Fiengo2007), who analyses interrogative sentences as expressing incomplete or truth-valueless propositions.

Further arguments in support of the view that conventionalised indirect speech acts like (27) or (29) are, at the literal and direct level, cases of saying without performing an illocutionary act, can be found in Terkourafi (Reference Terkourafi, Brabanter and Kissine2009c), who provides an interesting discussion of the relevant experimental data.⁸ On the one hand, experiments reveal that the putative direct force of a Could you___ request like (27), namely the illocutionary force of questioning, seems to be ignored in favour of the intended speech act of request. On the other hand, the interrogative form of the sentence is processed, as shown by answers such as ‘Yes, I can’ or by the fact that the interrogative form is recalled (for an extensive discussion and references, see Terkourafi Reference Terkourafi, Brabanter and Kissine2009c).

We thus arrive at the following picture of the possible relationships between locutionary and illocutionary acts:

1. a) The locutionary content corresponds to the content of the primary, direct speech act. This is the ordinary case.

2. b) The locutionary act does not constitute any direct speech act; the only speech act performed by the utterance is indirect. Here, two further sub-categories must be distinguished.

   1. i) The utterance is non-literal; S does not endorse the locutionary content. For instance, this content ‘echoes’ an utterance or a thought of another (possibly virtual) person (e.g. Sperber and Wilson Reference Sperber, Wilson and Cole1981; Wilson Reference Wilson2006). The important point is that, in such cases, S's performance of the indirect speech act cannot be reconstructed as an inference taking as one of its premises the performance of a primary speech act that shares its content with the locutionary act.

   2. ii) The utterance is literal, but the content of the only illocutionary act the utterance constitutes is distinct from that of the corresponding locutionary act. However, it is possible to reconstruct the interpretation process as starting from the performance of an illocutionary act whose content is identical with that of the locutionary act.

The contrast between ironic utterances – point b(i) – and conventionalised indirect speech acts – point b(ii) – deserves a little more discussion. Morgan's (Reference Morgan and Cole1978) idea in treating conventional indirect speech acts as ‘short-circuited’ implicatures is precisely that even though the indirect speech act can be derived from the putative performance of a direct speech act, such an inference does not actually take place: instead, A jumps directly to the indirect speech act. Take (27).

1. (27) Could you pass the salt? [repeated]

One possible Gricean reconstruction of A's interpretation of (27) runs as follows (cf. Searle Reference Searle, Cole and Morgan1975b):

1. (51)

   Step 1:

   S is asking me whether I have the capacity to pass the salt;

   Step 2:

   S probably knows that I have this capacity;

   Step 3:

   S knows that I know that she knows that;

   Step 4:

   So, S believes that I understand that she does not want to be informed as to my capacity to pass the salt;

   Step 5:

   We are at a dinner, and it is possible that S needs salt;

   Step 6:

   Being able to pass the salt is necessary in order to pass the salt;

   Step 7:

   So, by asking me whether I am able to pass the salt S requests me to pass her the salt.

However, A's exposure to conventions of language use allows him to jump directly from his recognition of S's utterance of an interrogative sentence of a certain form – which does not amount to a question – to interpreting S's utterance as a request to pass the salt. The important point about the rational reconstruction of the interpretation of an indirect speech act like (27) is that the first step – S's performance of the primary speech act of questioning – remains compatible with the last step – S's performance of a request. This reconstruction parallels that of genuinely indirect speech acts like (28).

1. (28) When will you close the door? [repeated]

The fact that, by uttering (28), S asks A when he will close the door is compatible with the fact that S, by means of this same utterance, requests A to close the door.

Now, contrast this with the ironic utterance of (7):

1. (7) This party is great. [repeated]

In order to understand what S really means by (7), A has to understand that S does not assert what she is saying. As in (51), let us try a rational reconstruction that would start with the premise that S performed a direct and literal speech act.

1. (52)

   Step 1:

   S is asserting that the party is great;

   Step 2:

   This party is all that S hates;

   Step 3:

   So, most probably, S does not believe that the party is great;

   Step 4:

   S is cooperative and would not violate conversational Maxims gratuitously;

   Step 5:

   S believes that I believe that S does not believe that the party is great;

   Step 6:

   S does not assert that the party is great;

   Step 7:

   S means that the party is awful.

Whatever the details, the important point is that, this time, and by contrast with (51), taking the last two steps requires falsifying the first one.

Note also that both rational reconstructions (51) and (52) presuppose that A is capable of making hypotheses about S's beliefs about A's mental states – i.e. that A has the capacity to attribute second-order or third-order mental states. Take first the reconstruction in (51). In order to get to the conclusion that S does not merely want him to say whether he can pass the salt, A must assume that S believes that A knows that S is not interested in (merely) knowing the answer. It would be irrational for S to use a question in order to request something, if she were fairly certain that A would not understand this. That is, A must attribute to S beliefs about A's beliefs about S's beliefs. Things are similar for (52): if A does not understand that S believes that A believes that S does not like the party, A cannot tell the difference between S's telling a lie and S's being sarcastic (for a more detailed discussion see Kissine Reference Kissine2008a).

From the foregoing, we can draw an important empirical reason for not taking rational reconstructions of indirect speech act interpretation as reflecting actual interpretive processes. Children do not master second-order mental state attribution before the age of 7 (Perner and Winner Reference Perner and Winner1985). Importantly, this cognitive ability seems to be required for understanding irony and for lying in an efficient way (Winner and Leekam Reference Winner and Leekam1991; Talwar and Gordon Reference Talwar and Gordon2007). By contrast, it has been repeatedly shown that well before 7, children respond adequately to and produce (conventionalised) indirect requests (e.g. Bates Reference Bates1976: 275–82; Shatz Reference Shatz1978; Reeder Reference Reeder1978; Carrell Reference Carrell1981; O’Neill Reference O’Neill1996), which reveals that this pragmatic ability does not require complex mind-reading skills, contrary to what is implied by rational reconstructions of the kind of (51).

9.6 Indirect speech acts and explicit performatives

Explicit performatives constitute one of the oldest and most vexing topics in the history of theorising about speech acts (for recent surveys, see Harnish Reference Harnish, Grewendorf and Meggle2002, Reference Harnish, Brisard, Meeuwis and Vandenabeele2004). While, again, I will not attempt an exhaustive review here, the fact, discussed in the previous section, that conventionalised indirect speech acts do not have more than one illocutionary force has an implication for the analysis of performatives that is worth considering.

Prototypical performative sentences have the form ‘I VP ___’, where the VP stands for the illocutionary act the utterances of these sentences constitute under normal conditions. Here are some examples:

1. (53) I order you to leave this room.

2. (54) I promise that I'll come to your party.

3. (55) I name this ship Queen Elizabeth.

Utterances of (53)–(55) are generally self-verifying: they constitute the act named by the matrix verb. By uttering (53), S makes it the case that she has ordered A to leave; by uttering (54), S makes it the case that she has promised to come to the party; by uttering (55), S makes it the case that the ship is named Queen Elizabeth.

The ‘self-verifying’ character of performatives like (55) can be explained by extra-linguistic, culture-specific institutions (for a more extensive discussion, see Kissine forthcoming). However, it is highly doubtful that the same treatment can be applied to (53)–(54) (for an attempt, see Searle Reference Searle1992a; for a cogent criticism, see Bach and Harnish Reference Bach and Harnish1992). The challenge consists in explaining how (53) is interpreted as an order and (54) as a promise without, at the same time, giving up commonplace semantics – i.e. by deriving the truth-conditional content of (53)–(54) through the same compositional semantic principles as for other structures of the form ‘I VP that___’, as in (56), or other present simple forms of order, as in (57).

1. (56) I hope that you are well.

2. (57) He orders you to leave.

That explicit performatives do not involve any extraordinary semantic features is further suggested by the observation that in some circumstances a sentence like (53) does not constitute a directive speech act at all, and has the content predicted by standard truth-conditional interpretation.

1. (58) A: Imagine that I light up a cigarette. What is your reaction?

   B: I order you to leave this room.

Another interesting fact is that the following examples (59)–(62) are as well suited as (53)–(54) to performing orders and promises.

1. (59) Leave the room, and that's an order.

2. (60) I'll come to your party, and that's a promise.

3. (61) Leave the room, I order you.

4. (62) I'll come to your party, I promise.

The analysis of examples in (59)–(60) is pretty straightforward. The speaker first utters a sentence, and then, with the second conjunct, indicates its illocutionary force, the demonstrative that picking up the first conjunct. As for (61)–(62), according to the account developed by Potts (Reference Potts2005; also Bach Reference Bach1999a), utterances with parentheticals express two propositions. The first is the main one – ‘at-issue’ – while the second is secondary. This second proposition may have a ‘procedural’ role (cf. Wilson and Sperber Reference Wilson and Sperber1993a), in that it facilitates the processing of the at-issue information. Following this widely accepted line of thought, in (61)–(62) the secondary propositions – the parenthetical ones – help attribute the illocutionary force to the main content.

The spirit of Davidson's (Reference Davidson and Margalit1979) ‘paratactic’ account of explicit performatives is quite similar. In his view, the surface form of, for instance (53), hides the following two utterances, where the first demonstrates the second one: I order you that. You leave the room. However, it is pretty clear that the that in (53) is not a demonstrative. Compare with French where there is no homophony between demonstratives and complementisers or relative pronouns.

Although it is not viable, Davidson's account is attractive in that it puts the explicit performatives in (53)–(54) and (59)–(62) on the same footing; the utterance's main informational content is assigned an illocutionary force through a comment on this content.

As for Bach and Harnish (Reference Bach and Harnish1979: 203–33; Reference Bach, Cohen, Morgan and Pollack1992), they claim that explicit performatives are standardised (conventionalised) indirect speech acts, exactly in the same way as indirect requests of the kind of (27).

1. (27) Could you pass the salt? [repeated]

Recall from the previous section that, on their view, although at the direct level (27) is a question, standardisation allows one to interpret it as a request without going through the inference from the direct to the indirect force. Likewise, they argue, explicit performatives like (53)–(54) are, at the direct level, ‘constative’ speech acts. The constative class includes those speech acts that aim at transferring information, assertion being a paradigmatic case. Bach and Harnish thus claim that by uttering (53) S asserts that she orders A to leave the room; this assertion is interpreted as an indirect order. However, because the inference from the direct to the indirect force is ‘compressed by the precedent’, (53) is a standardised indirect order exactly in the same way as (27). Bach and Harnish thus predict both that the content of (53) is the proposition [S orders that A leaves] and that (53) is an order.

Reimer (Reference Reimer1995) objects to Bach and Harnish's account that it does not seem intuitively plausible that S asserts anything by means of (53); the constative part of performatives is introspectively inaccessible. However, it is not obvious that the (alleged) question part of indirect requests like (27) is always accessible either. If, when solving a problem, I ask you ‘Can you tell me the square root of 16?’, it would be highly counter-intuitive to interpret my utterance as a question (e.g. Clark H. Reference Clark1979). In any event, nothing implies that conventionalised indirect speech acts have two illocutionary forces. We have seen above that indirect speech acts do not necessarily have a direct illocutionary force; in conventionalised indirect speech acts the content of the locutionary act is different from that of the illocutionary act, but only one illocutionary act is performed. Things are not different with explicit performatives. Recall that locutionary acts are endowed with a full-blown propositional content. Therefore, at the locutionary level explicit performatives have exactly those truth conditions that would be predicted by regular compositional semantics. The propositional content of the locutionary act performed in (53) is [S orders A to leave]. Yet, this does not imply that S asserts this content. The only illocutionary act performed in (53) is the order that A (should) leave. We can reconcile two seemingly conflicting intuitions: on the one hand, an explicit performative has the same propositional content as would an assertion constituted by an utterance of the same sentence; on the other hand, it does not constitute such an assertion.⁹

An interesting example in favour of the account of explicit performatives just sketched can be drawn from Allan (Reference Allan2006a). He points out that (65) is adequately paraphrased by (66):

1. (65) In the first place I admit to being wrong; and secondly I promise it will never happen again.

2. (66) The first thing I have to say is that I admit to being wrong; and the second thing I have to say is that I promise it will never happen again.

In (66) the second thing does not refer to a second act of promising, but to the second act of saying, that is to the explicit performative I promise it will never happen again; ditto for secondly in (65). That explicit performatives are acts of saying is consonant with the analysing of saying as the performance of a locutionary act. This is not to say that explicit performatives cannot be subject to a rational reconstruction, starting from the premise that S asserted that she, say, ordered A to leave the room. However, to repeat, rational reconstructions do not aim at modelling actual interpretive processes. Note that the fact that the interpretation of explicit performatives can be reconstructed on the basis of primary assertions does tell us something: that S is ready to endorse the content of the locutionary act – that she's not ironic. That is, it tells us that explicit performatives can be interpreted as assertions; but it remains that, in most cases, explicit performatives are not interpreted as assertions.

9.7 By way of conclusion: illocutionary force attribution

A natural question to ask, at this point, is how illocutionary forces are attributed. We have seen that developmental evidence makes it unlikely that illocutionary force attribution is underpinned by Gricean inferences about multi-layered communicative intentions (for a classical version of such an account, see Bach and Harnish Reference Bach and Harnish1979). Very young children are good at attributing illocutionary forces to utterances well before being able to attribute complex mental states required by Gricean reconstructions. As for contemporary pragmatic theories, as I said in section 9.1, they are fairly elusive on psychological mechanisms underlying illocutionary force attribution.

In Kissine (Reference Kissine2009) I argued that speech acts should be thought of as (not necessarily effective) reasons for S to believe that the propositional content is true (assertives), for A to bring about the truth of the propositional content (directives) or for S to bring about the truth of the propositional content (commissives).¹⁰ An important feature of this analysis is that it does not require advanced mind-reading skills for attributing illocutionary forces, except for complex communicative moves like irony (cf. Kissine Reference Kissine2008a).

Whenever the content of a speech act corresponds to the content of the constitutive locutionary act, this speech act is literal and direct. We have seen that utterances often constitute locutionary acts, that is they express propositional contents under a certain mood of presentation. The type of the mode of presentation constrains the range of the possible direct speech acts the locutionary act may constitute. For instance, if the imperative mood expresses an attitude characteristic of desiderative mental states, the potential direct illocutionary force will be a directive one.

However, there does not necessarily exist a one-to-one correspondence between sentence types and modes of presentation of the propositional content. In particular, it is fairly possible that the indicative mood is neutral in this respect (e.g. Recanati Reference Recanati1987; Allan Reference Allan2006a). An indicative sentence like (67) can be performed either as an assertion that A leaves the city on Monday or as an order that A leave the city on Monday.

1. (67) You leave the city Monday.

Let us assume that the indicative mood does not constrain the locutionary-act type. Illocutionary force attribution and the interpretation of the utterance type as a locutionary act with a certain mode of presentation will thus mutually influence each other. Note also that, in both cases, the propositional content is the same; it thus follows that the order in (67) is as direct as the one in (68); cf. Recanati (Reference Recanati1987) for the same conclusion.

1. (68) Leave the city tomorrow morning.

This is in tune with the widely acknowledged fact that most interpretive processes operate on-line. The precise ways sentence types, locutionary modes of presentation and illocutionary forces interact constitute a rich matter for future research.

10.1 Introduction

A person who listens to, understands and accepts the utterance of another acquires new information. There may be new beliefs that are about the main event described in the utterance, but crucially understanding and accepting it entails obtaining new information about the intentions of the other speaker. And if the intention is to inform, it also entails obtaining new information about the world. As long as there is no intention to correct old information, the new information can be described as an update of the beliefs of the hearer (and as a ‘downdate’ followed by an update if the new information is a correction of existing information). The direct description of the utterance as an update of an existing information state and the claim that this cannot be reduced to more fundamental operations and is the suitable framework for natural language semantics are characteristic of update semantics.

Probably the most basic way to think of update semantics in relation to pragmatics is as a natural environment for formalising pragmatics. Nearly everything in pragmatics can be described as (partly or completely) determining a change to an information state, the information state of the hearer, of an idealised hearer, of the speaker or the speaker's model of the hearer or of the common ground or the common ground according to one of the speakers. It seems fair to say that a pragmatic theory that cannot be described as characterising aspects of the up- or downdates in a context that models (real or imagined) hearers in a context interpreting the utterance is hard to imagine. On the other hand, many pragmatic approaches are still not theories that make precise and testable predictions and formalisation is very much needed to turn this area of study into proper science. So update semantics and update-semantical approaches to pragmatics are an important field of study even if e.g. one rejects the ideology of update semantics or some of the ideas that have been formalised in it. It still would seem the natural format for formalising any pragmatic ideas, even if one would like to include new mechanisms or reinterpret some of the mechanisms contained in the literature.

In this sense, up- and downdating information states is just a technical tool for being precise about what utterances do to the various information states involved in communication (the common ground between speaker and hearer, the hearer's and the speaker's knowledge and beliefs) and the natural successor of the commitment slates of Hamblin (Reference Hamblin1971) or the version of those that is employed by Gazdar (Reference Gazdar1979) or by Karttunen (Reference Karttunen1973) in the definition of filters. The tool has the advantage that whatever semantic assumptions one needs to make for one's pragmatics can be formulated directly in the same framework and in the same way.

What information states one wants to update depends on one's pragmatics. For semantic applications, it does not seem to matter. It suffices to assume somebody exposed to a series of trustworthy speech acts who is updating her own set of beliefs by accepting all these utterances and who happens to be an ideal hearer perfectly understanding the utterances she is exposed to. For pragmatical applications, the obvious candidate seems to be the common ground between speaker and hearer in a conversation. Both parties can update that and it would be possible to follow the progress of the conversation by taking in the effect of both parties’ moves. This would be third-party conversational analysis as proposed by Sacks et al. (Reference Sacks, Schegloff and Jefferson1974), but it is well known that it does not quite work for real conversations since misunderstandings and deceptions occur frequently. The best choice would be to have two separate versions of the common ground (or more if there are more speakers) and to try to track the evolution of all these common grounds. The common ground according to the speaker seems the most relevant information state from a linguistic point of view, since it can account for the speaker's choices in formulation and for the effect the speaker expects her utterances to have on the hearer. It also accounts for the hearer's choices in interpreting the utterances and for modelling the hearer in making choices in her formulation of those utterances, but in this second role, it does not account for the variation in linguistic form studied in linguistics. But there is no extra problem involved in formalising the third-party perspective: one just takes the model of the common ground of both parties.

Any serious semantic and pragmatic theory makes empirical predictions. Update semantics would be the theory with the shortest distance between the theoretical construct and what can be tested. What people assume after an utterance which they did not assume before is what can be tested almost directly. Much more than a truth-conditional account, update semantics seems to be about the primary function of linguistic utterances: to transfer information from one person to another (this also covers speech acts other than assertion).

Information states can be and have been modelled in many different ways. A simple and effective way is to start from a logical formalism with a proper model theory. Information states are consistent theories (sets of formulas that can be satisfied on at least one model for the theory) and updating is just adding new formulas to the set. Most of the older approaches like Hamblin's or Gazdar's follow this pattern, and Discourse Representation Theory (Kamp and Reyle Reference Kamp and Reyle1993) can be seen as an instance of this approach with a more sophisticated interface with syntax and maximal decomposition of formulas. It is also not difficult to make connections with tableau-based theorem-proving and to think of information states as sets of open branches with positive and negative literals (a formalism like that of Dekker (Reference Dekker, Dekker and Stokhof1992) comes close to that). Literals are here atomic formulas with constants, variables and skolem constants as the set of allowed terms. This may well be the best road towards implementing update semantics. Updating becomes the closure of those branches that become inconsistent under the new information and the addition of the necessary new literals to the branches that remain. Purely eliminative notions of update semantics are reached if the information state is a set of models, a set of possible worlds or a set of possibilities. Conceptual simplicity is typical for this last group of approaches; concreteness and thereby simple search procedures are typical for the first group, while mixed systems are best for implementational purposes. But it is not difficult to do translations between different formalisms. In this chapter, an elimination approach will be introduced in preference to other approaches because of its inherent simplicity.

10.2 Semantics and pragmatics

If pragmatics is incorporated in update semantics the distinction between semantics and pragmatics seems largely lost: in both cases one has updating operations. One can do some classification: some update operations are eliminative and distributive and can be called classical, but semantics seems to go further than just that.

For example, the analysis of may in Veltman (Reference Veltman1996) makes it a test on information states: does the information state have worlds in which its complement holds? If so, [mayϕ] is the trivial update; otherwise it is inconsistent information and leads to an absurd information state.

This would be half of the minimal analysis of assertion by Stalnaker, which requires assertions to be informative and consistent with the common ground. A variant of the analysis in which bad assertions lead to the absurd information state can be defined using may as: σ [may ϕ][mayϕ][ϕ].

The same holds for a proposal by Beaver (Reference Beaver2001) to analyse a presupposition trigger like bachelor as σ [pres (adult(x) ∧ male(x))][married(x)]. The semantics for pres ϕ proposed is also a test: σ[presϕ] = σ if σ[ϕ] = σ and undefined otherwise. This is close to the semantics for must that is most in line with the semantics for may. Being non-classical therefore does not mean being ‘pragmatic.’ The semantic/pragmatic distinction appealed to in attributing may a semantics is the distinction between lexically coded properties and properties that belong to the use of the expression. But it is questionable whether that distinction can be maintained. Conventional implicatures are lexically coded and so are presuppositions. Many particles are best seen as markers of rather unclear semantic features like additivity, formal contrast, denial of expectation etc. Their functional explanation seems related not to the expressive possibilities of the language, but to the difficulties of maintaining complex ranges of facts in memory.

One way to save the distinction would be to revert to a historical picture. An important feature of grammaticalisation processes is the semantics of the grammaticalised items becoming more pragmatical (‘semantic bleaching’). The existence of modal operators with ‘pragmatic’ meaning could then be explained as the outcome of such a process. The ‘ability’ readings with which ‘may’ started – as is generally assumed – would be classical updates. The process by which ‘ability’ became ‘epistemic possibility’ turned the classical update into a non-classical one. And this holds for the other cases as well. It is a standard assumption in historical linguistics that words start as either deictic elements or as descriptive (‘truth-conditional’ or property-expressing) elements. The elements that give trouble must therefore have come invariably from elements whose truth-conditional characterisation is unproblematic.

10.3 Presupposition

Karttunen's study of presupposition projection starting in the 1960s and carried on into the 1970s can be thought of as the origin of update semantics, but without the name and the slogan ‘the meaning of an expression is the change it brings to an information state’. For Karttunen, presuppositions are the prerequisite for using their triggers: a context that does not entail the presupposition cannot accept the use of the trigger that presupposes it. In Karttunen (Reference Karttunen1973) this leads to a classification of contexts of triggers into plugs, holes and filters. Plugs are sentential operators that do not let any presuppositional requirement out (e.g. verbs like ‘x says that p’), holes let all the presuppositional requirements out (not, maybe, x knows that) and filters are the ones that let some presuppositions out and not others (and, or and if…then are examples). To understand this properly, one must realise that Karttunen is interested in presuppositional requirements.² The requirement of a trigger like ‘is glad that’ can be fulfilled under ‘says that’ as in (15), and so it would be wrong to make the truth of the complement a requirement of the whole sentence.

1. (15) John said that Mary left and that Bill is glad that she did.

Filters are formalised in Karttunen (Reference Karttunen1974) in terms of information state, as in (16).

1. (16) X allows A_p iff X ⊨ p

   X allows A iff X allows A

   X allows A & B iff X allows A and X + A allows B

   X allows A → B iff X allows A and X + A allows B

   X allows A∨B iff X allows A and X + A allows B

The last of these has given some trouble, due to examples like

1. (17) Either the toilet is in a funny place or this house does not have a toilet.

There is still no fully satisfactory solution.

From the point of view of update semantics, this is an attempt to characterise presupposition in terms of properties of the input context for their triggers: the use of the trigger is inappropriate unless the presupposition holds in the input state. Update semantics can capture this ‘allowing’ by update restrictions along the lines of (18).

1. (18) σ[T] is undefined unless σ ⊨ p

The notion is problematic however. For some triggers (too and many other particles; definite descriptions) the place where the presupposition is true can be other than the local context of the trigger. In (19), B's parents presumably have no opinion about whether A is in bed, and John can think that Tim likes him, without having any idea that Tim is the hearer's brother.

1. (19) A: My parents think I am in bed.

   B: My parents believe I am in bed too.

   John thinks that your brother likes him.

And it ignores accommodation. Often – but not always and not for all triggers – the missing presupposition can be added, but when this happens it is more likely to happen in the global context than in the direct local context of the presupposition trigger. What one can say is that if the presupposition is not just missing but overtly absent (the speaker says ‘I do not know whether p’) or if it holds in the context of the trigger that p is false, the trigger is not acceptable.

Irene Heim realised that this can be stated better as an update semantics.³ In this way, truth conditions and allowance conditions are captured by one single definition.

Heim brings a second improvement to Karttunen's system, the addition of global accommodation. If a trigger is updated into an auxiliary information state and the presupposition p can be added either locally or globally, it is assumed that either the local state or the global state can be changed to be X[p], with a preference for the global state. This captures the predictions of Gazdar.

While Karttunen uses a concept of update semantics to explain filters, Gazdar's aim was the modern update semantics goal of stating the full change that utterance will bring to an input context. His contribution is flawed, because of an attempt to replace Karttunen's insights with an only seemingly equivalent formulation in which filtering is the effect of clausal implicatures. This does not work, witness (20).

1. (20) If John has a son, his child must be happy.

(20) would only produce clausal implicatures that the speaker does not know whether John has a son. This is not inconsistent with the presupposition of the definite description ‘his child’ which therefore does not get cancelled. So Gazdar predicts – against intuition – that ‘John has a child’ is projected.

It is, however, not difficult to amend Gazdar's idea with Karttunen-style filtering, as in (21). Here X is the given context (the information state in which the trigger is used).

1. (21)

   1. (i) The potential presuppositions of a simplex formula A under X are those given by the triggers in A that are not entailed by X.

   2. (ii) For O one of , □, ⋄, x Vs that, the potential presuppositions of OA under X are those of A under X.

   3. (iii) The potential presuppositions of A → B and QAB (for Q a positive quantifier) under X are the potential presuppositions of (A ʌ B) under X and those of A∨B under X those of (A ʌ B) under X.

   4. (iv) The potential presuppositions of A∧B under X are the potential presuppositions of A together with those of B that are not entailed by A under X.

The exception to the cumulative hypothesis given in Langendoen and Savin (Reference Langendoen, Savin, Fillmore and Langendoen1971) can be fully explained by Gazdar's view of triggers as signs that the speaker accepts the presupposition. Such a sign has no significance if the speaker has locally entailed that she accepts the presupposition in the context. The conversational contribution of utterance u can now be simply stated as

1. (22) X∪{u}∪!potIMP(u)

where potIMP(u) collects the potential scalar implicatures of u, the potential clausal implicatures of u and the potential presuppositions of u and ∪! is Gazdar's satisfiable incrementation: add an element x of potIMP to X∪{u} unless x makes a set consisting of X∪{u} and other members of potIMP inconsistent. The global projection of presuppositions in this view is treated quite rightly as yet another implicature,⁴ nowadays attributed by most who share the view that these projections are implicatures to Grice's maxim of relevance.⁵

While this is an attractive reformulation of Gazdar that improves on Karttunen and Heim by restricting global accommodation in interesting ways (conflicts with implicatures and other presuppositions), it falls short of Heim and Van der Sandt in one point: a cancelled presupposition only stays as a local entailment of the simplex clause in which the trigger occurs.⁶ This is probably too inflexible and the theories of van der Sandt (Reference Sandt1992) and Heim (Reference Heim, Barlow, Flickinger and Westcoat1983) offer the full space of auxiliary contexts necessary to put the presupposition in other places.

Van der Sandt's theory assimilates presupposition to the dynamic treatment of anaphora and is formulated in the DRT (a representational version of update semantics, as assumed in this paper). It incorporates the idea that presupposition triggers (just as pronouns) can only update the context if the context contains the presupposition (an antecedent for an anaphoric pronoun). Van der Sandt obtains the effects of Gazdar by allowing the insertion of a missing presupposition in any context to which the presupposition trigger has access and to which the presupposition can be consistently added, with a preference for the highest available one, a stronger assumption than in Heim, who just postulates a preference for the main context.⁷

All the theories treated so far fail on one important presuppositional phenomenon: partial resolution. This is the situation where an antecedent is present but not quite. Spenader (Reference Spenader, Kuehlein, Rieser and Zeevat2003) investigated corpora for too and Kamp and Rossdeutscher (Reference Kamp and Rossdeutscher1994) formalised some specific texts for the German wieder, ‘again’. In both studies, most antecedents turn out to be incomplete, i.e. one finds material that is intuitively the antecedent, but information needs to be added to it to make it a proper antecedent as required by the trigger. Kamp and Rossdeutscher (Reference Kamp and Rossdeutscher1994) have interesting suggestions about how to make the amendments and the treatment of these is more feasible in a representational theory than in an eliminative update semantics. Important is also to prevent it happening. For example, John is walking in the park. Mary will have dinner in New York too. does not lead to a repair in which John will have dinner in New York. (The example is due to Nicholas Asher, p.c.) This partial resolution happens with other triggers as well, as in McCawley's famous example (23).

1. (23)

   1. a. LBJ dreamt that he was a homosexual and everybody knew that his foreign policy was a failure.

   2. b. LBJ dreamt that he was a homosexual and everybody knew that he was waiting for boys in the restroom of the YMCA.

The examples show that (23b) is an example of resolution, since the presupposition does not project out. The processing that a sufficiently informed human interpreter of the sentence seems to go through seems to involve weighing several possibilities against each other, e.g. as in (24).

1. (24) Is LBJ waiting for boys in restrooms? Seems rather unlikely. Does everybody know this in his dream where he is a homosexual? Quite possibly.

This weighing of possibilities is important. (25a) is an example from Beaver (Reference Beaver2001).

1. (25)

   1. a. When Spaceman Spiff lands on planet X he will be surprised that he weighs more than on Earth.

   2. b. When Spaceman Spiff lands on planet X he will notice that he weighs more than 200 kilos.

In the (25b) example, Spaceman Spiff can weigh 200 kilos (in his own perception of his weight) when he lands on planet X, but this can also be his normal weight. In the latter case (this is serious overweight of the kind that happens), it would be extremely unlikely that he would be chosen to be a spaceman. The more local accommodation, that he weighs 200 kilos on planet X, is just as preferred as in the (a) case, but unlike the (a) case, it depends merely on probabilistic reasoning: the other possibility is consistent.

10.4 Implicature

Gazdar's formalisations of clausal and scalar implicatures are the first in an update perspective. They consist in the cancellable additions of the negations of scalar lexical alternatives further up the Horn scale and in the addition of Kϕ and Kϕ for clauses ϕ whose truth value is left undetermined by the matrix sentence in which they occur.

While there are many alternative treatments for the scalar implicatures, the treatment of clausal implicatures does not seem to have drawn much discussion. An obvious criticism (Rob van der Sandt, p.c.) is that the proposal appears to be too strong. It is perfectly possible to believe A or that A is false and at the same time state that John believes that A. But this is less of a problem in the context of the revised Gazdar solution from the last section, where it is easy to get cancellation of these implicatures. The implicatures can be connected to either quantity (if A is left undecided, while the speaker knows that A or that not A, he is being economical with the truth) or relevance: using the clause A is raising the question whether A and leaving it open is a sign that one cannot decide the matter. This also seems relevant for the example (26) discussed in Karttunen and Peters (Reference Karttunen, Peters, Oh and Dinneen1979), where in (26b) the medical speaker would take into account that there could be a view of the matter around under which the measles is not a possibility.

1. (26)

   1. (a) If he has the measles, he would have precisely these symptoms.

   2. (b) If he had the measles, he would have precisely these symptoms.

There are now many versions of update semantics where questions are used to obtain exhaustivity effects, including scalar implicatures. The idea combines an analysis of the topic/focus distinction in terms of answers to a question (Klein and von Stutterheim Reference Klein and von Stutterheim1987; van Kuppevelt Reference Kuppevelt1995) with the exhaustive interpretation of wh-questions (Groenendijk and Stokhof Reference Groenendijk and Stokhof1984). The trouble of course is that very often these questions are not overtly given, while at the same time the scalar implicatures and the exhaustivity effects are intuitively there.

Zeevat (Reference Zeevat, Aloni, Butler and Dekker2007) uses question updates to assign an exhaustive value to the wh-variable involved. In this way, John has two pigs can answer questions like How many pigs does John have? What animals does John have? Who has pigs? Who has two pigs? and give exhaustivity entailments like: the number of John's pigs is 2, the animals that John has are pigs, the one who has pigs is John, the one who has two pigs is John. The context should make it clear which of these questions apply.

Another way of going about this problem is to structure the set of possibilities in an information state into partitions (e.g. Jäger Reference Jäger, Dekker and Stokhof1996) and more recently in inquisitive semantics (an enterprise that goes well beyond just a treatment of these implicatures) by maintaining a possibility structure over the set of indices (Groenendijk Reference Groenendijk, Bosch, Gabelaia and Lang2009).

Another approach (Schulz Reference Schulz2007) integrates another aspect of relevance, utility for action, into the topic question. The treatment of relevance as a dimension of information structure is probably the biggest theme of update semantics.

Inquisitive semantics achieves a breakthrough by addressing the representation of questions in a way which also deals with the natural emergence of questions out of assertions. John is in Paris or London will lead to the information state that also represents the question whether John is in Berlin or Paris. And this should be generalisable to a similar mechanism for existentials: not just the information that there is one, but also the question which one.

Theories of topic and focus and many accounts of particles can profit from mechanisms that raise questions as well as give information. One would like to have predictive theories that add questions to the information state, such as a causal question if the information is surprising, a justification question if there is good reason for doubting the information, a background question if the setting is too unclear, more information questions when an object or a concept is not clear etc. This is often not a question of only logic.

10.5 Speech acts and discourse relations

There is surprisingly little work on speech acts within update-semantical pragmatics.⁸ Yet, the work of Austin (Reference Austin1962) and Searle (Reference Searle1969) fits in remarkably well. First of all, speech acts come with preconditions which can be captured much like other presuppositions, including accommodation. Then they can be seen as manifestations of intentions. The speaker wants to inform that p, wants to know who ϕ's or whether ψ, wants to commit herself to a course of action p, wants the hearer to do p, etc. Then the decision to use a speech act indicates a belief in its possible success, which normally can be cashed out as attitudes of the hearer: the hearer does not believe that p, the hearer may know whether p, the hearer would want the speaker to do x, the speaker has the necessary power over the hearer to order that p etc. Finally, the intention may connect to known goals of the speaker and allow further inferences. Especially if the utterance does not stand alone but is connected to an earlier utterance by the same or another speaker.

Then various forms of reaction to the speech act can be characterised in a similar way. The common ground that results from these speech acts may contain shared information, shared goals, shared emotions, contracts, shared questions, shared experience and shared plans.

10.6 Probabilities

A problem that has been equally ignored in update semantics as in other dynamic frameworks is the problem of pronoun resolution. It is normally put out of sight by coindexing or by employing the same variable in different updates. Or equally problematic: the choice between different readings of a word, for which – if the problem is noticed – a similar enrichment of the update with subscripts is chosen. What if one tried to integrate it? After all, the context should not just make the antecedents available but also serve as the content that allows a good choice between different antecedents. The context is normally supposed to play a similar role for word senses.

It is not easy to imagine something which is not similar to the picture that comes out of recent psychological investigations: the parallel activation of many readings with elimination of the less likely on the fly. An approximation is to have all the possible readings and a mechanism that computes conditional probabilities based on the probabilities of the predications on types (hit_ag(person, stone) vs hit_th(person, stone) vs hit_th(stone, person)) and combines them with information coming from the context. The construction requires a theory of how the information state can influence the probabilities and it is not directly clear how this can be achieved, though quite a number of things can be tried. The principle seems to be that if there is a clear probabilistic winner, it is the update, otherwise the update blocks (and feedback is sought). It is an unusual ingredient of a semantics, but it fits directly into the picture of an update semantics trying to model what happens to a hearer when exposed to utterances. It is also a necessary ingredient of any pragmatics. The process that selects word meanings, parses, antecedents, discourse relations and other unexpressed ingredients of the intended meaning is pragmatic because it stands under the normal pragmatic constraints. In particular, it stands under a constraint which makes the global interpretation as probable as possible, among the set of possible global interpretations.

It can be argued that this entails several operations that have traditionally been studied as part and parcel of pragmatics. One is a preference for connecting interpretations -- formulated in different traditions as ‘Do not miss anaphoric possibilities!’ (Hendriks and de Hoop Reference Hendriks and de Hoop2001), ‘Do not accommodate!’ (Blutner and Jäger Reference Blutner, Jäger, Fabricius-Hansen, Lang and Maienborn2003), ‘*New’ (Zeevat Reference Zeevat2009), ‘The Inertia Principle’ (Hamm and van Lambalgen Reference Hamm and van Lambalgen2005), ‘Minimal Models’ (Schulz Reference Schulz2007) -- over interpretations that are less connected. It can be constructed as consisting of preferring anaphoric (repeating) referents over bridging referents over brand-new referents, with a preference for connections to most activated referents. An anaphoric referent is the referent of some earlier expression, a bridging referent is a proper part of a given plural referent or a proper part of a given singular referent, or the cause of a given event. This principle has a surprising number of applications in the area of presupposition treatment and in the area of pronoun resolution, but also in the area of discourse structure.

It is not correct that a principle of this kind increases probability on the level of content. If p and p′ only differ in that p makes more contextual connections, p may well have a lower probability given the context. But on the level of what speakers are likely to intend with utterances such as p, they have a clear advantage: speakers seem to operate under the principle that where an identification is not intended, it is overtly marked that the entities are different. The result is that identifying interpretations are more probable, since they have been made impossible if unintended.

An explanation for this speaker strategy may lie in the fact that any kind of perception seems biased to identification and integration with the preceding experience. Natural language interpretation is just a special case of perception, so that speakers need to work against overidentification and overintegration.

A second principle is known as ‘Relevance’ (also as ‘Strongest Interpretation’ (Blutner and Jäger Reference Blutner, Jäger, Fabricius-Hansen, Lang and Maienborn2003), ‘Maximise Discourse Coherence’ (Asher and Lascarides Reference Asher and Lascarides2003), ‘Minimal Models’ (Schulz Reference Schulz2007) in many traditions, though the different traditions give different accounts and it is not directly clear that everybody intends the same. The author's favourite formulation is that interpretations that can be constructed as addressing a raised question win over minimally different interpretations that do not and lose out to identical interpretations that are taken as settling that question (‘Strongest Interpretation’ -- there are a number of instances of a strict interpretation of this principle that are counterexamples, and it should probably not be taken as a correct characterisation; see Beaver and Lee Reference Beaver, Lee, Blutner and Zeevat2003). ‘Maximise Discourse Coherence’ (Asher and Lascarides Reference Asher and Lascarides2003) seems to work in the same direction as this principle. Minimal Models either minimises the domain of the model (new objects are assumed only under force majeure, possibly by assuming that the new object is part of an old one), or alse ‘minimal’ refers to the size (extension) of certain relations.

The effect of the relevance principle is restricted by the conservativity principle: one cannot freely invent answers to a raised question. In the formulation given, it will deal with a wide range of phenomena again: scalar implicatures (27a), a side effect of settling quantity questions; exhaustivity implicatures (27b), a side effect of settling identity questions; presupposition projection (27c), a side effect of settling the question whether the presupposition holds taking into account that the speaker gives a signal with her use of the trigger that it does; certain explicature implicatures (27d); strengthening discourse relations (27e); relevance implicatures in the sense of Grice (27f) and even most of the other effects of relevance in Relevance Theory. (The symbol ‘+ >’ in (27) stands for ‘conversationally implicates’.)

1. (27)

   1. a. John has 2 cats and a canary. + > John does not have three cats.

   2. b. John has 2 cats and a canary. + > John does not have other pets.

   3. c. John hopes that his niece will visit. + > John has a niece.

   4. d. John had a drink. + > John had an alcoholic drink.

   5. e. John pushed Bill. He fell. + > Bill fell because of the pushing.

   6. f. There is a garage around the corner. + > The garage can help you to get some petrol.

Settling raised questions produces more information than keeping these questions open and so again works against probability. The set of minimal models is also a part of the set of all models, so again, probability can at best go down. But again it holds that speakers who want to avoid such inferences should mark against them, i.e. that not marking against them is sufficient for being taken to intend them. Again, the tendency of interpreters to make them can be reduced to a bias in perception (seeing what you want to see).

The author would conjecture that approximations to content probability together with symbolic implementations of the two principles discussed above would give promising results.

The given information state functions three times in such a programme: as a source of information which determines the relevant conditional probability that would decide between different interpretations of the new contribution; as a source of the objects given in the context and of their activation levels, for the second principle; and as a source of activated questions and their activation levels, for the third.

The last role highlights once again the importance of integrating the other speech acts into update semantics. The information state that is shared between speaker and hearer must meet the functional demands arising from the various roles that it has to play.

10.7 Conclusion and prospects

Probably the most important conclusion about update semantics is that it is a central format for any theory of interpretation. Even if one holds that the basic mechanisms of pragmatics are not related to updating, the predictions of any new proposal can only be tested if they are brought into the format of update semantics, since only this will give the predictions of the new proposal.

Second, update semantics seems central to accounts of pronouns, presupposition and speech acts, to the extent that alternative treatments tend to be unnatural and problematic. Whatever is interesting about E-type analyses of pronouns can be easily captured in an update semantics. The view that pronouns refer to an object of their local context is easier and more natural and much closer to the historical origin: deictic pronouns. Presupposition requires a distinction between old and new, a distinction which is the architectural backbone of update semantics. In other frameworks (cf. Schlenker Reference Schlenker2008) this distinction has to be carefully reconstructed. While non-controversial accounts of accommodation and partial resolution are not currently available, the formalisation of relevance on which it seems to depend is the focus of study in recent update semantics, since it is a central aspect of pragmatic updating. Relevance seems much further away in classical semantics.

Third, there is another side to our problematic attempts to separate semantics from pragmatics. In the classical picture, the historical shifts from a classical notion such as ability to a pragmatic notion such as epistemic modality would necessarily be something of a saltus: one moment there is a truth criterion, the next moment there is only an assertability condition. This does not sit well with an evolutionary account of these changes in which certain numerical parameters would slowly change. Update semantics offers continuity for changes of this kind and would predict in general the possibility of slow and continuous shifts in the meaning of lexical items that can move from descriptive to more and more pragmatic.

Finally, there is a philosophical point akin to the eliminability of a representational level for semantics in Montague's conception of natural language semantics. Update semantics is about changes to information states and these are sufficiently accessible to speakers and hearers to allow for experimental validation. This would be a level of representation that is as robust as the ability to communicate nonsense words through speech and hearing: performance can be tested. This does not apply to any of the intervening levels in standard accounts of language such as syntactic trees, logical forms of various flavours, proto-logical forms without resolution of pronouns, representations of ambiguous words. These levels have been assumed in accounts of grammar or of the interpretation or of the formulation and speaking process but lack any direct evidence of their own. Update semanticists are not committed to any of these intermediate representations, with the exception of phonetic representation, though – of course – they face the burden of explaining how the assumed processes might work just as much as anybody else and may have to adopt some of these levels in their accounts.

Current update semantics falls short of modelling linguistic agents in various ways: they have information (including anything that can be modelled as information, such as perception and desires), defaults, obligations and questions. There is no good reason, it seems, why they should not have other aspects as well, such as emotions, goals, intentions and decisions. It can be surmised that the future update semantics will not be about information states anymore, but about agents that are updated by linguistic communication (or about all the aspects of an agent that can be changed by communication).

Another element that was mentioned above is the need for more probabilities in these models (this is also part and parcel of a shift from information states to agents). Current update-semantical pragmatics is not able to weigh different resolutions of the same pronoun against each other, is not able to decide between accommodating and partly resolving a presupposition, and certainly not capable of the subtle weighing of the literal meaning of an utterance against an ironic interpretation. This last task is difficult. It would be wrong just to proceed with a probabilistic reinterpretation of the bivalent notion. In our perception and in linguistic communication, there is perhaps always an element of doubt, but on the whole our intuitions place us firmly in one single environment.

11.1 Discourse and normativity

It is not too controversial to say that discourse may interact with normative relationships among people, relationships such as obligations or entitlements. Some speech acts may produce novel normative links: if you promise your friend to return the money he lent you, then your speech act institutes an obligation on your part, the obligation to return the money. Some speech acts may presuppose specific already-extant normative relationships: for example, to speak at the banquet of a scientific conference presupposes some entitlement yielded by the speaker's status – such as his being the head of the organizing committee or perhaps the dean of the faculty backing the conference.¹

In some cases, a speech act may both presuppose and institute normative relationships. A typical case in point is ordering (commanding): this purports to institute an obligation on the part of its addressee; but it succeeds in instituting it only if it meets the condition that the actor's position is in a relevant sense superior to that of the addressee. Moreover, it would seem that ordering can be completely characterized in terms of the changes of the normative links it brings about: ordering, we can say, is simply the act which, when carried out by an entitled actor, creates a specific obligation on the part of the addressee. However, ordering is not usually thought of as a particularly typical speech act; language, so the usual story goes, is more a matter of something like “encoding and decoding information” or “communicating ideas and feelings”, while giving orders, or other ways of building on or establishing normative links, is little more than a by-product of this.

In this chapter I want to explore the possibility that normativity is far more crucial to language than this. An idea flickering in the theories of several twentieth-century philosophers of language (and seen earlier in Immanuel Kant) is that a certain kind of normativity is constitutive of our distinctively human mind (aka reason), founding our concepts and infiltrating the semantics of our language. If this is true, then normativity is not only an accidental element of some of our speech acts, but rather their essential ingredient. Here we want to expose the motivations supporting this view and search out its consequences.

In section 11.2, we will reconsider the traditional picture of communication as essentially a matter of transferring information and the related picture of language as a collection of representations (of ideas or things of the outer world); and we will consider, in section 11.3, alternative pictures. This will not only result in giving pragmatics pride of place over semantics in explaining the nature of language, but, moreover, endorsing a peculiar version of pragmatics, which we will call pragmatist pragmatics and which will take words to be first and foremost means of achieving practical ends. However, this will be only an intermediate station before our ultimate terminus: inspection of normative versions of pragmatics. We will reach it in section 11.4 after we reject the possibility of erecting the pragmatist picture of language on the concept of disposition and thus will be driven to what we will call the normative turn. In sections 11.5 and 11.6 we will then consider the consequences of this turn.

To foreshadow this approach to language and discourse, let us take as a basic thesis that meaningfulness is not a (naturalistic) property of a type of sound or an inscription, but rather a propriety: saying that an expression means thus and so is saying that it is correct to use it thus and so, that it is governed by a certain set of rules. The mechanism is supposedly similar to that animating games and sports: saying a piece of rubber is an ice-hockey puck is not ascribing the piece a peculiar (naturalistic) property, but citing its role vis-à-vis the rules of ice hockey. Hence from this viewpoint, discourse is like a kind of game: it is governed by rules (though in contrast to ice-hockey, not necessarily explicit rules) and meaningfulness is the effect of the rules.

It is crucial to realize that rules, as we understand the term here, are something different from what is studied in Chomskyan linguistics. The latter are certain principles implemented in human brains and thus inevitably governing our linguistic behavior; whereas rules in the sense adopted here guide us always “evitably” – it is a hallmark of the rules in our sense that they can be, as a matter of principle, violated. A rule construed thus is a social fact, consisting in the collective awareness that something ought to be thus and so, manifested in the corresponding behavioral regularities (not only the more or less regular compliance with the rules, but a more or less regular prosecution of deviations, etc.)

11.2 Transferring information?

Consider the received wisdom that language is a matter of transferring information. There seems nothing particularly normative about such transfers (aside from the trivialities that it can be exercised, as any other kind of act, correctly or incorrectly, well or badly, appropriately or inappropriately, etc.). What is going on is the moving of something from one head to another – a perfectly naturalistic enterprise.

But is this truly so? Consider the act of “transferring information” accompanying the act of assertion. What happens when I assert something, tell something to somebody, or inform somebody about something? I emit a certain sound (let us disregard the written form of language, for simplicity's sake), and it reaches my audience's ear. What can be effected by a mere sound? Seeing it as merely a sound, rather than an expression, we can imagine that its capabilities might include being able to scare the audience or attract its attention; but using the specific kinds of sounds that constitute language enables us to do much more complicated and very specific things: for example, we might cause somebody at a distant place to open a particular door, go into a room and take something from a particular drawer. How do we achieve this?

There is the temptation to have all the explanatory work done by the terms information and, indeed, meaning. The sounds that constitute language differ from other kinds of sounds, and acquire their almost miraculous abilities, because they have been furnished with meanings. (How? Perhaps – we pull another rabbit out of the hat – by means of a convention!) This grants them the ability to transfer information from one head to another, and as potential pieces of information constitute an incredibly large and fine-grained spectrum, we can achieve, by sending them to our audience, a very large and differentiated spectrum of reactions.

However, using such concepts as information and meaning as unexplained explainers begs the question. Almost everybody would agree that the talk about transferring information by means of words is merely metaphorical, that there are no such things as actual units of information literally hanging on the words that flow from my mouth to your ear.² But what is it a metaphor of? An immediate answer might be that our brains have the ability of dealing with language in the sense that they discern a vast number of sounds and somehow recognize them as “codes” encoding something. Hence the speaker uses the ability of his brain to “encode” information and the hearer uses it to “decode” it. Thus, nothing must literally hang on the words, it “hangs” on them merely metaphorically in that they encode it.

But in what precisely does the encoding/decoding ability of our brains consist? Does the brain contain some huge “code table,” which allows it to translate information to and from the linguistic codes? This seems to be precisely the idea put forward by Chomsky, who urges that language is a huge system of pairings of sounds and meanings.³ But how do such “code tables” materialize in our brains?

Clearly, this must happen during the process of learning the language in question. (It cannot be innate, for the codes, the words, differ from society to society.) A forthright picture of how this happens was described by Augustine: we are shown things and told their names, thereby building the table.⁴ This has developed into what can be called a semiotic or a representationalist picture of language, according to which meaningfulness consists, first and foremost, in standing for something. But is this picture plausible? Two problems immediately surface: First, although we can be shown things or people, and perhaps also, less straightforwardly, we can be introduced to qualities such as redness or manhood by being shown red things or male persons, there still remains a plethora of words where it is utterly unclear what we can be shown (think, for example, of adverbs, such as rapidly or always). Second, to speak a language is to have use of an unlimited number of expressions, so inevitably no amount of pointing can furnish them all with their meanings.

These problems are usually not considered insurmountable. As for the first, it is assumed that only some of the words of our language are learned by ostension and thus come to function as proxies for things, whereas the rest of them constitute some kind of auxiliary scaffolding for employing the “core words” (see, e.g., Weinreich, Reference Weinreich, Householder and Saporta1962: 36). As for the second, learning a language is seen to consist of learning the meanings of words plus mastering ways of composing meanings of complex expressions out of those of simpler ones.

Note, however, that this essentially compromises the picture of our semantic competence as a matter of the possession of a code table. As for the first problem, the remaining worry concerns what it is that is stored in our brains along with words like rapidly or always. As for the second, we can certainly say that what effects the encoding/decoding should be seen more as an algorithm than as a table, which would account for the potential infinity of meaning–sound pairs; but what about the infinity of meanings themselves? Certainly if we are to encode a potentially infinite number of entities, we need to have the entities (if only merely potentially). But how do we generate all the meanings expressible in a language? Of course, we can generate them via generating all the expressions expressing them; but if they are to be encoded by the expressions, then it seems they should be available before the encoding happens. And if we have an unlimited number of meanings, what sense does it make to think that we have them as being assembled in our brain for a potential encoding?

We may try to meet this challenge by claiming that the meanings are something we are born with – that nature and our genes endow us with a “language of thought” that does not consist of expressions codifying meanings, but rather of expressions that directly are meanings. This seems to be the strategy of Fodor (Reference Fodor1975; Reference Fodor2008) and his followers. However, if we see a language as a collection of sound- or inscription-types that become meaningful via their engagement with our complicated discursive practices, then such a detachment of meanings from expressions is basically problematic: once again it leads to unexplained explainers that stand in essential need of explanation themselves.

11.3 The pragmatist turn

At this point, the picture of communication as principally a matter of exchanging encoded information loses its initial plausibility, encouraging us to consider some other approach. And as many philosophers and linguists have recommended, a useful alternative might be to see language not in terms of representing things and encoding/decoding information, but rather in terms of practical ends to which linguistic items can serve as means, to see expressions as tools rather than as codes. This insight is characteristic of philosophical pragmatism (see Haack and Lane, Reference Haack and Lane2006), but it has found its way into various other philosophical and linguistic conceptions of language in the later twentieth century. It is central to the neopragmatist theories of Quine and Davidson, it animates the later Wittgenstein's theory of language games and it is partly also present within the speech act theory initiated by Austin and Grice. I will call this view of language and communication pragmatist pragmatics. From this perspective, questions of meaning and information, of course, are radically altered.

Consider a toolbox (the metaphor for language favored by the later Wittgenstein⁵). I may learn various ways of using the tools the toolbox contains; and not only each of them alone, but also in combination: the hammer and a nail, the screwdriver and a screw, a bolt and a nut…I may accomplish various useful things. The more skillful I am with the tools, the more practical tasks they can help me solve. Moreover, they render it possible for me to do things I would never have even considered before: to build and mend things I would previously have been unable to imagine. Thus, although sometimes I use tools to cope with tasks that would have faced me independently of whether I had a toolbox or not, very often I use them to do things that I would not come to think about were it not for my experience with the toolbox – it is the skill of using the tools that makes many tasks that can be solved by its means visible in the first place. For this reason it would be absurd to think of the tools and their combinations as responses to tasks given beforehand.

Switching to this pragmatist view of language also prompts us to re-evaluate the traditional view of the semantics/pragmatics boundary. Recall that the traditional definition of the boundary between semantics and pragmatics, as given by Morris (Reference Morris1938) and Carnap (Reference Carnap1942),⁶ was conceived within the representationalist framework. Semantics was considered to address questions central to the framework, namely what things our words stand for, while pragmatics was relegated to the sphere of peripheral, residual questions of how we use words – how the words endowed with meanings get also peripherally endowed with our habits, moods, or fancies.

Given the pragmatist turn that leads us to see language more as the vehicle of an activity, semantics is effectively swallowed up by pragmatics – everything is a matter of how we use language. Of course, we can still have a semantics/pragmatics boundary, but now pragmatics will not be an unessential appendage of semantics, but rather semantics will be a slightly artificially demarcated part of pragmatics (such as, for example, the part which deals with truth-conditions, as Stalnaker (Reference Stalnaker1970) suggests).

Hence, the pragmatist turn may help us overcome the obstacles generated by the earlier idea that what we principally do with language is transfer information, which led to the consequent code conception of language.⁷ From the pragmatist viewpoint, language is not a mere instrument of dealing with the extralinguistic world: while in some cases it may help us cope with problems we would face independently of us having or not having a language, more often than not we will use it for tasks which only came into being with the birth of language – discussing theoretical questions, reciting poetry, buying a book, and so forth. Hence it seems that the tasks for which expressions are fitted co-evolve with language. For this reason it seems to me to be misleading to imagine expressions as codes of something given independently of them.

Note that, although we claimed that this pragmatist approach to pragmatics is not alien to the speech act theory of Austin and Grice (see their perlocution dimension of the speech act), ultimately it is an alternative to the specifically Gricean approach to pragmatics based on the concept of intention, which dominates the current pragmatics scene. Just as this avoids taking either the concept of information or that of meaning as an unexplained explainer, so it also avoids taking intention as such.⁸ In this way it is more thoroughly naturalistic.

Hence the impasse into which we were brought by the code conception might be overcome if we take the pragmatist turn; but from the opening comments of this chapter it follows that we want to consider one more turn, namely a normative one. Why is this? Why should we not rest content with the pragmatist turn?

11.4 What ties an observation report to what it reports?

Consider the sentence This is a dog. What does the fact that it means what it does in English consist in? It would seem that whatever else might come into its meaning, what is essential is that English speakers use this sentence when confronted with a dog, and not when confronted with, say, a horse.

However, is this true? For a start, we can think of cases where competent speakers might utter This is a dog when confronted with a horse – as a result of bad sight, of jocularity, irony, or poetic inspiration, etc. But perhaps such cases may be dismissed as statistically insignificant, and more substantial is the observation that the majority of English speakers when confronted with a dog would be extremely unlikely to actually utter This is a dog. In fact, I suspect the number of positive cases may well be statistically insignificant. (As Chomsky (Reference Chomsky, Davidson and Hintikka1975) conjectured, the statistical probability of uttering almost any specific expression in a given situation will not be significantly higher than zero.)

Hence we are back at the conundrum of what establishes the connection between This is a dog and dogs. The usual rabbit that philosophers and linguists pull out of their hats here is the concept of disposition. This is a dog means that this is a dog (and not, say, a horse) because speakers are disposed to utter it when confronted with a dog. This disposition sometimes provokes the overt utterance, but more often than not it remains covert.

But this rabbit, I feel, is just a trick to delay clarification. What does it mean that a speaker has a covert disposition to utter This is a dog? It amounts to the counterfactual claim that the person would utter it were it not for some hindrance. What would substantiate such a claim? Certainly, it would be well substantiated were we able to identify some mechanism in a person's brain which tends to lead to the utterance, which may be obstructed by certain factors, and which may be shown to be so obstructed in the case in point. But at present we are party to no such mechanisms.

Alternatively, we can interpret the claim as not a claim about an inner mechanism, but rather about empirical regularities. We may report that This is a dog is always uttered in the presence of a dog unless certain “hindering” factors occur. But for such a claim to have empirical content (to be, for example, testable) we would have to be able to specify the relevant factors. Otherwise the claim would be empty: any evidence would be compatible with it. (We would never be able to object that the claim is at variance with the facts: cases of speakers reacting to dogs with This is a dog would be in order, and cases of those not reacting in this way could always be explained away with reference to unspecified hindering factors.)

Are we able to give an exhaustive catalog of things or events that stop one exclaiming This is a dog in the presence of a dog? Can we say, for example, that principal factors are unwillingness to talk, preoccupation with other matters, or not noticing the dog in question? Hardly; we can clearly think of any number of others. Hence it seems that invoking the concept of disposition here is a mere illusion of explanation. Does this mean that there is, despite appearances, no intelligible connection between This is a dog and dogs after all? I do not think so; but I think that we tend to look at this connection in the wrong way. What I think is the case is that the connection is normative rather than causal. This is to say that the link between the occurrence and the utterance is not a matter of any causal mechanism connecting the two, but rather of the fact that to utter This is a dog when a dog is in focus is correct.⁹

However, without further ado, this would be at most a gesture towards an explanation (if not merely another trick). What is a non-causal, normative connection? Should one imagine some kind of supernatural fiber leading from dogs (and other potential objects of reference) to the minds of speakers? It is clear that unless we give a viable clarification, this alleged explanation would not really be useful. Hence it is the clarification to which we now turn our attention.

11.5 Correctness

To foreshadow where we expect our normative turn to lead us to, let us consider an activity seemingly very different from using language: the game of football. What is important for us is that playing football amounts to enjoying a spectrum of actions that are not available for us outside of its framework: get into an offside position, foul an opponent, or (joy of joys!) score a goal. How do such actions become available for me? Obviously because I, as well as my team-mates and our opponents, submit to the rules of football – it is the rules, and in particular the collective submission to them, that open up the space for the new kind of actions. And the thesis I want to put forward and discuss is that linguistic actions, actions that we tend to describe as cases of meaningful talk, transfer of information, or stating facts (or whatever else one can do with meaningful language), arise analogously: namely as a result of our collective recognition of the rules of language.

This recognition means nothing over and above the fact that we take certain linguistic utterances for correct, and others for incorrect. (This may be the case on several disparate levels – an utterance may be, e.g., grammatically correct while being incorrect as an assertion.) We know that it is correct to say This is a dog when pointing at a dog (whereas that this is incorrect when pointing at a bus); we know that it is incorrect to dissent from This is an animal while assenting to This is a dog; and we know that it is correct to raise one's hand when assenting to Will you raise your hand? Thus, a rule in this sense is a matter of a collective awareness, of an awareness that something is correct and something else is incorrect, leading to the appropriate behavior (praising the correct and trying to do away with the incorrect).

Let us start from the question of how we recognize the presence of a normative link of the kind discussed above. (After all, as Quine reminded us, we all learn to speak by means of observing our elders and peers and as what we can perceive is exclusively behavior – hence, we can say, there cannot be anything in meaning that was not in behavior before.) How does a language-learner recognize that This is a dog is “normatively linked” to dogs (rather than horses) and so grasp the meaning of the sentence (and especially of the word dog)?

When learning a language we may witness a demonstration of using This is a dog as accompanying pointing at a dog. However, though this can indicate the existence of a link, it cannot intimate the nature of the link, let alone that it is a normative link. Given our genetic tendency to imitate, we may come to utter This is a dog when pointing at a dog ourselves; but nothing apparently stops us from uttering it when pointing at things other than dogs – say, all furry things, or even at anything whatsoever.

The decisive step here is that we must learn that using it when pointing at something not a dog is incorrect. How do we learn this? By experiencing some kind of “social friction,” by facing “corrective reactions” of other speakers to such misuse (our own or somebody else's). What constitutes such a “corrective reaction”? It may be anything from a true punishment to a mild dissatisfaction or puzzlement. Anyway, one of the principal “social skills” anybody must master to be an integral part of his or her society is to recognize which kinds of behavior count as “corrective” (and being able to feel this kind of “friction” appears to be one of the most essential social skills).

Hence the original encounter with the normativity of meaning is in this “must not”: We do not learn what we should do, but rather what we should not do. This is important, for this may help clarify one of the most frequent misunderstandings regarding the normativity of meaning: the normativity does not rid us of our freedom in using language and hence does not contradict the obvious fact that using language is a spontaneous activity – it merely restricts the freedom, still leaving a vast number of possibilities. (We will return to this later.)

This also indicates that the normativity of meaning is somehow carried by the corrective, or as we may say more generally, normative attitudes of speakers to other speakers’ pronouncements. And this may bring us to a suspicion that we have not done away with the concept of disposition we deemed suspicious above, but merely shifted it one level up: for do we not need the concept for the characterization of the concept of normative attitude? Can we say that normative attitudes consist in a “corrective behavior” – or do we have to say that they consist in the disposition to corrective behavior? After all, not everybody who uses language incorrectly faces correction by others!

It is true that though the occurrence of a generic “corrective behavior” is more easily predicted than individual pronouncements, it is not, of course, the case that we cannot say that a pronouncement is wrong only if some such behavior occurs. But we may avoid the concept of disposition in the same way as before: instead of saying that an utterance is incorrect either if it faces corrective behavior, or if others would have the disposition to the “corrective behavior” towards it, we can say that it is incorrect if the corrective behavior towards it would be correct.

However, this definitely looks like a trick, and in this case a particularly naive one. (Am I criticizing my colleagues for pulling rabbits out of their hats only to end up pulling one out myself?) Is it not merely shifting the whole problem to a third level and thus possibly starting an infinite regress? The attempt to reduce correctness to behavior or to dispositions to behavior would indeed lead us to an infinite regress; but my answer is that this is to be taken as indicating that the reduction is impossible.

But is this, then, not a mere reductio ad absurdum of the whole idea of the “normativist turn”? Am I suggesting that behind (or above) human behavior (and the whole network of causal relationships, in which it is embedded) there looms some different, supernatural stratum of reality where we can encounter correctnesses? Of course not, though admittedly it may sometimes be useful to invoke this picture as a metaphor. The point is rather that there are no such things as correctnesses. Why they seem to be here is that we seem to state facts about them; but what looks like declarative statements about such correctnesses – I will call the statements normatives – are not always really declaratives. Hence what is behind the untranslatability of the normative idiom into the declarative one (and hence the reduction of “norms” to “facts”) is not the incommensurability of a “normative” and a “factual” stratum of reality, but simply the more mundane fact that many of the normatives are not genuine declarative sentences at all, but rather belong to a different species of speech act. They are, as Wilfrid Sellars (Reference Sellars1962: 44) put it, “fraught with ought.”

Return to the case of football; and consider the statement Hands should not touch the ball. This is a normative. There are two ways of employing a statement of this kind. First, one can state the fact that this kind of rule is in force in some community. This is, as it were, an “outsider” statement; a statement made by a disengaged observer describing the practices of the community in question. Second, one can state this as an “insider”, which does not amount to (or only to) stating a fact, but also to upholding the rule, urging its propriety or at least confirming its legitimacy. And true normatives are normatives posed precisely from this perspective. It follows that to say that rules exist is strictly speaking a metaphor: they do not exist, of course, in the way rocks, trees, or dolphins do. To say that a rule exists is to take some true normatives people use for ordinary declaratives. It seems to be our human way to do this; but we should be aware of the fact that this is a sense of existence different from the one in which we use the word when we talk about the existence of spatiotemporal particulars and their constellations.

And here we come to the mystery of how correctnesses, or proprieties, can exist relatively independently of our attitudes, and yet without being situated in some independent stratum of reality. The point is that any verdict we reach regarding correctness is at best tentative, it belonging to the nature of the concept that the verdict is considered as always amendable by our successors. They can find out that what we hold for correct is in fact not correct – but unlike in the case of blue, indivisible, or animate, we do not always see such cases as errors in the application of the concept, but rather as discovering the true nature of the concept of correctness.

Perhaps we can say that a rule or a norm is always an unfinished and open project (see Gauker, Reference Gauker2007). Usually it grows out of our current practices, is continuous with them, sometimes to the extent that we can understand its statement as a description of the practices as they, as a matter of fact, are; but the statement of a norm is also usually its prolongation, an urge “And we should go on like this!” Hence a rule is never a completed whole, it is always open to the future, not only to prolongations, but also to modifications and amendments. It is like a track that we must go on extending to ever new horizons.

11.6 Norms and convention

I have mentioned the concept of convention as a “rabbit” which is usually pulled out of the semanticist's hat when we need to say what ties an observation report to the observation it reports. But at this point someone might want to wonder whether the theory I have been developing in terms of the concepts of norms and normativity is not about what semanticists have long been referring to by means of the term “convention.” Well, in one sense it is. However, this is due to the fact that the term is largely ambiguous; and some of its senses do refer to normative phenomena.

As Rescorla (Reference Rescorla2007) points out in his overview article, “in everyday usage, ‘convention’ has various meanings, as suggested by the following list: Republican Party Convention; Geneva Convention; terminological conventions; conventional wisdom; flouting societal convention; conventional medicine; conventional weapons; conventions of the horror genre.” He offers a useful list of possibilities that the “conventional” may be contrasted with: “the natural; the mind-independent; the objective; the universal; the factual; and the truth-evaluable.” This indicates that to deal with the concept of convention, we need to start with disambiguation. I think there are at least three senses of the term relevant for the theory of language. In the first sense, “convention” is something like a habit; in this sense, “the conventional” is, in the words of Goodman (Reference Goodman and Krausz1989: 80), “the ordinary, the usual, the traditional, the orthodox as against the novel, the deviant, the unexpected, the heterodox”. In the second sense, convention is what has been established by man and has not been part of nature all along; in this sense, “the conventional” is, using Goodman's words again, “the artificial, the invented, the optional, as against the natural, the fundamental, the mandatory” (ibid.). In the third sense, “the conventional” is what has been explicitly agreed upon.

I think that the attractiveness of the term “convention” stems largely from the conflation of the three senses. When we encounter a problem concerning the way language hooks onto the world, we often invoke the term in the second sense. Surely the sentence – that is, the sound-type – “This is a dog” is not naturally connected with dogs, hence the relation is conventional. So far so good; but aside from giving the relationship a label nothing has been explained yet. However, the next step often is that we do not really need to explain anything, for the concept of convention is more or less self-explanatory. However, whereas this might be true for “convention” in the third, or perhaps also in the first, sense, it is definitely not true for “convention” in the second sense, for this sense remains blatantly neutral with respect to how conventions come into being and what their nature is. And it is clear that if we use the term to account for how language hooks onto the world, then it cannot be generally convention in the third sense of the word: language cannot be based on this kind of convention, for language is presupposed by this kind of convention.¹⁰

Could it be “convention” in the first sense, convention as a habit, that holds together the sentence and dogs? Habits do not seem to be too unperspicuous; so maybe it is “convention” in this sense that those who use the term as the universal unexplained explainer use it. However, “conventions” in this sense clearly do not overlap with our normative account of language and discourse: habits as such do not have any normative dimension. If my habit is to go for a walk every evening, it may be surprising that I do not go out today, but it is in no sense wrong.

Habits, to be sure, may evolve into norms. Once people start to take the habitual as not merely what usually happens, but rather what should happen, there emerges a norm – or, you may want to say, the habit becomes a norm. But here the latter step is crucial. For consider chess or football, which we use as our models of our discursive practices. People may acquire the habit not to take the ball into their hands; but the game cannot really get off the ground until this starts to be felt as what should not be done and until those who keep doing this start to be penalised. Hence the habitual substrate is surely not everything that makes up norms.

When you look into the writings of Austin (Reference Austin1961; Reference Austin1962), Grice (Reference Grice1989), Searle (Reference Searle1969) or other speech act theorists, what you find there is that the terms “convention” and “conventional” are among the most frequently used words, but despite this none of the authors pay much explicit attention to the question of what conventions are. Of course there are some hints: Austin (Reference Austin1961: 64), for example, when mentioning a “semantic convention” adds a parenthesis “(implicit, of course)” which indicates that what he has in mind is not “convention” in our third sense of the word. Elsewhere he talks of an assertion justified “not merely by convention, nor merely by nature,” which in turn indicates that he uses “conventional” in opposition to “natural”; hence that he uses it in the second of our three senses.

There is a flagrant disproportion between the huge explanatory work the concept of convention is supposed to do in such writings and the absence – or near absence – of its own explanation. This can be, to a great extent, justified by the fact that these authors use the terms “convention” and “conventional” in a sense that they do not see as being in need of explanation, hence, I would think, mostly in our second sense. But sooner or later, then, we must face the question of how this kind of conventionality comes into being.

The first person who realized that this is a serious problem was David Lewis (Reference Lewis1969). He clearly realized that the concepts of convention and conventionality that occur so frequently in writings about language cannot be generally construed as explicit agreements, and set himself the task of showing how “tacit conventions” can come into being. His solution of this problem is based on two assumptions: conventions come into being to solve coordination problems, and the solution of such problems can evolve spontaneously along the lines envisaged by game theory.

I think that despite the fact that Lewis laudably brought the nature and emergence of conventions into the focus of attention and showed how some tacit conventions may emerge spontaneously (thus breaking from the vicious circle into which we would fall if we wanted to base language on explicit conventions), his approach is not general enough. In particular I think that by no means all the norms language is based on can be seen as deriving from conventions solving coordination problems – at least not unless we generalize the concept of coordination problem to the extent that it will no longer be explainable in Lewis's simple game-theoretical terms.

Consider chess or football again. Can we say that their rules are a matter of conventions? Obviously, we can; in fact it would seem that the rules of games or sports are prototypes of what we would call conventional. (As we saw, there might be some terminological disputes over whether we should say that the rules themselves are conventional, or whether they evolved from conventions, but this is not important now.) But can we see them as solutions to coordination problems? This does not seem to be too plausible.

The fact is that, as I have argued elsewhere (see Peregrin, 2011 from the game-theoretical viewpoint the basic kind of rules relevant for language is more akin to those governing games of the kind of the Prisoner's Dilemma (solving genuine conflicts) than those of coordination.¹¹ Thus I do not think that the sector of game theory Lewis took into account is general enough to account for the problems standardly addressed under the heading of normativity.

On the other hand, I repeat that I think that it is possible to see a large overlap between “normativity” and “conventionality.” However, one thing to keep in mind is that this is due to the catholic nature of the concept of conventionality. Moreover, if we adhere to the sense of conventionality for which this overlap obtains, this sense of convention is not sufficiently explained and the present considerations may be seen as a contribution to its explanation.¹²

11.7 Normative speech acts theory?

The ideas exposed in the previous sections have led to the project of normative pragmatics that was first explicitly formulated – on a very general level – by Brandom (Reference Brandom1994: Chapter 1).¹³ Brandom's tenet is that

In this sense semantics must be “answerable to pragmatics,” namely to normative pragmatics.

When Searle (Reference Searle1969), in his classic book about speech acts, elaborated on the Gricean and Austinian speech act theory, his major example, discussed in the third chapter of the book, was the act of promising. His incipient characterization of this act reads as follows (Searle Reference Searle1969: 57–61):

1. 1. Normal input and output conditions obtain.

2. 2. S expresses the proposition that p in the utterance of T.

3. 3. In expressing that p, S predicates a future act A of S.

4. 4. H would prefer S's doing A to his not doing A, and S believes H would prefer his doing A to his not doing A.

5. 5. It is not obvious to both S and H that S will do A in the normal course of events.

6. 6. S intends to do A.

7. 7. S intends that the utterance of T will place him under an obligation to do A.

8. 8. S intends (i-I) to produce in H the knowledge (K) that the utterance of T is to count as placing S under an obligation to do A. S intends to produce K by means of the recognition of i-I, and he intends i-I to be recognised in virtue of (by means of) H's knowledge of the meaning of T.

9. 9. The semantical rules of the dialect spoken by S and H are such that T is correctly and sincerely uttered if and only if conditions 1–8 obtain.

This characterization involves a normative notion, namely the notion of obligation (point 7). Then Searle points out that his characterization leaves no room for insincere promises; so he then proposes to replace condition 6 with 6a:

1. 6a. S intends that the utterance of T will make him responsible for intending to do A.

Thus he introduces the normative notion of responsibility. (As the concepts of obligation and responsibility may be interdefinable, maybe it is not another normative notion, but merely the reiteration of the original one.)

However, is the list, and especially the role of the normative notions in it, formulated adequately? Does one, making a promise, intend to be placed under an obligation? Of course, as we assume that a typical promise is an intentional act, we would tend to consent; but is this inevitable? Suppose that I agree, say in a written form, to return somebody some money he lends me. Then suppose that I do not do so and when my creditor sues me, I tell the court that I did not really intend to have this obligation, hence that my act was not really a promise. (And let us suppose this is all true.) Am I likely to win the trial on the basis of proving that I have not promised anything? (It is hard to imagine how my declaration about my intention – if we understand the term “intention” in the sense of Grice and Searle as a basically internal act – could be challenged, for I alone have direct access to it.)

In view of this fact we can consider replacing 7 with 7*:

1. 7*. The utterance of T will place S under an obligation to do A.

Then, it would seem, some of the other entries on the list may become superfluous. Consider, for that matter, 6 or 6a. Suppose that someone promises me to give me some money, but in fact does not intend to be responsible for it. Does it mean that what he does is not promising?

In fact, it would seem that the rest of the list might also dwindle (if not vanish completely). True, the inflation of the point 7 to 7* results from using a more robust concept of obligation than the one used by Searle: insofar as I understand him, what he has in mind is obligation as a matter of psychology, whereas the one suggested by me is obligation in the sociologico-institutional sense. Hence inflating the normative dimension of the act also involves moving the act “out of the head,” into the open. This means that some or all of the differences between Searle's account and the proposal made here may be terminological.

This does not mean, though, that the difference between them is insubstantial. The normative twist given to speech act theory involves a significant reinterpretation of the whole enterprise – instead of having psychological states as its direct concern, it now concentrates on normative statuses. Unlike its traditional version, it is wholly broken away from psychology of communication, which is the result of the conviction that language, and especially meaning, is more an institution than a psychological reality, and that psychology of communication is no more directly connected to communication than psychology of chess to chess.¹⁴

Moreover, a normative speech act theory must make it plausible that not only speech acts like promising or ordering, which it can handle relatively easily, but also such speech acts as asserting can be characterized in normative terms. This is a much harder nut (see, e.g., Pagin, Reference Pagin2004, for a skeptical viewpoint). The idea here is that making an assertion is nothing over and above assuming the commitment to provide a specific justification; and to entitle anybody else to reassert the sentence in question while deferring its justification to you.

Kukla and Lance (2009) proposed a normative version of speech act theory, according to which speech acts are generally characterized by the normative conditions of their appropriateness and the normative outcomes of their occurrence. Thus, for example, ordering is appropriate if the orderer is in some sense superior to the orderee; and in such a case its normal felicitous outcome is the commitment on the part of the orderee to do what he or she was ordered. Using this unusual key to the classification of speech acts yields an unusual classification: for example, the usual category of assertions divides into declaratives and what Kukla and Lance call observatives. The two acts differ in their conditions. Declaratives, assertions like There is a pig in the yard, are indiscriminatingly available to anybody; whereas only certain people are entitled to observatives, assertions like Lo, a pig!

Another peculiar kind of speech act which has surfaced after Kukla and Lance put our linguistic practices under the normative lens is what they called vocatives, acts of the kind of Hey, you! While observatives are characterized by having specific normative conditions (not everybody is entitled to make them), but general normative outcomes (they entitle everybody to make use of them), with vocatives it is the other way around: they have general normative outcomes (everybody is entitled to them), but their outcome is specific (they entitle a specific individual to enter the ongoing language game). Kukla and Lance claim that the identification of such speech acts, which do not surface in traditional speech act theory, significantly advances our understanding of language.

All of this, of course, presupposes that we accept that there is no meaningfulness without a normative dimension. This is, recall, the result of taking the Wittgensteinian picture of our discursive practices as a cluster of language games at face value, not only in the sense that the practices are heterogenous, but also in the sense that they are essentially underlain by rules. Given this, any speech act is individuated by the way in which it fits into the normative scaffolding that constitutes the space which provides the necessary substrate for such speech acts. And given this, in turn, our perspective on discursive practices shifts significantly, and may illuminate aspects hardly discernible from the perspective of Austin and Grice.

11.8 Conclusion

The traditional approach to language was based on the assumption that we must first explain how a word means something (which was, in turn, taken as tantamount to explaining how it can stand for that something), and then we would be able to explain language as a product of the synergic effect of an assembly of meaningful words. The pragmatic turn in the twentieth century (especially in its pragmatist variety) inverted the perspective: we must explain, the credo has come to be, directly how language works, i.e. our linguistic practices, using the concept of meaning at most as an expedient of this enterprise.

What I have been exploring here is the possibility of this pragmatic turn being given a normative twist: of meanings being explained as roles vis-à-vis rules of language. Let us return to football. As I noted above, once you accept its rules, you can do things which you were not capable of doing before. Note that this does not mean: some things you were capable of doing before (like kicking a round thing through a gate-like thing) now receive new descriptions (“scoring a goal”). Scoring a goal is not reducible to kicking. I and my team-mates might do precisely the same movements we do now, but without being caught in the normative network constituted by the rules of football they would not be scoring goals and they would not have many of the effects they have now (like making us happy, making our opponents annoyed, bringing money to those who laid bets on us, while causing those who betted against us to lose their money, etc.). In short, the rules of football constitute a new spectrum of actions not available to us before. And likewise, the amazing spectrum of things we can do with words is created analogously – by means of the rules of language.

When Carnap and Morris presented their division of the theory of language into syntax, semantics, and pragmatics, they gave, in effect, pride of place to semantics (relegating syntax to the auxiliary role of honing the vehicles that only semantics discloses as carrying meanings; and relegating pragmatics to the marginal role of sidekick to semantics). Moreover, Carnap then reconstructed semantics as a mostly logico-mathematical, armchair enterprise: the semantic theories he presents in his Introduction to Semantics (1942) or Meaning and Necessity (1947) do not seem interconnected, in any significant way, with any empirical investigations of natural languages. I think this was an unhappy development (and it justified the revolt of the many theoreticians of language who subsequently made pragmatics, rather than semantics, the centerpoint of the study of language¹⁵), and I want to ensure that the normative turn discussed here should not lead to a similar consequence.

It is true that normativity seems to be a tool that only philosophers have in their philosophical toolbox. Linguistics, one is tempted to say, is a down-to-earth science, and science describes how things are, not how they should be – so what use for normativity is there? The present chapter has tried to offer an answer, an answer as down-to-earth as possible. Human activities, be it chess or football or some much more complex and socially important ones, are governed by rules – indeed they are constituted by the rule-governance. This is clearly nothing mysterious or at odds with science – it is simply an empirical fact. And here I want to suggest that insofar as this applies also to language (which presupposes seeing language as a social institution, rather than, say, a psychological reality), we may come to see that this enterprise has an important normative dimension and that to understand this dimension may be essential for understanding language and discourse.

12.1 Introduction

In this chapter I shall discuss only the lexicon of English, but the general principles seem to apply to many, if not all, other languages even though the minutiae do not. By “lexicon” I mean a rational model of the mental lexicon or dictionary. Although the way a lexicon is organized depends on what it is designed to do, it is minimally necessary for it to have formal (phonological and graphological), morphosyntactic (lexical and morphological categorization), and semantic specifications. Relations are networked such that formal specifications are (bi-directionally) directly linked to morphosyntactic specifications that are directly linked to semantic specifications – which, for the moment, subsumes pragmatic specifications. A lexicon must be accessible from three directions: form, morphosyntax, and meaning; none of which is intrinsically prior. Each of these three access points is, additionally, bi-directionally connected with an encyclopaedia. Haiman (1980: 331) claimed “Dictionaries are encyclopaedias” and certainly many desktop dictionaries contain extensive encyclopaedic information (e.g. Hanks Reference Hanks1979; Kernfeld Reference Kernfeld1994; Pearsall Reference Pearsall1998). The position taken here is that a lexicon is a bin for storing listemes,¹ language expressions whose meaning is (normally) not determinable from the meanings (if any) of their constituent forms and which, therefore, a language user must memorize as a combination of form, certain morphosyntactic properties, and meaning. An encyclopaedia is a structured database containing exhaustive information on many (perhaps all) branches of knowledge. It therefore seems more logical that the lexicon forms part of an encyclopaedia than vice versa, but the actual relationship does not significantly affect this chapter. I assume that encyclopaedic information is typically, if not uniquely, pragmatic.

A lexicon is a bin for storing listemes for use by language speakers in any and all contexts. This is not to deny that new listemes are occasionally created, but the coining of a new listeme is a rare event and the resources of a lexicon are normally adequate for all contexts that a speaker faces. Consequently the meanings of listemes are expected to be adapted by semantic extension or narrowing both concretely and figuratively by speakers in utilizing them and hearers in interpreting them. Such lexical adjustment can be illustrated by the various meanings of the related listemes cut in (1).

1. (1) cut grass, cut hair, cut steel, cut the thread, cut the cards, cut your losses, cut out the middle man, cut the ties, to cut and run, cut the cackle, cut a class, cut someone socially, be a cut above, she's all cut up by the breakdown in her marriage, be cut to the quick, cut through the obfuscation, cut my finger, cut the tyres, cut the cake, cut a disk, a railway cutting, cut through the back lane, cut a [fine] figure

Most, if not all, of these seem to derive from a basic notion of severing, interpreted in various ways according to what is severed and/or the manner of severing (this could even apply to cut a figure). Similarly, it is well known that a color term may extend to shades very far from the focal color (Berlin and Kay Reference Berlin and Kay1969; MacLaury Reference MacLaury1997) as selected from, say, the Munsell Color Array; we can attribute this to the elasticity that language needs to have in order that it can usefully be applied to the world around us. In certain domains and in certain formulaic expressions color terms are used of hues vastly distant from the focal color. Take the domain of human appearance: terms like white, black, yellow, and brown have all been used to characterize the skin pigmentation of people of different races, often dysphemistically. These color terms are descriptively appropriate not so much in relation to the focal colors as in relation to each other: a white person is typically paler than the others and a black person darker; a yellow person is typically yellower than the others. The peoples of southeast Asia and Austronesia are often referred to as brown, despite the fact that peoples labeled black are often of similar brown skin color. So brown, too, functions by contrast with white, black, and yellow in this domain. In the domain of oenology, red wine does have a (usually dark) red tinge but white wine is only white by virtue of being paler than red wine; white wine is normally pale yellow or pale green. Clearly what determines the meanings of these particular sets of color terms is their comparative function: by means of very rough approximation to the focal color, they distinguish within a semantic field between different species of the kind of entity denoted by the noun they modify.

Pragmatics within the lexicon is largely an addition to the semantic specifications; for instance, it is useful to identify the default meanings and connotations of listemes. Default meanings are those that are applied more frequently by more people and normally with greater certitude than any alternatives. Bauer (Reference Bauer1983: 196) proposed a category of “stylistic specifications” to distinguish between piss, piddle, and micturate, i.e. to reflect the kind of metalinguistic information found in traditional desk-top dictionary tags like ‘colloquial’, ‘slang’, ‘derogatory’, ‘medicine’, ‘zoology’; such metalinguistic information is more encyclopaedic than lexical. So too is etymological information. Pustejovsky (Reference Pustejovsky1995: 101) specifies book as a “physical object” that “holds” “information” created by someone who “write[s]” it and whose function is to be “read.” Certainly, there is a relation between book, write, and read that needs to be accounted for either in the semantic specification or pragmatically – Pustejovsky represents it in terms of a network and networks are also used in frame semantics (Fillmore Reference Fillmore1982; Reference Fillmore and Brown2006; Fillmore and Atkins Reference Fillmore, Atkins, Lehrer and Kittay1992; FrameNet at http://framenet.icsi.berkeley.edu) and by Vigliocco et al. (2009). Category terms like noun, verb, adjective, and feminine are part of the meta-language, not the object language; but they also appear in the lexicon as expressions in the object language and there needs to be a demonstrable relation from object language to meta-language (and vice versa). It would seem incontrovertible that encyclopaedic data is called upon to interpret non-literal expressions like Ella's being a tiger; likewise, to explain the extension of a proper name like Hoover to denote vacuum cleaners and vacuum cleaning or the formation of the verb bowdlerize from the proper name Bowdler. I assume that, because many proper names are shared by different name-bearers, there must be a stock of proper names located either partially or wholly in the lexicon, even if they are stored differently in the brain (see section 12.9). The production and interpretation of statements like those in (2)–(3) requires pragmatic input.

1. (2) Caspar Cazzo is no Pavarotti!

2. (3) Harry's boss is a bloody little Hitler!

(2) implies that Caspar is not a great singer; we infer this because Pavarotti's salient characteristic was that he was a great singer. (3) is abusive because of the encyclopaedic entry for the name Hitler that carries biographical details of a particular infamous name-bearer. Such comparisons draw on biodata that are appropriate in an encyclopaedia entry for the person who is the standard for comparison but not appropriate in a lexicon entry; the latter should identify the characteristics of the typical name-bearer, such as that Aristotle and Jim are normally names for males, but not (contra Frege Reference Frege1892) the biographical details of any particular name-bearer – any more than the dictionary entry for dog should be restricted to a whippet or poodle rather than the genus as a whole.

One of the earliest investigations of lexical pragmatics was McCawley (Reference McCawley and Cole1978); McCawley (correctly) argued that a listeme (such as pink or kill) and a semantically equivalent paraphrase (such as pale red or cause to die) are subject to different pragmatic conditions of appropriateness that give rise to different interpretations, which he thought could be captured by general conditions of cooperative behavior such as Grice's cooperative maxims. He did not tackle the question of whether pragmatics intrudes on lexical entries. Nor does Blutner (Reference Blutner1998; Reference Blutner, Horn and Ward2004; Reference Blutner and Cummings2009). Blutner discusses pragmatic compositionality, blocking (if a listeme already exists to express a meaning, do not construct another one without good reason to do so),² and pragmatic anomaly (recognized as early as Apollonius Dyscolus in Peri Suntaxeōs III.149; see Uhlig Reference Uhlig1883). The closest Blutner comes to pragmatics within the lexicon is discussing the interpretation of certain adjectives and institute-type nouns (Blutner Reference Blutner1998).

Carston (Reference Carston2002: Ch. 5), then Wilson and Carston (Reference Carston and Frápolli2007) discuss lexical narrowing (e.g. drink used for ‘alcoholic drink’), approximation (e.g. flat meaning ‘relatively flat’), and metaphorical extension (e.g. bulldozer used to mean ‘forceful person’). They argue that the same interpretive processes as are employed for literal utterances are used for narrowing, broadening, through to approximation and figurative usage in hyperbole and metaphor. Interpretation is triggered by the search for “relevance” constrained by the principle of least effort: “An input is relevant to an individual when it connects with available contextual assumptions to yield positive cognitive effects (e.g. true contextual implications, warranted strengthenings or revisions of existing assumptions)” (Wilson and Carston Reference Carston and Frápolli2007: 245). Inferences deriving from “explicature,” “implicature,” and context-based assumptions satisfy the expectation of relevance, which causes the interpretive process to stop at whatever interpretation a hearer judges satisfactory in the context of utterance.

Huang (Reference Huang2009) also deals with lexical narrowing, lexical blocking, and pragmatic anomaly and, in addition, contrastive focus reduplication. But (despite his title “Neo-Gricean pragmatics and the lexicon”) he has very little more to say about pragmatics in the lexicon than is found in Blutner or Wilson and Carston.

Copestake and Lascarides (Reference Copestake and Lascarides1997) identified the importance of noting in the lexicon the frequency of particular word senses, in a manner very similar to that independently proposed for a broader range of data by Allan (Reference Allan and Peeters2000; Reference Allan2001 and again in this chapter). Copestake and Lascarides (Reference Copestake and Lascarides1997: 140) write “For example, in the bnc [British National Corpus] diet has probability of about 0.9 of occurring in the food sense and 0.005 in the legislature sense (the remainder are metaphorical extensions, e.g. diet of crime).” In section 12.2 below I introduce a credibility metric like that of Copestake and Lascarides which applies to (some) nonmonotonic statements within the lexicon. I argue the case for nonmonotonic statements in the lexicon in entries for nouns in section 12.3 and for verbs in section 12.4. In section 12.5 I discuss the pragmatic intrusions into the interpreting of collectives and collectivized nouns. This leads naturally to a consideration in section 12.6 of the entries for animal nouns that may refer to either the animal's meat or its pelt (after Allan Reference Allan1981; Nunberg and Zaenen Reference Nunberg, Zaenen, Tommola, Varantola, Salmi-Tolonen and Schopp1992). Section 12.7 takes up the dictionary entry for and; 12.8 discusses the pragmatic component of lexicon entries for sorites terms; 12.9 looks at the place of “prefabs” or “formulaic expressions” in the lexicon and 12.10 tackles ways in which connotation might be incorporated into entries for listemes. Section 12.11 summarizes the chapter.

12.2 A credibility metric

In some of what follows it will be helpful to use a credibility metric for a proposition. The truth value of a proposition p hinges on whether or not p is, was, or will be the case. What matters to language users is not so much what is in fact true, but what they believe to be true.³ The credibility of p is what is believed with respect to the truth of p, or believed is known, or is in fact known of its truthfulness. Because most so-called “facts” are propositions about phenomena as interpreted by whomever is speaking, we find that so-called “experts” differ as to what the facts are (for instance, with regard to global warming, or what should be done about narcotics, or what is the best linguistic theory). Whether ordinary language users judge a proposition true or false depends partly on its “pragmatic halo” (Lasersohn Reference Lasersohn1999): in any normal situation Sue arrived at three o'clock is treated as true if she arrived close to three o'clock; the slack afforded by the pragmatic halo is restricted by a pragmatic regulator such as precisely or exactly in Sue arrived precisely at three o'clock or Sue arrived at exactly three o'clock.⁴ Mostly, though, truth or falsity is assigned by the ordinary language user on the basis of how credible the proposition is, and this is reflected in the way that language is produced and understood. There is a credibility metric such as that in Table 12.1, in which complete confidence that a proposition is true rates 1, represented as cred = 1.0, and complete confidence that a proposition is false rates cred = 0.0; indeterminability is midway between these two: cred = 0.5. Other values lie in between. (□ is the necessity operator, ⋄ is the possibility operator, symbolizes exclusive disjunction, p means “not-p.”)

Table 12.1. The credibility metric for a proposition

In reality, one level of the metric overlaps an adjacent level so that the cross-over from one level to another is more often than not entirely subjective; levels 0.1, 0.4, 0.6, 0.9 are as much an artifact of the decimal system as they are independently distinct levels in which I have a great deal of confidence. Nonetheless, I am certain that some variant of the credibility metric exists and is justified by the employment of the adverbials (very) probably, (very) possibly, and perhaps in everyday speech. This metric is needed in some lexical entries, as we shall see.

12.3 Semantic specifications for bird and bull

Birds are feathered, beaked, and bipedal. Most birds can fly. Applied to an owl this attribute of flight is true; applied to a penguin it is false. Birds are sexed and a normal adult female bird can lay eggs. It is a defining characteristic that members of the female sex carry ova; I’ll label this function sxF (which can be glossed ‘sexual female’). Where they don't, or the ova are non-viable, the organism can count for our purposes as a gendered female, genF, but not sxF. Mostly, sexual females are gendered females too; see (4) where → indicates semantic entailment.

1. (4) Most(x)[sxF(x) → genF(x)]

Although we do speak of human eggs, nonetheless the default egg is from an oviparous genus such as a bird, so I’ll assume this characteristic ought to be noted in the lexicon.⁵ Based on Allan Reference Allan2001: 252, I propose that the semantic part of the lexicon entry for bird be (5), where ∧ symbolizes logical conjunction, +> indicates (defeasible) nonmonotonic inference (NMI), which could perhaps be referred to as an implicature and which is cancelled for species such as emus and penguins.

The lambda-operator is useful to identify an individual as having a number of properties jointly, e.g. being a member of the set of creatures that are at the same time feathered and beaked and bipedal. In (5) the line bird(x) +> ⋄fly(x) identifies that a bird is most probably capable of flight with a credibility rating of 0.7. In the case of a sparrow, the semantic component of the lexicon entry may look like (6); for a penguin, like (7).

For both (6) and (7) the oviparity of sxF sparrows and penguins is an entailment of their being birds. The credibility of a sparrow being able to fly is estimated at cred ≥ 0.99 (it might be injured), whereas the credibility of a penguin flying is 0 (its not-flying has a credibility of 1).

The first entry under bull in the Oxford English Dictionary (1989) is “The male of any bovine animal; most commonly applied to the male of the domestic species (Bos Taurus); also of the buffalo, etc.” Part of this is more formally stated in (8).

1. (8) ∀x[λy[bull(y) ∧ animal(y)](x) → λz[male(z) ∧ bovine(z)](x)]

I will ignore the facts identified in (9).

1. (9) male(x) → genM(x) +> sxM(x)

(8) is inaccurate because the noun bull is not restricted in application to bovines; it is also properly used of male elephants, male hippos, male whales, male seals, male alligators, and more. The initial plausibility of (8) is due to the fact that it describes the stereotypical bull. The world in which the English language has developed is such that bull is much more likely to denote a bovine than any other species of animal. Peripheral uses of bull

are examples of semantic extension from bovines to certain other kinds of large animals; consequently they require that the context make it abundantly clear that a bovine is not being referred to. This is often achieved by spelling it out in a construction, such as bull elephant or bull whale, which is of greater complexity than the simple noun bull used of bovines – a difference motivated by the principle of least effort (Zipf Reference Zipf1949). There is no regular term for “the class of large animals whose males are called ‘bulls’, females ‘cows’, and young ‘calves’” so in Allan Reference Allan2001: 273 I coined the term *bozine to label it.⁶ The semantics of English bull is given in (10) from which the NMI of bovinity will be cancelled where the animal is contextually specified as giraffid, hippopotamid, proboscid, pinniped, cetacean, or crocodilian.

Once again we see a default interpretation being recorded as an NMI in the lexicon because of the salience of this particular characteristic, namely bovinity, of the default reference (i.e. the denotatum) for bull

. (At first sight a salient meaning should be almost the opposite of a default meaning: something that is salient jumps out at you; by contrast a default is the fall-back state when there is no contextual motivation to prefer any other. On a second look, what qualifies a state to become the default is its salience in the absence of any contextual motivation to prefer another.) The credibility of ≥0.9 is based on my intuition. A search of ten corpora totaling about ten million words (the Australian corpus of English; Australian ICE; the Lancaster–Oslo/Bergen corpus of British texts; the London–Lund corpus; the Freiburg corpus of British texts; the Freiburg corpus of American texts; the Brown corpus of American texts; the Wellington corpus of written New Zealand texts; New Zealand ICE; Kenya–East Africa ICE) revealed no applications of bull to animals other than bovines, nor indeed were such searches useful in confirming or disconfirming any of the other credibility ratings in this chapter.

In this section I have shown that a lexicon entry can be constructed to indicate the necessary components of meaning for the entry and also the most probable additional components of meaning that obtain for most occasions of use but which may be canceled as a function of contextual constraints. These can be seen as prototype effects that, for instance, help distinguish cup from mug and bowl (see Labov Reference Labov, Farkas, Jacobsen and Todrys1978). Traditional Arab and Turkish coffee cups are small bowls with no handle, very similar in configuration to Chinese porcelain teacups. The typical Western teacup or coffee cup has a handle and is accompanied by a saucer. All these types of cup are bowl-like in shape though they are smaller, usually have higher sides, and serve a different function than most bowls. Cups are intended to be put to the lips to convey liquid to the mouth whereas liquid in food bowls is spooned into the mouth; otherwise a bowl is used for food preparation. These kinds of conditions (that distinguish cup from mug and bowl) are encyclopaedic and pragmatic rather than purely semantic.

For each lexicon entry the semantic identity of the listeme is presented as a meaning postulate – cf. (10); for instance, the noun bull is semantically represented by the predicate bull ranging over a variable for the entity denoted. Predicates like bull, animal, male, and bovine are not decomposed into semantic primitives but give rise to certain inferences some of which are necessary semantic entailments, others are probabilistic nonmonotonic inferences. Similar conditions apply to the verb climb, as we see in section 12.4.

12.4 Climbing

Jackendoff (Reference Jackendoff1985) identified some interesting characteristics of the verb climb. From (11) we understand that Jim climbed up the mountain – contrast (11) with (12). We also understand that he used his legs and feet – contrast (11) and (12) with (13).

1. (11) Jim climbed the mountain.

2. (12) Jim climbed down the mountain.

3. (13) Jim climbed (down) the mountain on his hands and knees.

Snakes, airplanes, and ambient temperature lack legs and feet they can use when climbing (which is presumably a metaphorical extension with these actors), and they can't climb down, some other verb must be employed.

In (17) the lexicon entry captures the fact that the default interpretation of climb presumes both upward movement, symbolized by ↑⁷and the use of feet (and therefore legs, too).

NMIs apply not just to nouns and verbs but potentially in any lexicon entry.

12.5 Collectives and collectivizing

Allan Reference Allan1976 and Reference Allan2001 discuss the semantics of collective nouns such as admiralty, aristocracy, army, assembly, association, audience, board, class, clergy, committee, crowd, flock, government, and collectivized nouns such as those italicized in (18)–(19).

1. (18) These three elephant my great-grandfather shot in 1920 were good tuskers, such as you never see today.

2. (19) Four silver birch stand sentinel over the driveway entrance.

A definition of collectivizing will be given shortly, but let's begin with familiar collectives.

Collective nouns allow reference to be made to either the set (collection) as a whole or to the set members. In many dialects of English (but not all) the different interpretations are indicated by NP-external number registration

; consider (20).⁸

Whereas singular NP-external number registration indicates that the set as a holistic unit is being referred to, cf. (21), the plural indicates that the set members are being referred to (22). In these and later examples, X and Y are (possibly null) variables for NP constituents; NP_sg is a singular NP, and NP_pl is plural; x, y, z are sets, either unit sets (individuals)⁹ or multi-member sets, so one should understand from (21) and (22) that ∀x[∃y[y⊆x]].

1. (21) ∀x[NP_sg[X N_head[λy[many(y) ∧ collocated(y)](x)] Y]

   → combined_membership(x)]

2. (22) ∀x[NP_pl[X N_head[λy[many(y) ∧ collocated(y)](x)] Y]

   → constituent_members(x)]

Thus, (23) identifies the composition of the committee, while (24) identifies dissension among the membership of the committee.

NPs denoting institutions, e.g. the company I work for, the BBC, the university must be singular (NP_SG in (27) and (28)) when the institution as a building, location, or single constituent body is referred to, as in (25), but can have plural NP-external registration when referring to the people associated with it (26).

The facts with respect to such collective nouns are represented in (27)–(29), where N₀ is the form of the noun unmarked for number.

1. (27) ∀x∃z[N₀[library(x)] → λy[many(y) ∧ book(y) ∧ collocated(y)](z) ∧ x⊇z]

   +> ∃x[NP_sg[X N_0,head[library(x)] Y] ∧ institution(x)]

2. (28) ∀x[NP_sg[X N_0,head[institution(x)] Y] → constituent_body(x) ∨ site(x)]

3. (29) ∀x[NP_pl[X N_0,head[institution(x)] Y] → staff_members(x)]

There is no evidence in (20)–(29) of probabilistic representation being required in the lexicon. The different interpretations are indicated through morphosyntactic choices.

Allan Reference Allan1976 and Reference Allan2001 identify a principle of N₀ usage for English, given in (30).

1. (30) N₀, the form of the noun unmarked for number, is used when the denotation for N is perceived not to consist of a number of significant similar units.

In a plural NP headed by N₀, the absence of plural inflexion on the head noun marks “collectivizing.” Collectivizing signals hunting, conservation, or farming jargon because N₀ is characteristically used of referents that are NOT perceived to be significant as individuals. Early users of the collectivized form were not interested in the individual animals except as a source for food or trophies. Consider the italicized nouns in (18)–(19) and (31)–(34), to which italics have been added.

1. (31) A three month shooting trip up the White Nile can offer a very good mixed bag, including, with luck, Elephant, Buffalo, Lion, and two animals not found elsewhere: Nile or Saddle-back (Mrs. Gray's) Lechwe and White-eared Kob. (Maydon Reference Maydon1951: 168)

2. (32) On the way back to camp we sighted two giraffe on the other side of the river, which were coming down to the water's edge to drink. (Arkell-Hardwicke Reference Arkell-Hardwicke1903: 285)

3. (33) These cucumber are doing well; it's a good year for them.

4. (34) The cat-fishes, of which there are about fifty distinct forms arranged in four families, constitute the largest group, with probably the greatest number of individuals per species. In some parts of the country where nets are little used and fishing is mainly done with traps and long lines, at least three-quarters of the annual catch is of cat-fish. (Welman Reference Welman1948: 8)

The plural NP “cat-fishes” at the beginning of (34) refers to species of cat-fish, whereas the N₀ at the end refers to individuals caught by fishermen. Collectivizing of trees and other plants is much less common than collectivizing animals – from which, perhaps, it derives. Vermin are never collectivized; though individual language users may differ over what counts as vermin. Early uses of the collectivized form were applied to animals hunted for food or trophies. Today, collectivizing occurs in contexts and jargons of hunting, zoology, ornithology, conservation, and cultivation, where N₀ is characteristically used of referents that, as I’ve already said, are not perceived to be significant as individuals. Two possible contributing factors to the establishment of N₀ as the mark of collectivizing are (1) the unmarked plural of deer – which once meant ‘wild animal, beast’, and (2) the fact that meat nouns are N₀ (discussed in the next section). Despite the fact that there is a good deal of variation in the data (see Allan Reference Allan1976: 100f.), collectivizable nouns should be marked as such in the lexicon. Reference will need to be made to the discourse domain being one of the contexts identified above and vermin will need to be excluded. The kind of entry I envisage is (35), which uses giraffe as an example.

1. (35) IF Domain = conservation THEN ∀x[NP_pl[X N₀[giraffe(x)] Y]]; cred ≈ 0.6

Clearly, more work is needed.

12.6 Animals for food and fur

In this section I take up a discussion from Allan Reference Allan1981. Consider the sentences in (36)–(37).

1. (36) Harry prefers lamb to goat.

2. (37) Jacqueline prefers leopard to fox.

Most likely you will interpret the animal product nouns in (36) to refer to meat, such that (36) is paraphrasable by (38), whereas the animal product nouns in (37) refer to animal pelts such that (37) is paraphrasable by (39).

1. (38) Harry prefers eating lamb to eating goat.

2. (39) Jacqueline prefers leopard skin to fox fur.

The converse interpretations are unlikely, especially Jacqueline prefers eating leopard to eating fox.¹⁰ The predicate prefer in (36)–(37) offers a neutral context permitting the default animal product to rise to salience. This suggests that the lexicon entries for lamb and goat, and that for other creatures (such as whale, see (40)) should include a specific application of the formula in (41).

1. (40) In Tokyo, whale gets ever more expensive!

The lexicon entries for leopard and fox should include a specific application of the formula in (43); so will all of the italicized animal product nouns in (42).

1. (42)

   1. a. Jacqueline was wearing mink.

   2. b. Elspeth's new handbag is crocodile, I think.

   3. c. This settee's made of buffalo.

   4. d. The tannery has loads of impala right now.

A mass NP headed by an animal noun will refer to the pelt of the animal denoted by that NP when there is in the clause an NP head or clause predicate describing apparel, accessories to apparel, furniture, the creation of an artifact, or any object likely to be made from leather and any place or process that involves pelts, hides, or leather such that these constrain the domain for the interpretation of N₀. Thus the nonmonotonic inference in (41) is canceled by the implications of the lining in (44); from (43) the NMI is canceled by the predicate eat in (45).

1. (44) I prefer the lining to be made of lamb, because it's softer.

2. (45) All we had to eat was leopard.

More subtle interpretations are required in (46)–(49).

1. (46) A plate of lamb can be worn by no-one.

2. (47) The girl holding the plate was wearing rabbit.

3. (48) The girl who wore mink was eating rabbit.

4. (49) Because she decided she preferred the lamb, Hetty put back the pigskin coat.

In (46) “plate of lamb” identifies meat. Although the most likely interpretation of a plate of steel is ‘a plate made of steel’ (cred ≥ 0.99), a plate of lamb is, with similar credibility, interpreted as ‘a plate bearing food’. The predicate “wearing rabbit” in (47) identifies the rabbit pelts as apparel (again, cred ≥ 0.99) and, likewise, “wore mink” in (48) identifies mink as apparel while the predicate in “eating rabbit” coerces the reference to rabbit meat. In (49) “the lamb” is most likely to be interpreted as meat (cred ≥ 0.8) until this is revealed as a “garden-path” misinterpretation, corrected by the preference for a porcine pelt in the second clause, which cancels the original NMI, replacing it with the coerced interpretation “lambskin coat”.

In this section I have claimed that animal nouns in mass NPs which denote a product from the dead animal typically refer to either the animal's flesh or its pelt, but this probabilistic inference can be canceled by certain contextual elements that condition the domain for interpretation. Credibility rankings can be assigned as shown in (50). However, in (50) these rankings are based on my intuition, although they ought to be made on the basis of the frequency of interpretations retrieved from large and diverse corpora.

1. (50) NP_mass [N[λy[lamb(y) ∧ animal(y)](x)]] +> meat_of(x); cred ≥ 0.8

   IF NOT meat_of(x) THEN pelt_of(x)

   NP_mass [N[λy[goat(y) ∧ animal(y)](x)]] +> meat_of(x); cred ≥ 0.7

   IF NOT meat_of(x) THEN pelt_of(x)

   NP_mass [N[λy[rabbit(y) ∧ animal(y)](x)]] +> meat_of(x); cred ≥ 0.7

   IF NOT meat_of(x) THEN pelt_of(x)

   NP_mass [N[λy[leopard(y) ∧ animal(y)](x)]] +> pelt_of(x); cred ≥ 0.9

   IF NOT pelt_of(x) THEN meat_of(x)

   NP_mass [N[λy[fox(y) ∧ animal(y)](x)]] +> pelt_of(x); cred ≥ 0.9

   IF NOT pelt_of(x) THEN meat_of(x)

   NP_mass [N[λy[mink(y) ∧ animal(y)](x)]] +> pelt_of(x); cred ≥ 0.9

   IF NOT pelt_of(x) THEN meat_of(x)

   NP_mass [N[λy[buffalo(y) ∧ animal(y)](x)]] +> pelt_of(x); cred ≥ 0.8

   IF NOT pelt_of(x) THEN meat_of(x)

   NP_mass [N[λy[crocodile(y) ∧ animal(y)](x)]] +> pelt_of(x); cred ≥ 0.8

   IF NOT pelt_of(x) THEN meat_of(x)

   NP_mass [N[λy[impala(y) ∧ animal(y)](x)]] +> pelt_of(x); cred ≥ 0.7

   IF NOT pelt_of(x) THEN meat_of(x)

It would seem obvious that there should be some generalization over nouns that can refer to either meat or pelts; one might refer to the degree of choice between these two alternatives being “graded salience” (Giora Reference Giora2003: 10 and this volume), but this notion is yet more relevant in the lexicon entry for and.

12.7 And

And may conjoin all sorts of sentence constituents and whatever is felicitously conjoined is grouped together such that there is always some plausible reason for the grouping. This “plausibility” valuation is a coherence metric and necessarily pragmatic because it relies on knowledge of whatever world is spoken of; later, I shall question whether it is relevant to the lexicon entry for and. With the exception of some conjoined NPs that I will refer to as NP-*com-Conjunction (and briefly exemplify in (61)–(65)), the conjoined constituents are synonymous with a conjunction of sentences, e.g. in (51e) ‘Two is a number ∧ Three is a number’.

1. (51)

   1. a. Sue is tall and slim.

   2. b. Eric was driving too fast and hit a tree.

   3. c. Elspeth always drove slowly and carefully.

   4. d. Joe and Harriet are tall.

   5. e. Two and three are numbers.

On the assumption that Φ and Ψ are well-formed (combinations of) propositions expressed as well-formed conjunctions in English, the semantics of Φ and Ψ is as presented in (52). There is, in addition, a series of nonmonotonic inferences that exemplify Giora's “graded salience” (Giora Reference Giora2003: 10); they are listed with the strongest contextually possible inference as the first to be considered.

1. (52) Φ and Ψ ↔ Φ ∧Ψ

   1. a. IF cred(Φ → Ψ) ≥ 0.9 ∧ cred(cause(Φ,Ψ)) ≥ 0.8

      THEN Φ and Ψ +> Φ causes Ψ (e.g. Flick the switch and the light comes on; cause ≺ effect¹¹) ELSE

   2. b. IF cred(enable ([do(Ø,Φ)],Ψ)) ≥ 0.9 ∧ cred(Φ → Ψ) ≥ 0.8

      THEN Φ and Ψ +> Φ enables the consequence Ψ ∨ Φ is a reason for Ψ (e.g. Stop crying and I'll buy you an ice-cream; action ≺ consequence) ELSE

   3. c. IF cred(Φ<Ψ) ≥ 0.8

      THEN Φ and Ψ +> Φ and then later Ψ (e.g. Sue got pregnant and married her boyfriend; Φ ≺ Ψ) ELSE

   4. d. IF cred(enable(Φ,[do(S,[say(S,Ψ)])])) ≥ 0.8¹²

      THEN Φ and Ψ +> Φ is background for Ψ (e.g. There was once a young prince, and he was very ugly) ELSE

   5. e. Φ and Ψ +> Φ is probably more topical or more familiar to S than Ψ (e.g. On Saturdays my mum cleans the flat and Sue washes the clothes)

Note the conditional relations in (53):

1. (53) (Φ causes Ψ) → (Φ is a reason for or enables the consequence Ψ) → (Φ temporally precedes Ψ)¹³

Whether the last two discourse-based implicatures of (52) are part of this sequence remains to be determined. However, it is arguable that if Φ is background for Ψ then Φ is prior to Ψ; and if Φ is more topical or more familiar than Ψ, then again, it is arguable that Φ is prior to Ψ; and should these rather tenuous claims be acceptable, then the fact that Φ precedes Ψ when they are conjoined is normally iconic. However, the choice of sequence is a matter of usage (or pragmatics) and is not obligatory, but it does seem to justify a general statement such as (54):

1. (54) Φ and Ψ ↔ Φ ∧ Ψ

   Φ and Ψ +> Φ is prior to Ψ; cred ≥ 0.9

Consider (from (52c)) Sue got pregnant and married her boyfriend: it is false (cred = 0) that Sue's getting pregnant literally causes her to marry her boyfriend, though it may be her reason for doing so, cred ≈ 0.4; but it is quite probable (cred ≈ 0.75) that her marriage to the boyfriend is a consequence of her being pregnant, whether or not he is the biological father-to-be. It is almost certain (cred ≥ 0.9), even though defeasible, that Sue's pregnancy precedes her marriage. Out of any natural context of use it is not possible to determine whether or not saying Sue got pregnant is a background for going on to say that she married her boyfriend. This aside, it has been possible to propose a (partial) lexicon entry for and which includes its implicatures in grades of salience. There seems to be no good reason to treat and as multiply ambiguous semantically when one core meaning can be identified (logical conjunction) and all other interpretations can be directly related to that as a hierarchy of nonmonotonic inferences processed algorithmically. As Ockham wrote: Numquam ponenda est pluralitas sine necessitate ‘Plurality should never be posited without necessity’ (Ordinatio Distinctio 27, Quaestio 2, Ockham Reference Ockham, Gál and Brown1967–88: I, K)

Is it possible to define a plausibility measure for Φ and Ψ that is semantically based? I suspect not. At first sight the acceptability of (55) as against the unacceptability of (56) seems explicable semantically because only living things eat and if Max is dead he is no longer living and this is a semantic entailment of die.

1. (55) Max ate a hearty meal and died.

2. (56) *Max died and ate a hearty meal.

However, the situation seems pragmatically determined in (57)–(60): it is a matter of conventional beliefs about death, going to hospital, and going to heaven.

1. (57) Max went to hospital and died there.

2. (58) *Max died and went to hospital.

3. (59) Max died and went to heaven.

4. (60) *Max went to heaven and died there.

In NP-*com-Conjunction, *com is a ≥2-place predicate with a sense ‘is added to, is mixed or combined with, acts jointly or together with, is acted upon jointly or together with’ (Allan Reference Allan and Peeters2000: 196). It is found in (61), which is not semantically equivalent to (62) – contrast the latter with (51e).

1. (61) Two and three are five.

2. (62) *Two is five ∧ Three is five

A revealing recipe-like paraphrase of (61) is (63), which accounts for the fact that (64) is a paraphrase of (61).

1. (63) Take two_x and take three_y, combine them (*com(x,y)), and you get five_w, cf. Mix flour_x and water_y to make paste_w or just Flour and water make paste.

2. (64) Two and three make five.

NP-*com-Conjunction is recognized when a conjunction of sentences either cannot apply or is unlikely to apply as in (61) and (65).

1. (65) Joe and his wife have a couple of kids.

The subject NP of (65) is most likely NP-*com-Conjunction whereas that of (66) is not. That these judgments are pragmatically rather than semantically plausible is seen by comparing them.

1. (66) Joe and his sister have a couple of kids.

(66) is, given social constraints on incest, most likely an infelicitous manner of expression where the conjunction is intended to be Φ and Ψ with the weakest of nonmonotonic inferences; preferred would be Joe and his sister each have a couple of kids. With respect to (65), although it is true that each of Joe and his wife has two kids, the sentence Joe and his wife each have a couple of kids suggests these derive from former relationships such that the married couple has four children altogether.

12.8 Sorites

Two horses don't constitute a herd nor do ten grains of sand constitute a heap. For collections such as these, denoted by sorites nouns,¹⁴ the number of constituents needed to render the description accurate depends on the nature of the constituents; for example, whereas the least lower bound on a herd of horses might be three, that on a heap of sand is probably more than a hundred. There are sorites predicates like be bald, be tall, be many and sorites adverbs like slowly, loudly. These are invariably gradable and contextually determined as may be seen from the contrasts in (67).

1. (67) tallfor a Pygmyversustallfor a North American basket-ball professional¹⁵

   manypeople thought George W Bush was a foolversusmany of my students didn'tattend class today

   a slug moves slowlyversusthe train went through the station slowly

There is a similar contextual relevance for the nouns: a herd of horses, elephants, or giraffes will typically have fewer members than a herd of wildebeest, though this is not necessarily the case; moreover, it has no bearing on the lexical meaning of herd. The least lower bound on a heap of beans is lower than that on a heap of sand, probably because of the size of the constituent members. Clearly these are facts about the world referred to but are they facts about the meaning of listemes? No, but they are relevant to the propositions in which the listemes occur: for instance, if speakers wish to report the speed at which a slug is moving they need to apply different criteria than when reporting the speed at which a train is moving. It appears from work reported by Hagoort et al. (Reference Hagoort, Hald, Bastiaansen and Petersson2004) that the brain is prepared to do exactly that kind of thing and that contextual information is integrated with semantic information from the start; see also Terkourafi (Reference Terkourafi, Haugh and Bargiela-Chiappini2009b). However, as I’ve said, although this is relevant to the meaning of propositions, we can dispense with such enriched interpretations in the lexicon because they are instances of lexical adjustment: they count as “ad hoc categories” (Barsalou Reference Barsalou1983; Carston Reference Carston2002; Wilson and Carston Reference Carston and Frápolli2007) dependent on a particular domain of discourse. What we see in (67) is a context-induced specification of the meaning for the sorites words. The same holds for bald: various degrees of baldness are characterized in (68)–(70).

1. (68) His hair is thinning / thin ≈ He is balding / going bald / has a bald patch.

2. (69) He is bald.

3. (70) He is completely bald.

The domain of baldness extends from thinning (head) hair to its almost complete absence. It is arguable that (69) is applicable in situations where (68) or else (70) would also hold true, even though the accuracy of (69) might be disputed in favor of either (68) or (70). So, how sorites words should be specified in a lexicon is highly controversial.

Although not directly concerned with the lexicon, there is a large number of proposals discussed in Williamson (Reference Williamson1994), Beall (Reference Beall2003), and N. J. J. Smith (Reference Smith2008). They include supervaluation, subvaluation, and plurivaluation. Smith suggests “talk of the meanings of some terms must always be relative to a group of speakers, whose dispositions regarding the use of those terms play an essential part in fixing those meanings” (Smith Reference Smith2008: 314). This is a recasting of Quine's “There is nothing in linguistic meaning beyond what is to be gleaned from overt behavior in observable circumstances” (Quine Reference Quine1992: 38). To return to (69): what I suggest for the meaning of bald is the minimal semantics of (71).

1. (71) bald(x) → [full_complement_of_hair(x)]

Two speakers, or the same speaker on different occasions, may differ as to what counts as ‘not a full complement of hair’ such that x is bald has a range of truth values; i.e. there is no single state of hair-loss for which it is invariably true of x that x is bald for all occasions and all speakers. A modification like (68) is appropriate to the least lower bound and (70) to the greatest upper bound; (69) applies to both.

Defining sorites terms often invokes alternative points on the relevant scale. For instance many implies a contrast with other points on a quantity scale; more precisely, less than most and greater than a few. In (72), |f∩g| can be glossed ‘the number of Fs that (are) G’.

1. (72) [many(x): Fx](Gx) → |f∩g|> [a_few(x): Fx]G(x)

   +> |f∩g|< [most(x): Fx]G(x)

(I assume that a few x > few x > one x.) The domain referred to significantly affects the actual numbers, as we saw in (67). It is notable that to establish the truth of (73) we cannot look to a specific number because even if that can ever be known, the precise number that justifies the use of “many” will differ for different speakers and even for the same speaker on different occasions.

1. (73) Many US citizens live in poverty.

Although the meaning of (73) falls under the definition in (72) there is also an implication, or perhaps connotation, that (according to the speaker) the number of US citizens living in poverty is greater than it ideally ought to be. Similar conditions hold for Many of my students were absent from class today, which does not imply that more than half of them weren't there, but that ‘more than one might have expected to be absent were in fact absent’ – and that could easily be as little as 5 per cent.

For sorites like tall and slowly it will be necessary to invoke, respectively, the height scale tall > average height > short and the speed scale slow < average speed < fast on condition that these apply to a particular domain or set of domains as shown in (67).

Sorites like herd and heap (in the sense of Eubulides’ soros) involve configurational criteria.

Suppose that three is the least lower bound for a herd or heap and often the number of constituents is many more, often vastly many more. There is

no upper bound. A heap of sand will typically have many more constituents than a heap of logs; though if the domain of discourse is an egg-timer on the one hand and a clear-felled forest on the other, there may not be such a discrepancy. There is no unique quantity that defines a heap, not even a heap of some particular substance; that is, there is no exact number that determines when a quantity of sand constitutes a heap; the roughly conical configuration is a necessary part of the requirement but is insufficient in itself – as is the condition on quantity. However, the semantic extension of heap(s) as in I have a heap of things to do and There were heaps of people at the party has lost all notion of a particular configuration and is roughly synonymous with lots of or many and must be defined in a manner similar to (72).

12.9 Formulaic language in the lexicon

“A formulaic sequence is a sequence, continuous or discontinuous, that appears prefabricated and stored as a chunk, rather than being generated afresh” (Wray Reference Wray2008: 94). Just as metaphor is pervasive in language, so are “prefabs” – a useful term succinctly defined by Erman and Warren Reference Erman and Warren2000 as “specific conventionalized multiword strings.” Especially in the spoken language, people use thousands of them (just look, for example, at www.phrases.org.uk/index.html); but they are also markers of oral literature, religious texts, best-seller scripts, and popular radio and TV shows (see Allan Reference Allan2001, Reference Allan2006b; Corrigan et al. Reference Corrigan, Moravcsik, Ouali and Wheatley2009; Donahue Reference Donahue1991; Goldman Reference Goldman1990; Jackendoff Reference Jackendoff1995b; Jensen Reference Jensen1980; Kuipers Reference Kuipers2009; Paraskevaides Reference Paraskevaides1984; Schmitt Reference Schmitt2004; Wray Reference Wray2002, Reference Wray2008). Prefabs can be classed into at least three groups.

1. Idioms are primarily figurative; they include: a bit of the other; Bob's your uncle; by and large; come a cropper; fuck off; go the whole hog; kick the bucket; put a sock in it; rain cats and dogs; set store by; sleep like a log; spill the beans; sweat blood; the key to.

2. Clichés are primarily nonfigurative; they include: be heavily compromised; be not very well; believe you me; don't do anything I wouldn't do; Good Lord; Happy Birthday! Hot-dog! [= great!]; ladies and gentlemen; out of sight out of mind; reading, writing, and (a)rithmetic; to make a long story short; un je ne sais quoi; you can say that again; you'd better [do A].

3. Catchphrases include: Beam me up, Scotty; Computer says ‘No’; Frankly, my dear, I don't give a damn; It doesn't amount to a hill of beans; Not that there's anything wrong with it; One potato, two potato, three potato, four…; Play it again Sam; S/he loves me, s/he loves me not.