Inferring and sharing

Sperber & Wilson's (Reference Sperber and Wilson1986) RT is an example of what C&L call radical contextualism, in that it takes the influence of context on propositional meanings to be ubiquitous (an aspect of the theory addressed and defended by Carston Reference Carston2002 in particular). That is, the intrinsic content of a linguistic form, such as a sentence, is taken to be regularly – perhaps always – sub-propositional. Particular propositional meanings are conveyed only in context, via the addressee's applying certain principles of reasoning about the speaker's intended meaning.

RT is a particularly coherent radical contextualist framework partly because it is not built primarily upon pre-existing semantic traditions (it is after all a pragmatic framework). Instead, it is developed from fresh reasoning over some basic observations about the context in which linguistic communication takes place: notably, the fact that such communication takes place between individuals who interact not only with each other's utterances but also simultaneously with their environments – a salient part of which is the other interlocutor and everything that may be known or gleaned regarding that interlocutor.

Another basic observation is that linguistic communication is what Sperber & Wilson term ostensive communication: by its very nature, a linguistic utterance conveys the fact that the speaker intends to communicate thereby. This provides the basis for a view of linguistic interpretation as a process of inferring the speaker's communicative intentions. The speaker provides a basis for doing so by formulating an utterance with an awareness of the communicative context. For the speaker this includes information that could reasonably be inferred about the addressee, which gives rise to expectations about how the addressee will react to certain kinds of linguistic stimulus and certain bits of information. RT therefore recognises communication to be a matter not of simply encoding and decoding messages, but partly of ‘mind-reading’ – and the goal of RT is to provide an explanation of how this mind-reading is achieved.

RT is thus an attempt to explain how addressees arrive at what, according to their subjective calculations, must be the communicator's intended meaning. It is a characterisation of what language users do when faced with any of a huge variety of conceivable circumstances, not of some abstract notion of communication in idealised circumstances. That is, RT is not a semantic theory (though it may impact upon semantic theory rather directly). As such it simply does not deal in objective truth in the way that semantic theories tend to. To criticise RT on the grounds that it cannot guarantee objective identity of content is therefore to miss the point of RT. This in itself constitutes, in broad outline, a response to C&L's ‘shared content’ argument. To flesh it out requires a brief presentation of key aspects of RT.

The structure of RT and semantic minimalism

In this section, I offer a skeletal outline of how RT works (for proper introductions to the framework, see Sperber & Wilson Reference Sperber and Wilson1986, Blakemore Reference Blakemore1992, Carston Reference Carston2002, Wilson & Sperber Reference Wilson, Sperber, Laurence and Gregory2004). My aims here are (a) to provide the basis for comparison to C&L's approach, (b) to show how much of RT is based on highly intuitive, minimally controversial observations and (c) to show that RT is not grounded in conventional semantic wisdom but does have ramifications for semantic practice.

RT in effect starts from two basic observations: (i) that what is understood to be communicated by practically any utterance is heavily dependent on context and (ii) that nevertheless people succeed in communicating with each other, commonly without any apparent problem. This means that there must be some set of principles that enables people to recognise the intended meaning with a considerable degree of reliability. I take these observations to be uncontroversial. Certainly C&L would agree with them – indeed, their (2006a) article on ‘content sharing’ is (independently) explicitly structured around observations (i) and (ii).

RT rejects the Gricean picture whereby, in effect, propositions are directly conveyed by sentences uttered (modulo somewhat mysterious processes of disambiguation and reference assignment) and the role of inference is confined to the generation of cancellable implicatures. There are two important reasons for this rejection. First, RT recognises the pervasive influence of context on what is understood to be conveyed by a given utterance – hence the radical contextualist tag. Second, RT starts out from a much more basic position on what kinds of meaning (or ‘content’) there must be: simply that some meaning is encoded in linguistic forms and some is inferred. The meaning encoded in a given linguistic expression is sometimes called its logical form, though I prefer to stick to the less loaded term encoded meaning. Rather than making a priori assumptions about the nature of encoded meaning, RT works from the opposite perspective, asking what principles of reasoning enable us to recover a speaker's full intended meaning, given the significant influence of context on what is understood to be conveyed and the fact that people cannot literally mind-read. The question boils down to what people can take to be evidence for communicative intentions.

One very important kind of evidence available to addressees is of course the encoded meaning of the linguistic expression used. But there must also be a means of evaluating the particular significance of this on a given occasion, in the light of some kind of further evidence. Such further evidence is extra-linguistic and may in principle be drawn from any information that is available to the individual. ‘Available’ here cannot mean ‘objectively knowable’ – what an individual has to work with is just whatever he has reason to believe, to some extent or other. In RT terms, some set of assumptions is manifest to an individual, where ‘manifest’ means roughly ‘accessible or deducible on the basis of available evidence’. A crucial fact for communication is that an individual's assumptions will include many assumptions about other people and about the assumptions that are manifest to them. A given pair of individuals ‘shares’ a given assumption insofar as that assumption is mutually manifest to them: i.e. to each person, the assumption is manifest, and so is the fact that it is manifest to both of them. The set of mutually manifest assumptions constitutes those individuals' mutual cognitive environment.

All this technical talk adds up to the observation that there are things that we feel we can more or less reasonably assume about the outside world, including about each other and about each other's perceptions of things. This much is surely a necessary basis for any theory of communication. The substance of RT therefore lies not in establishing this view of ‘shared content’, but in its proposals for how we home in on just those mutually manifest assumptions that are relevant to a given act of communication. Clearly, it would be inefficient to the point of paralysing communication to require interlocutors to examine every mutually manifest assumption in the course of producing and interpreting utterances. Equally clearly, one of the problems with this would be that some assumptions that are manifest (i.e. potentially accessible) would require considerable mental effort to actually recover and process.

Reasoning on this basis leads to the fundamental mechanism of RT: interpretation proceeds according to a ‘presumption of optimal relevance’, where relevance is conceived of as a ratio of communicative rewards to processing effort. These ‘rewards’ are what RT calls cognitive effects: strengthening or weakening/eliminating existing assumptions or providing new assumptions that interact with existing assumptions to allow for the deduction of further assumptions (these deduced assumptions are known in RT as contextual implications; the apparently rather complicated definition given here merely reflects the intuitive fact that new assumptions with no connection to existing ones are not relevant to an individual).

Such technical details of RT are less important to the arguments of this article than the general picture that RT presents, as follows. Language users have at their disposal the words and structures of a given language, which clearly carry some very specific kinds of meaning but also may regularly be taken to convey a variety of different things, depending on the assumptions that are brought to bear on the process of interpretation. That is, linguistically encoded meaning constrains but does not determine what is conveyed on a given occasion. Which assumptions are brought to bear on interpretation depends on the relative accessibility of different assumptions and more generally on the processing effort involved in reaching a given interpretation – with addressees only considering more costly lines of enquiry if this is necessary to reach a worthwhile interpretation. Communicators, meanwhile, have to put in some effort to anticipate the outcome of such a strategy of interpretation and to tailor their utterances accordingly (this is not altruistic – the communicator buys his own form of reward thereby: getting his message across).

What does this tell us about shared, or shareable, ‘content’ in RT? As the above discussion makes clear, there are ultimately just two key levels at which one could identify ‘content’ in RT. Fundamentally, RT is a single set of principles for calculating intended meaning given the evidence of (i) the encoded meaning of linguistic expressions used and (ii) the ‘broad context’ (i.e. the mutual cognitive environment of the interlocutors).Footnote ³ Therefore, the only truly distinct notions of ‘content’ that RT makes available are the input to and outputs of this process: encoded and inferred meanings.Footnote ⁴

This makes RT strikingly similar to C&L's professed position, which they call semantic minimalism combined with speech act pluralism. They too recognise just two significant levels of meaning: what they call minimal semantic content and speech act content. And, crucially, they too reject the Gricean (and, more generally, moderate contextualist) association of linguistically encoded semantics with intuitively perceived truth-conditional meanings – i.e. with ‘what speakers say (assert, claim, state, etc.)’ by using given linguistic forms in given contexts. The rejection of this is so fundamental to C&L that they term it the Mistaken Assumption. They define this as follows (elsewhere, they call it the Speech Act Conception of Semantics; C&L Reference Cappelen, Lepore and Noel2007, Reference Cappelen, Lepore, Ernest and Barry2006a):

What C&L apparently fail to recognise is that RT quite explicitly rejects this too (see further section 5.1). It may be that relevance theorists tend to express the idea in rather different terms – for example, ‘linguistically encoded semantics falls short of truth-conditionality’ – but since truth conditions are conventionally identified with what the speaker is taken to be committed to (i.e. to have asserted, claimed or stated) by uttering some sentence, this is in fact just another way of expressing the very same point.

As mentioned above, C&L's minimal semantic content is claimed to be propositional and this is a clear difference to RT's claims about encoded meaning. As discussed in section 4, this does relate to certain genuine differences between the two approaches, but in other ways the importance of this contrast has been overstated. The fact that C&L's particular notion of propositionality is significantly divorced from usual definitions causes a certain amount of confusion. More importantly, the propositionality or otherwise of minimal semantic content fails to distinguish it from RT's ‘encoded meaning’ with regard to the role each plays in the sharing of content – and hence with regard to the evidence that C&L attempt to use against RT.

Given how closely aligned these two theoretical positions seem underlyingly to be, it is just as well for C&L that their own arguments against RT fail – as I show in the following sections.

‘Non-shared content’ in reality and in theory

C&L's arguments rest on the idea that RT, like other radical contextualist approaches, is committed to the proposition that content is simply not shared (in some important sense). They derive this on the basis of quotations from relevance theorists like the following.

Expanding on this, C&L claim that RT is crucially based upon a notion of similarity of content, instead of actual content sharing (i.e. identity of content). In this they appear, on the face of it, to be justified by remarks such as the following:

There are at least two problems with C&L's characterisation of RT in this regard. First, it embodies claims that are stronger than anything claimed in RT (or in the above quotations): RT recognises that the addressee may not grasp an intended propositional meaning p; there is no claim that p will not be recovered. Individual relevance theorists may possibly differ over just how reliable they believe linguistic communication to be, but RT is at heart an attempt to explain just how commonly addressees do seem to recover intended propositional meanings (see the quotation from Carston below). It is rather important to the plausibility of RT that people should have reason to believe that they can recover each other's communicative intentions – otherwise it might not be worth trying – and if it were not the case that they regularly do so (or come negligibly close to it) they would probably soon find reasons not to believe any such thing. Hence RT in effect presupposes that, by almost any definition, content is regularly shared to a significant extent. However, the precise degree of reliability attributed to linguistic communication is not crucial to the present argument. This is because C&L's position rests on the idea that anything short of a guarantee of shared content is disastrous for communication, so the validity of their arguments ultimately depends on the truth or otherwise of this idea (which I come to in section 3.2).

The second, more important problem with C&L's position is that they imply that ‘similarity’ is an active component of RT. It is not. C&L's criticisms look at content sharing from the point of view of an omniscient third party. In this objective sense, intended meaning and recovered meaning may indeed be only similar. RT, however, is concerned with the addressee's understanding of what must be the intended meaning of an utterance (and how this understanding comes into being). It is this subjective assessment of intended meaning that drives RT as an explanatory framework and (as outlined below) this is also all that is required to understand the kinds of communication that C&L bring up in their arguments against RT.

It seems clear to me that Carston, in the quotation above, refers to similarity only from the point of view of an ‘omniscient’ analyst, without suggesting that this is part of the explanation of how communication is achieved. This is made quite plain by material from earlier in the same paragraph:

The actual work that relevance-driven mechanisms are claimed to carry out in communication, against which the success of RT as a framework must be evaluated, has nothing to do with omniscient third parties, because such handy telepaths are not available for language-users to consult, and RT is a theory of what language-users do. Put another way, I do not read Carston here as claiming that the addressee says to himself, ‘I reckon that interpretation q is reasonably similar to the intended interpretation p and therefore communication has been successful’. Rather, he says to himself, ‘Having employed the best means available, I calculate that the speaker must have intended to communicate q’ – even if an omniscient third party might be able to identify that the speaker really intended to communicate the similar proposition p. It is only by failing to distinguish between these two ways of understanding the role of ‘similarity’ that one can imagine C&L's arguments to have any force against RT.

Some of C&L's misunderstanding seems to come from different notions of ‘successful communication’, a notion appealed to in the first quotation from Carston, above. Here one might have some sympathy: there is a sense in which one would like to reserve the phrase ‘successful communication’ for cases in which (at some crucial level) the intended meaning has been conveyed precisely. If we were to make this terminological decision, then we might be forced to conclude that, on the RT view, communication is rarely perfectly successful. However, even stating that communication may often be less than wholly successful in this sense is not to say that RT rules out its ever being wholly successful. Furthermore, it seems to me that a position that recognises the frequent imperfection of communication reflects our everyday experience much more accurately than one that relies on infallible content sharing to characterise human communication.

Just as importantly, as the above quotation from Sperber & Wilson states clearly, successful communication simply isn't defined in this way within RT. In RT, the goal of communication is to enlarge the interlocutors' mutual cognitive environment; that is (very roughly), communication is successful iff through it two people can tell, on the basis of the evidence available to them, that they have some more assumptions in common than they had before. This does not require the exact reproduction of thoughts (or of intended propositional meanings) – and, again, my intuition is that this is just as well: the exact reproduction of thoughts seems fantastically unlikely and even if it does happen we have no way of recognising it with certainty, given the typical absence of telepathic semanticists from real communicative situations. Nevertheless, the process that just might get people to this happy situation does depend on the addressee's assumption that she can get at what the communicator must have meant to convey (so that it is worth trying to do so).

I do not mean to suggest by this that addressees are never aware that their understanding may be imperfect. We all know the feeling that we may or may not have fully grasped what the communicator meant (not least when reading the work of our fellow academics). This is in effect a failure of communication (or perhaps a problematically limited success, in Sperber & Wilson's sense of success) and it is only in such cases of partial failure that addressees consciously fall back on perceptions or hopes of mere similarity to the intended meaning. However common this may be, it is part of the outcome of relevance-theoretic reasoning in certain circumstances; it is not a part of that reasoning process. To suggest, then, as C&L do, that similarity constitutes a crucial part of the relevance-theoretic explanation of communication is to confuse RT, an explanation of what interlocutors do, with the possible effects of applying RT (as viewed by an omniscient third party).

This is important because any approach that actually took similarity to be an intrinsic part of the process of interpretation would quickly lapse into incoherence. C&L go to some lengths to show this (especially in C&L Reference Cappelen, Lepore, Ernest and Barry2006a, Reference Cappelen, Lepore and Noel2007), and in so doing they merely succeed in very convincingly taking apart a straw man, at least where RT is concerned. Most of their discussion centres around the truth conditions of utterances that are reports of or reactions to another utterance. This gives grounds for suspicion from the outset, as it implies a semantic argument (i.e. one grounded in issues of abstract truth) made on the basis of essentially pragmatic issues (i.e. reports of or reactions to speech acts; see Gross Reference Gross2006 for related argumentation).

As a counter to RT, these arguments do not stand up any better in matters of detail. Take the following argument from C&L (Reference Cappelen, Lepore and Noel2007). Imagine that Naomi has uttered a sentence S and someone else (call him Donald) reports that Naomi said that p, thus conveying that Donald believes Naomi to have conveyed the proposition p when she uttered S.Footnote ⁵ C&L then observe that, in a radically contextualist framework with no direct sharing of content, ‘the proposition attributed to Naomi will (typically) be similar to, but not identical to, the proposition that Naomi intended to communicate with her utterance’ (2007: 119). As discussed above, a more appropriate wording would here replace ‘will (typically) be’ with ‘may be’, but C&L's subsequent remarks are more important: ‘A proponent of [such a framework] … has but two options: (a) she can say that reports like [Naomi said that p] are literally false, or (b) she can say that [Naomi said that p] can be true even though Naomi did not intend to communicate the proposition that p’ (ibid.). Options (a) and (b) are what C&L proceed to call the Revisionist and Conservative strategies, respectively.

The so-called Conservative strategy (b) is what C&L (Reference Cappelen, Lepore, Ernest and Barry2006a) attribute to RT and other radical contextualist approaches, and they proceed there to show how incoherent such a strategy would be. It would, after all, amount to re-defining our very notions of truth and identity such that ‘similar’ propositions could be identified as being truth-conditionally equivalent. As C&L rightly point out, this would wreak havoc with all kinds of inferential procedures. They also point out (2007) that it would be essentially lacking in foundation, as no suitable definition of similarity could exist without a traditional notion of identity.

No matter; this strategy bears no relation to RT. I can only guess that C&L were led to consider it a conservative development of RT on the basis of relevance theorists' references to ‘similarity’ in successful communication, such as those in the quotations from Carston and Sperber & Wilson given above. But recall that the RT definition of successful communication makes no reference to objective evaluation of truth.

Returning to the example, Donald's report of speech and attribution of belief are based on his subjective judgement of the identity of the proposition that Naomi intended to communicate. It is quite clear that the speech report is true as such iff Donald's judgement is right: i.e. iff p is (objectively) identical to Naomi's intended assertion. Otherwise, it will be, strictly speaking, false. But either way he can carry out the report sincerely and effectively. Ultimately, whether the objective truth or falsity of his statement really matters depends on a number of independent factors: just how far off Donald's judgement is and just how important it turns out to be in practical terms for Donald, Naomi and others to have understood one another very precisely. If it is important enough, it will doubtless become clear sooner or later just what was falsely assumed – their subsequent interactions and experiences will very likely provide them with ample evidence. If it is not so important, Donald and Naomi and any interested third party may well get by just fine, never knowing that he has made a report of her speech that is strictly speaking false in some way. A theory of the cognitive processes involved in communication has no responsibility to account for such extraneous factors or their effects (though RT does predict that the addressee will generally get at least close enough to the intended meaning for practical purposes, since he has a good set of principles to guide interpretation).

The same point applies to C&L's other examples in this vein, which involve phenomena such as attribution of belief based on previous utterances and explicit assessment of other people's assertions. RT does not, for example, claim that Naomi believes that p is equivalent to ‘Naomi believes something similar to p’ or that What you said is correct/wrong/outrageous is equivalent to ‘Something similar to what you said is correct/wrong/outrageous’. What it does claim is that because of the pervasive context-sensitivity of language, the speaker of Naomi believes that p may be, to a greater or lesser extent, wrong even if he believes he has heard Naomi express that p, and that sometimes a proposition judged to be correct, wrong or outrageous is not exactly the one that the original speaker intended to convey. Again, the degree to which this mismatch between speaker's thoughts and addressee's understanding actually matters is itself dependent on the nature of the context (as it develops). Considerations of objective truth (or of objective identity or similarity) simply do not enter into the process of communication and as such bear no relation to the only really important definition of ‘content sharing’.

To summarise, C&L's ‘similarity’ argument is confused in several ways. It is an attempt to criticise RT as if it were a semantic theory, when RT is in fact a pragmatic theory (in the particular sense of a theory of the cognitive principles involved in communication). Furthermore (as discussed in section 4), it is self-contradictory for C&L to attempt to run a strictly semantic argument off data such as reports or assessments of utterances, since these all involve implicit reference to ‘what has been said’. This clashes with C&L's rejection of the Mistaken Assumption, whereby the semantics of a natural language is not responsible for intuitions about ‘what is said’. This kind of argument is therefore simply not available to C&L.

The R-strategy

The above rejection of the Conservative strategy clearly encapsulates a version of the so-called Revisionist one. Since (I maintain) this is the standard position within RT, there could hardly be a greater misnomer, so I shall refer to it as the R-strategy (for either ‘Relevance’ or ‘Revisionist’, according to taste).

The R-strategy seems to me to be backed by strong intuitions and all kinds of anecdotal evidence. We can sincerely report the speech of others and, unknowingly, falsely convey that they meant to say something they didn't. Likewise, we can and surely do inadvertently attribute beliefs falsely as a result of misapprehending the intended significance of someone else's statements – even following very careful and extensive communication, as the histories of philosophy and linguistics surely demonstrate amply.

However, for C&L (Reference Cappelen, Lepore, Ernest and Barry2006a, Reference Cappelen, Lepore and Noel2007) the obvious intuition is a different one: that people understand each other really rather precisely, rather a lot of the time, and that this is the basis of many kinds of complex behaviour (though elsewhere C&L do show apparent recognition of more problematic aspects of verbal communication; 2005: 206). This too is obviously true, in some sense. As C&L point out, people often follow instructions, laws and regulations that are conveyed linguistically and show appropriately co-ordinated responses across a range of individuals and contexts. People work together, often successfully, to achieve particular goals; this clearly requires shared understanding of those goals and effective discussion of potential solutions. Very generally, people act on what others communicate linguistically: we form beliefs and justify actions based on what others have said, and, relatedly, we hold others responsible for what they have said. Even relevance theorists behave as though language use is content sharing: they write (more or less) carefully-worded tracts, attempting to specify precisely what they mean (see C&L Reference Cappelen and Lepore2005: chapter 9, for the charge that RT is self-contradictory in this way).Footnote ⁶ From such observations, C&L conclude that the R-strategy cannot be right: there must be some kind of content sharing that is reliable; language must indeed allow us to ‘reproduce thoughts’.

Putting together all the intuitive and anecdotal evidence of the above two paragraphs may seem to lead only to a contradiction: people seem regularly to communicate with such striking effectiveness and with so much at stake that there is simply no room for error, yet we also feel that as communicators, as elsewhere in life, we are only human – we know that fundamental misunderstandings can occur. Though this may appear at first sight to represent a stalemate with regard to the arguments of the present article, it in fact constitutes an argument in favour of the R-strategy (and hence in favour of conventional RT).

To see why, consider the overall nature of C&L's arguments. At first they make out that they stand in opposition to a complete lack of shared content in communication. This implies that they believe that successful communication is explicable as long as content sharing (in the sense of reproducing communicators' thoughts) is possible. In this case, as I have argued, they should have no argument with RT: recall that Carston calls the principles of RT a ‘reliable but fallible’ means to get at intended meanings (but not ‘hopelessly fallible’ or ‘inevitably inaccurate’). Therefore, if C&L see their observations as arguments against RT, they must either have misunderstood RT or believe that even the possibility of failing to grasp someone else's intended meaning entirely cannot be admitted: i.e. content sharing must be complete and guaranteed.

In fact, it seems that both of these things are the case. C&L really do seem to have missed the idea that RT is meant principally to explain how addressees regularly do recover speakers' meanings (even while predicting that they will sometimes fail to do this to a greater or lesser degree). Thus, when C&L refer to the basic RT process of inferentially developing the logical form of an utterance, they claim ‘We have no way to predict in advance which development of these logical forms various [addressees] will end up with … developments of these logical forms can be radically different and it would be a minor miracle if they were not’ (2007: 131). To a relevance theorist, this is precisely a description of the problem that RT sets out to solve. One may argue about how well RT does this, but one cannot criticise RT as though it fails to address this problem, when that is its very raison d'être. All the apparatus of RT, from the notions of manifestness and mutual cognitive environments to the cost-benefit calculations prompted by the presumption of optimal relevance, exists to provide an explanation of what would indeed be a minor miracle otherwise: the recognition of speakers' communicative intentions. Note that everyone, C&L included, requires an explanation of this, since it is uncontroversially required at least for the derivation of implicatures.

At the same time, C&L do seem to require complete and guaranteed content sharing – all their discussion points to this as being the only way to explain communication as we know it. I have already indicated that I find that the idea of guaranteed sharing of thoughts through language simply fails to reflect common experience. Furthermore, RT's recognition of a lack of guarantees merely puts communication into the context of very general facts about human processing of data about the outside world. We have only indirect access not just to other people's minds, but to the outside world as a whole: the evidence for all that we know or believe is filtered through our physical senses and through our ability to process and reason over the data that they provide, both of which have been skewed in certain directions by our evolutionary history and are even then decidedly fallible. And yet we plainly have some pretty good ways of distinguishing more and less reliable judgements about how things are. The striking reliability of much linguistic communication is merely a part of this: we do not have direct access to our interlocutors' communicative intentions, but we plainly have some means of recognising and assessing evidence for what these intentions are.Footnote ⁷

Ultimately, the observations that C&L appeal to – linguistic communication as a basis for collective action, judging others' beliefs, etc. – logically show just the following: people can recognise each other's communicative intentions (i.e. share content) well enough to do all these things. RT not only accepts this, it suggests a mechanism by which it occurs.

Alongside the rather sweeping dismissal of RT (and of all radical contextualism) discussed above, C&L (Reference Cappelen, Lepore and Noel2007) do briefly attempt to explain why they think the particular mechanisms of RT fail to explain how communication is achieved. Their point (though wrapped up in much irrelevant discussion of ‘similarity’, as discussed above) is essentially that different contexts – both external circumstances and individual language users' cognitive states – can be just too different. Attempting to adopt RT terminology, they state that ‘[what] you end with as a result of applying [relevance-theoretic reasoning], starting with the logical form of the sentence you're interpreting … depends on what cognitive effects the utterance has on you, and that again on what beliefs (etc.) that the interpreter has’ (sic; 2007: 131).

This is a misrepresentation of RT, implying that the assumptions of different interlocutors are quite divorced from each other and subject to unconstrained, idiosyncratic variation; that there is no basis for relating one's own assumptions to those of another language user. This ignores the central role of mutual manifestness in RT. Far from depending on just any ‘beliefs, etc.’ that an individual addressee may entertain, interpretation is based on that which interlocutors have good reason to believe that they share (something that we need for many purposes other than linguistic communication). This always exists, though it will be richer in some cases than in others – a fact that affects the way one expresses oneself in different contexts (even in C&L's example of their writing for an unidentified reader they evidently have a kind of reader in mind, one with certain assumptions that they can rely on). C&L's description of the ‘cognitive effects an utterance has on you’ suggests a kind of free association, the quirky sparking of an individual's imagination. But RT is concerned with the ways we recognise just those effects that are intended by the speaker of an utterance. Once again, C&L's attempted criticism of RT works better as a description of the kind of problem that RT sets out to solve – in other words, how things are if one does not assume something like RT.

At this point, a supporter of C&L might retreat to another position, claiming that even if RT hasn't been refuted, either approach is still possible. It might be conceded that RT offers one kind of explanation for the recognition of others' communicative intentions (in contradiction to C&L's claim that radical contextualism ‘makes communication impossible’; 2005: chapter 8), but also noted that C&L's notion of content sharing directly through language offers another kind of explanation, and arguably a much simpler one. This might be the case, were it not for the fact that C&L's notion of content sharing is self-contradictory, as shown in the following section.

Minimal semantic content as shared content

C&L place three key requirements on their level of minimal semantic content:

1. (i) that it be the meaning that is constant across all uses of a given linguistic form, in all contexts;

2. (ii) that it be propositional;

3. (iii) that it be the kind of shared content on which are based such activities as the evaluation of assertions, reporting of indirect speech acts, justification of subsequent actions, and so on.

In the following section, I show how (iii) cannot be correct. In section 4.2, I discuss (ii). This is the source of a substantial difference between C&L's minimal semantic content and RT's encoded meaning. However, as argued below (and at length by Carston Reference Carston2006, Korta & Perry Reference Korta and Perry2006, Borg Reference Borg, Preyer and Peterto appear), it is ultimately incompatible with (i) – at least given C&L's notion of propositionality. Furthermore, C&L's heuristics for minimal semantic content fail to identify content that is necessarily propositional. In this section, I concentrate on the implications of defining minimal semantic content in terms of (i).

That minimal semantic content should be constant across all contexts is stated explicitly by C&L (Reference Cappelen and Lepore2005: 143) and is implicit both in their arguments about the nature of shared content and in the way they state minimal semantic content truth-conditionally. When C&L speak of truth conditions, they do not refer to the common (moderate contextualist) conception of ‘what makes the proposition intuitively expressed by a given utterance true’, but rather something that is indeed considerably more minimal. (1) is an example:Footnote ⁸

1. (1) ‘Steel isn't strong enough’ expresses the proposition that steel isn't strong enough and is true iff steel isn't strong enough (C&L Reference Cappelen and Lepore2005: 61)

Though this kind of statement appears less than dazzlingly enlightening, C&L insist that this is the level at which truth conditions should be stated, arguing that it is the business of metaphysics, not semantics, to determine conditions on truth any more precisely.Footnote ⁹ The point for current purposes is that (1) shows how C&L's strictly semantic argumentation – in common with their essentially communication-based ‘shared content’ arguments – is committed to the idea that the semantic content of a given sentence is just whatever remains common to its meaning across all contexts of utterance. There is no other way to interpret statements like (1).

By this definition, C&L's minimal semantic content is entirely equivalent to the notion of encoded meaning in RT. This is necessarily the case, since encoded meaning is by definition just that which is conveyed by a given linguistic form, irrespective of context. Note that to say that some meaning is encoded is entirely neutral with respect to questions such as whether it has other properties, such as being propositional. Thus, RT's notion of encoded meaning is equivalent to minimal semantic content as long as the latter conforms to (i) (as it must, to carry any of C&L's arguments), but is free of the problems that (ii) brings, as discussed below.

Given this, it is unsurprising that C&L's discussion of the psychological and communicative roles played by minimal semantic content fails entirely to distinguish their position from that of RT. In an entire chapter dedicated to these issues (2005: chapter 12) they merely reach the irrelevant conclusion that RT's arguments against Gricean theorising would not constitute an argument against their notion of ‘semantic minimalism’ (C&L Reference Cappelen and Lepore2005: 180; this is, of course, consistent with my argument that the similarities between C&L's position and RT are substantial and outweigh the differences). In the same chapter, C&L show that they are aware that parallels between their position and RT might be drawn and are (for some reason) very keen to deny their validity. They draw up a list of five differences between the two approaches (2005: 181). All but one of these rest upon the supposedly crucial role of the propositionality of minimal semantic content.Footnote ¹⁰ However, this plays no role in C&L's discussions of communication and brings significant problems at other levels, as outlined below.

Minimal semantic content as propositional content

Differences between ‘minimal semantic content’ and ‘encoded meaning’ arise because of C&L's second requirement (or assumption) regarding minimal semantic content: that it is propositional. Tellingly, this is also where both logical and empirical problems enter into C&L's story.

C&L insist that the propositionality of minimal semantic content is crucial to its psychological and communicative role (e.g. 2005: 181, and the reply to Korta & Perry in C&L Reference Cappelen and Lepore2006b), but it is quite unclear why this should be, given that (as outlined above and shown below) all the communicative phenomena that C&L invoke depend on the existence of a crucial level of shared content, not on its logical properties (Borg, Reference Borg, Preyer and Peterto appear, suggests some reasons why philosophers might care about the propositionality of semantic content, but notes that these are independent of the analysis of communication). It is of course necessary that some linguistically conveyed meaning be propositional, as shown by activities like assessing the truth or plausibility of what someone else says. However, I show below that this cannot be the same kind of meaning as that which is constant across different contexts (and is thus ‘shared’ in C&L's sense), so this observation provides no support for C&L's assumption that minimal semantic content must be propositional.

C&L's self-imposed commitment to the propositionality of minimal semantic content forces them to make this content less minimal than it could be, as Borg (Reference Borg, Preyer and Peterto appear) and Korta & Perry (Reference Korta and Perry2006) point out. It also introduces a difference between their conception of minimal semantic content and RT's encoded meaning, as Carston (Reference Carston2006, Reference Carstonto appear) stresses. In order to stick to what they recognise to be propositional meanings, C&L have to assume that minimal semantic content contains no unsaturated indexicals or ambiguities. The problem is that this immediately makes minimal semantic content something other than that which is shared across different contexts: the saturation of indexicals is a context-dependent process. C&L note that this means that their minimal semantic content ‘cannot be characterized completely independently of context’ (2005: 143; emphasis in original), but do not explain how this is to be squared with the proposed role of minimal semantic content as infallibly shared content.

Here C&L seem to be following the convention in semantics of simply assuming that there exists some self-contained process of assignment of reference to indexicals and that the semanticist has no responsibility to characterise what this is or how it relates to other context-dependent areas of linguistic interpretation (see Larson & Segal Reference Larson and Segal1995 for a typical non-exposition of this idea, for example). This is curious, given that C&L's general concern is meant to be precisely the matter of context-dependence in language. It is particularly strange that C&L should include the referents of overt indexicals in that meaning that is infallibly shared, since this is one of the more obvious sources of misinterpretation in language use (given an utterance of He's ready, surely a response like Who? Mark or Paul? is at least as likely as Ready for what?; yet according to C&L's position, the former question should be pre-empted by ‘shared content’, while the latter is not).

In this sense, then, C&L's particular formulation of minimal semantic content is insufficiently minimal to achieve one of the key roles that they assign to it. I refer the reader to the above-cited works for more fully developed and philosophically grounded versions of this argument. My current concern is with the degree to which the propositionality of minimal semantic content plays any genuine role in C&L's arguments regarding content sharing – and the degree to which propositionality is in any case demonstrated in their heuristics for minimal semantic content. If propositionality is not necessary in either of these ways, the obvious step for C&L or those who otherwise have sympathy for their arguments is to adopt the RT notion of encoded meaning, which provides the logically necessary level of some infallibly shared content without the problems associated with C&L's propositional version.

In fact, no part of C&L's argumentation on the role of minimal semantic content in communication actually seems to depend on its propositional status. Consider the following quotation, which is typical of C&L's reasoning in this area:

The argument here turns upon there being some constant meaning attached to a given linguistic form (such as a sentence). It does not depend upon this constant meaning having any particular properties, such as being propositional (in case it seems that being the complement clause of say may guarantee propositionality in some pertinent sense, see section 4.3). Consequently, the role of minimal semantic content, as envisaged by C&L themselves, does not distinguish it from the encoded meaning that RT makes use of, at least as far as content sharing is concerned.

To summarise this section, C&L's minimal semantic content is torn in two by the conflicting requirements that they place on it. The first is that it should be context-independent meaning. Any level of meaning that genuinely fulfils this criterion is of necessity equivalent to RT's encoded meaning, since this is definitional for encoded meaning. As such, C&L have no room to criticise RT for lacking minimal semantic content, on this definition. The second is that minimal semantic content must be propositional (in C&L's minimal sense). While playing no discernible role in the explanation of communication, this conflicts with both the first requirement and, as shown in section 4.3, the ways in which C&L attempt to identify such content.

Tests for minimal semantic content

C&L (Reference Cappelen and Lepore2005: chapter 7) offer certain tests for minimal semantic content, which are important in numerous ways. They represent C&L's concrete methodological alternative to working only with the moderate contextualist's intuited propositions and they supposedly constitute C&L's most substantial objection to radical contextualism (their other objections, as in 2005: chapters 8 and 9, amount to a version of the case against RT already discussed above). Moreover, they provide the only concrete evidence for propositionality in minimal semantic content. These tests are close relations of some of the tests for shared content mentioned above: again, the basic strategy is to examine reported speech, thus turning a judgement about the content of an original utterance into a judgement about the truth of a second utterance, the speech report.

C&L's primary concern in presenting these tests is to take on contextualism as a whole. They are keen to point out that the use of speech reports distinguishes uncontroversial, overt indexicals, such as pronouns, from other cases of alleged context-sensitivity. Thus, (2b) would be false as a report of (2a) because of the context-sensitivity of the pronoun I (C&L Reference Cappelen and Lepore2005: 89).

1. (2)

   1. (a) Utterance made by George Bush, June 3, 2003: ‘I wasn't ready yesterday.’

   2. (b) Utterance made by Ernest Lepore, June 5, 2005: ‘Bush said that I wasn't ready yesterday.’

In contrast, had Bush uttered a sentence like John is ready, it could be truthfully reported by Lepore with the utterance Bush said that John is ready. This is despite the fact that John is ready intuitively falls short of specifying full truth conditions (it invites the question Ready for what?), so is context-sensitive by contextualist reasoning.

C&L's tests thus undoubtedly demonstrate that items like pronouns (what C&L call the Basic Set of indexicals) are different to other linguistic material. This is problematic for the kind of moderate contextualist who would deal with the ‘incomplete truth conditions’ of strings like John is ready by positing covert indexical elements, but not for genuinely radical contextualists such as relevance theorists. By definition, radical contextualist approaches take the influence of context to be pervasive. Hence it is perfectly consistent with a radical contextualist position to accept that there are some linguistic items that specially encode a particular relationship to context (true indexicals), while others are prey to the influence of context in other ways (see Atlas,Reference Atlas, Jay, Preyer and Peterto appear, for a careful elaboration of the general point that recognising a case of context-sensitivity does not entail a claim of indexicality of any kind). C&L's argument regarding true indexicals and context-sensitivity therefore only has force as an objection to moderate contextualism, even though they present it as an objection to the radical kind.Footnote ¹¹

Still, perhaps these tests do demonstrate something important: the key difference between C&L and RT, which is C&L's claim that encoded meaning is propositional. How else, the argument goes, could we make judgements of truth and falsity about indirect speech reports? Intuitively, when we say Nina said that John is ready (unlike Nina said, ‘John is ready’) we express a proposition concerning Nina's having committed herself to something – something that is true only under certain conditions. Therefore, it seems that any utterance which can be indirectly reported in this way itself expresses a proposition, not just a ‘propositional radical’ or any other form of sub-propositional encoded meaning.

To give another perspective on the same idea, C&L suggest imagining two contexts C1 and C2 in which Nina has pronounced the words John is ready. Even if these contexts are sufficiently different that John was ready for quite different things, we can say that the sentence In both C1 and C2, Nina said that John is ready is true. C&L take this to be an argument against radical contextualism: if ready had to be enriched in order to create a proposition, then in this sentence the complement of said would have to express two distinct propositions simultaneously. This is incompatible with the existence of a single truth value for the whole sentence.

C&L's reasoning here rests on a crucial empirical question that they do not investigate: do indirect speech reports consistently work this way? In particular, can we really make reliable judgements of truth and falsity about such reports? Note that this consistency is crucial to C&L's arguments; if the answer to the question is no, we can no longer assume that the complement of said in such cases is necessarily propositional. In that case, being the complement of said is simply not a test for propositionality – with the consequence that indirect speech reports cannot be used to argue for the propositionality of encoded meaning that putatively distinguishes semantic minimalism from RT.Footnote ¹²

There are various ways in which indirect speech reports of the kind that C&L rely on can fail to report a single proposition. For example, being the complement of said does not force disambiguation of homonyms. I can truthfully utter the sentence In C1 and C2, Nina said that John went to the bank even if I know that Nina was referring in C1 to the financial bank and in C2 to the watercourse bank. Of course, I might well continue … but in C1 she meant that he went to the HSBC whereas in C2 she meant that he went to the edge of the river. Significantly, one may easily do this with regard to C&L's example also: In C1 and C2, Nina said that John is ready – but in C1 she meant he is prepared for his exam, while in C2 she meant he has his raincoat on. I can see no difference in the form of these utterances that would allow us to treat them as identifiably different grammatical phenomena (i.e. the one an ambiguity, hence potentially expressing different propositions under any analysis, and the other a single proposition in C&L's sense). Therefore, unless C&L want to argue that an un-disambiguated homonym contributes to the creation of a proposition, they cannot maintain that indirect reports necessarily contain propositions in their complement clauses.

Similarly, C&L's test fails to disambiguate use/mention ambiguities: consider (3) (where ‘Human Resources’ and ‘Personnel’ are two names for the same thing):

1. (3) (It is true that) in C1 and C2, Elspeth said she didn't want to work in Human Resources

   … but in C1 she meant she wanted to work for a department called Personnel, while in C2 she meant that she didn't want to work in that area at all.

I doubt that C&L would wish to claim that any proposition exists before such an ambiguity is resolved, but they would have to do this in order to maintain the supposed significance of their indirect speech report test. Again, note that C&L do not have the option here of simply dismissing such cases as being intuitively different from their examples. Their avowed purpose in introducing linguistic tests is to get away from reliance on intuitions (2005: 87); the tests they propose in their place cannot themselves depend on intuition but must be purely linguistically defined and must have consistent effects in order to convince.

It seems that indirect speech reports are in a sense not indirect enough to do the job that C&L ask of them. They cannot be relied upon to identify a propositional minimal semantic content and therefore they do not demonstrate that minimal semantic content is necessarily propositional. Without this demonstration, we are essentially left in RT territory: we know there is encoded meaning, but we can't prejudge the nature of it for any given linguistic expression (and, consequently, we cannot rule out the influence of inferential pragmatic processes in constructing any pre-defined kind of meaning).

So much for the theory of minimal semantic content. Where C&L's (Reference Cappelen, Lepore, Ernest and Barry2006a, Reference Cappelen, Lepore and Noel2007) empirical arguments are concerned, there is an equally important problem: the role of minimal semantic content simply does not extend to the functions of shared content that are set out in these arguments. These must involve a different kind of content altogether.

The content that is shared

Consider again the kinds of observations that C&L claim make shared content indispensable: co-ordinated action, collective deliberation, linguistic communication justifying beliefs, holding people responsible for what they say, and so on. Is minimal semantic content capable of supporting such things? If so, then an engineer who says Steel isn't strong enough is held responsible just for that proposition (if we call it that) which is expressed by any and every utterance of this sentence. If the roof collapses, it should be considered reasonable of the engineer to point out that I just said steel is strong enough; I never said strong enough to support the roof. Such utterances would usually be described using words other than ‘reasonable’. The example, simplistic as it may seem, is not a trivial one in the context of C&L's discussion of content sharing: it shows that if ‘being held responsible for what one says’ is evidence for content sharing (as they claim it is), then the relevant kind of content is not C&L's minimal semantic content.

Similar arguments apply to C&L's other examples of content sharing. If Beth says to me The apple is red, I will not be tempted to conclude that ‘Beth believes [the minimal proposition that is true iff the apple is red]’ – something that, as C&L recognise, could cover a huge number of possible conditions, in different contexts (e.g. the apple is wholly red, partially red, red on the inside, painted with red stripes, mostly a colour that would otherwise be called brown but therefore ‘red’ as opposed to being a green apple, etc.). I might well, however, conclude that ‘Beth believes that the apple is red in the particular sense that it is partially red on the outside’ (for example). Similarly, if I am part of ‘a CIA task force concerned with whether Igor knows that Jane is a spy’ (C&L Reference Cappelen, Lepore and Noel2007: 122), I will assume that we are ‘collectively deliberating’ about some rather more specific proposition than the minimal semantic content of Igor knows that Jane is a spy (it probably won't concern us if Igor has figured out that Jane is revealing the secrets of one knitting circle to the members of another, as long as he hasn't noticed that she's also a CIA spook – but both of these situations are compatible with this minimal semantic content). As a final example, consider again C&L's evidence from people's assessments of others' assertions. The Speech Act nature of the relevant ‘shared’ content here is shown by the fact that such assessments can be made about implicatures just as easily as about what is taken to be directly asserted, as shown in (4):

1. (4)

   1. A: Is John really such a bad cook?

   2. B: Well, he's English.

   3. A: That's outrageous/a fair point.

It is clear that minimal semantic content is not what must be shared in order to facilitate these activities. What is needed is, minimally, some notion of ‘what is said’ by a given utterance. As we have seen, C&L's rejection of the Mistaken Assumption makes it quite clear that this is not a significant part of their theory. Instead, ‘what is said’ is merely a part of Speech Act (henceforth SA) content. Therefore, what is necessarily shared here is SA content.

Once this is established, all of C&L's arguments regarding RT and content sharing simply fall apart. If the basis for their examples of the necessity of content sharing is the sharing of SA content, then these examples fail to distinguish in any way between the position of RT and that of C&L themselves. C&L are quite explicit on the fact that, for them as in RT, the recovery of SA content is indirect, context-dependent and indeterminate:Footnote ¹³

Also just as in RT, C&L argue that SA content is in part determined by the same context-dependent mechanisms that generate implicatures (this is their eighth principle of Speech Act Pluralism; 2005: 204). Since C&L's SA content encompasses what most (moderate contextualist) semanticists would call the propositional content of a sentence, this is a striking convergence with RT on a position that opposes dominant assumptions.

To summarise, the basis for collective deliberation, inferences of belief, and so on is demonstrably SA content, not minimal semantic (or encoded) content. C&L and RT agree that SA content is not infallibly shared. It follows that the ‘sharing’ of content in these activities is counter-evidence to RT only to the same extent that it is counter-evidence to C&L's own position. Fortunately for both, we have seen that the arguments made around these phenomena are in any case misdirected: perfect content sharing in this sense is not essential to communication or to the theory of communication.

Given their position on the nature of SA content, C&L in effect presuppose an approach that is at least broadly like RT. Ostensibly, they avoid committing themselves to any particular account of how SA content is derived, claiming that ‘there can be no systematic theory of speech act content’ (2005: 190; emphasis in original) and thereby washing their hands of the responsibility to formulate one. Again, I see a failure to distinguish an objective, external characterisation of individual acts of communication from a characterisation of the internal competences that allow people to make the necessary subjective judgements to communicate effectively. While a fully formalisable mechanism showing the derivation of SA content is most likely impossible, one can theorise about the principles that people employ when required to derive such content. C&L (Reference Cappelen, Lepore, Ernest and Barry2006a) do note that their position requires answers to questions such as ‘How is speech act content determined?’ and ‘how interpreters focus sometimes on one aspect of the speech act content, sometimes on another’. It is unclear why they fail to see RT – and indeed the Gricean and post-Gricean pragmatic traditions as a whole – as an attempt to answer exactly such questions.

The importance of the radical/moderate divide

C&L's distinction between moderate and radical contextualism is an important one. On the other hand, as should now be clear, the reason their arguments against RT fail is essentially because they fail to recognise significant similarities between true radical contextualism and true semantic minimalism. These are quite logical and predictable: the more minimal one's notion of semantic content is, the more work there is to be done by non-semantic, context-dependent processes. C&L's hostility to radical contextualism is therefore rather surprising. I believe that it is based upon certain points of confusion as much as on substantive disagreements of principle. One of the key points of confusion is C&L's idea that both moderate and radical contextualists (including relevance theorists) subscribe to the Mistaken Assumption. By their own definitions, no one who did would actually be a radical contextualist.

The only reason C&L give for lumping together moderate and radical contextualists in this respect (apart from a single quotation from the radical contextualist – but not relevance theorist – Charles Travis) is their common use of what C&L call Context Shifting Arguments (CSAs). This refers to the strategy of showing that what is intuitively asserted by some sentence varies from context to context and on the basis of this claiming that the semantics of the sentence in question is context-dependent. C&L believe that a commitment to the Mistaken Assumption follows from the use of CSAs:

There is a straightforward answer to C&L's rhetorical question, from the radical contextualist's perspective: the intuitions appealed to in CSAs are relevant to the debates in which we are engaged just because almost everyone else assumes them to be theoretically important. Radical contextualists’ use of CSAs can be viewed as implicitly doing the the very same job as C&L's own (Reference Cappelen and Lepore2005: chapter 5) arguments about moderate contextualism: they make a conditional argument. If you believe that ‘what is said (asserted, claimed, stated, etc.)’ is the appropriate level at which to state truth conditions, then you will soon find that you cannot maintain a direct relationship between encoded meaning and truth-conditional meaning – contextual influence is simply too pervasive. In other words, the Mistaken Assumption is a premise used in a reductio of moderate contextualism by both C&L and radical contextualists.

Another way to understand the issue is to recognise that what C&L call the Mistaken Assumption in fact has two distinct parts. The first is the assumption that ‘what is said (asserted, claimed, stated, etc.)’ is the appropriate level of meaning at which to invoke truth conditions. C&L deduce that radical contextualists hold the Mistaken Assumption on the grounds that they seem to subscribe to this first assumption (this is clearly the nature of their commentary on a quotation from Travis, for example; C&L Reference Cappelen and Lepore2005: 56). However, this amounts to a mere terminological point unless one accepts the second part of the Mistaken Assumption: that the semantics of linguistic expressions should account for this level of meaning directly. It is in the very nature of radical contextualism to deny this second assumption and, as such, to deny the Mistaken Assumption. It is somewhat immaterial whether or not one thinks that ‘what is said’ might be spoken of as ‘the truth conditions of an utterance’ if one's express purpose is to show that this level of meaning is not directly related to the semantics of the linguistic expressions uttered.

If C&L's distinction between moderate and radical contextualism has any substance, then what sets apart the radical kind is surely the claim of the generality of contextual influence – or rather (to get to the heart of the RT perspective) the pervasiveness of the influence of extra-linguistic processes on the interpretation of linguistic forms. This means that truly radical-contextualist arguments are intrinsically incompatible with the crucial part of the Mistaken Assumption, which depends upon contextual influence on ‘what is said’ being restricted in such a way that a direct connection between it and encoded meaning can be maintained. It follows that any argument that involves the Mistaken Assumption is by definition a moderate contextualist argument.

Other misconceptions may have arisen, ironically enough, from the very fact that relevance theorists and C&L have a common enemy in moderate contextualism. The problem is that they approach this common enemy from opposite perspectives (in a way that is doubtless related to the origins of RT in linguistics and anthropology, while C&L are philosophers) and hence may fail to realise just how much of a common cause they have.Footnote ¹⁴ Thus, relevance theorists tend to push the notion that contextual influence is pervasive, in contrast to the limited, grammatically triggered contextual influence allowed for in the dominant moderate contextualist paradigm. As part of this, considerable emphasis is placed on the idea that propositional meanings are partly derived by context-dependent inference. Since at this stage the point is to take on moderate contextualism, the propositional meanings referred to here are those intuited truth-conditional forms that moderate contextualists deal in.

C&L's notion of proposition is quite removed from that of moderate contextualists (they put a good deal of effort into arguing that what moderate contextualists consider to be sub-propositional is for them fully propositional; C&L Reference Cappelen and Lepore2005: chapter 5). Thus, we clearly have two separate notions – we could call them Propositionality_CL and Propositionality_MC. When a relevance theorist denies that encoded meaning is propositional (in line with C&L's definition of radical contextualism), this refers to Propositionality_MC.

As we have seen (in section 4.2), Propositionality_CL is not ultimately compatible with RT either. However, this is only because of C&L's line on the saturation of recognised indexicals. This is a genuine and important difference, but nowhere near as important as that between moderate contextualism and either C&L or RT. As long as the discussion concerns examples without recognised indexicals, such as (1), there is little or no conflict between RT and C&L, because both recognise a level of meaning that consists of just whatever is common to all uses of a given linguistic form. If this is Propositionality_CL, then (where these examples are concerned) RT logically cannot be arguing for contextual influence in the creation of Propositionality_CL. Meanwhile, C&L recognise that there is contextual influence in the creation of Propositionality_CL if the sentence does contain a recognised indexical (2005: 143). Thus C&L are in fact in total agreement with RT on the issue of which levels of meaning are subject to contextual influence: everything above and beyond encoded meaning, including the saturation of indexicals. Genuine differences between the two begin with C&L's idea that there is a single, minimal, theoretically and psychologically significant level of meaning that mixes context-independent and context-dependent aspects of meaning. However, this point of divergence is not the logical or ostensible basis of C&L's arguments against RT, which instead rest upon the false notion that RT sees contextual influence where C&L see none.

Moderation as a cardinal sin

Having thrown a lot of criticism at C&L for their take on RT, it is time to give due recognition to their analysis of pivotal assumptions in semantic/pragmatic theory. The nature of my argument has been that RT is immune to the more coherent parts of C&L's criticism just because RT is truly radical (i.e. thorough) in its contextualism. It follows that any approach that strays into moderate contextualism becomes a legitimate target for their minimalist objections. Given this, I believe that C&L's notions of the radical/moderate divide and of the Mistaken Assumption can serve as important points of reference for the assessment of both theory and practice in linguistic analysis (whence my ready adoption of these terms in the present work).

In many ways, the dominant moderate contextualist (e.g. Gricean) view sets the agenda in linguistic semantics and pragmatics. This can skew discussion, leading towards excessive concerns over ‘truth-conditionality’ (i.e. Propositionality_MC), even sometimes among those analysts who clearly fall on the ‘radical’ side of C&L's distinction.Footnote ¹⁵

Relevance theorists are not immune to the tug of moderate contextualism. One case in point is Carston's (Reference Carston, Robyn and Seiji1998) analysis of scalar inferences involving cardinal numerals (surprisingly, since Carston's work is typically characterised by what is in effect thorough argumentation for genuine radical contextualism). Carston's main aim here is to argue against the (unfortunately still widely assumed) neo-Gricean analysis of numerals (originally due to Horn Reference Horn1972), whereby they are taken to encode ‘at least n’ readings and take on ‘exactly n’ readings only by implicature. She produces many convincing arguments to this effect, but moves on to propose a categorical distinction between numerals and other expressions that trigger what would traditionally be termed ‘quantity’ inference (after Grice's Reference Grice, Peter, Jerry and Grice1975 Maxims of Quantity) – including upper-bounded readings of non-numeral quantifiers like many as well as inferences of ‘exhaustive listing’ of clearly ‘non-scalar’ objects.

The problem with this is the nature of the reasoning that Carston employs. On the basis of examples like (5) and (6) (originally due to Horn Reference Horn1992), she argues that the upper-bounded (‘exactly two’) reading of numerals constitutes part of the truth-conditional meaning that the speaker is committed to, while the upper-bounded reading of non-numerals (e.g. ‘many but not all’) is mere cancellable implicature.Footnote ¹⁶

1. (5)

   1. A: Do you have two children?

   2. B₁: No, three.

   3. B₂: Yes, (in fact) three.

1. (6)

   1. A: Are many of your friends linguists?

   2. B₁: No, all of them.

   3. B₂: Yes, (in fact) all of them.

Carston then proposes a ‘sense-general’ lexical semantics for numerals – i.e. one which is systematically underspecified for ‘at least’, ‘exactly’ and ‘at most’ readings (see also Atlas Reference Atlas and Jay2005). She proposes a rough formalisation of this as [X[THREE]]. As Carston explains,

This introduction of a ‘contextual variable’ is recognisably the formal strategy of many moderate contextualists.Footnote ¹⁷ Indeed, Carston explicitly draws on moderate contextualist proposals for quantifier domain restriction (which are subsequently defended at length by the ‘hidden indexicalists’ and critics of RT, Stanley & Szabó Reference Stanley and Szabó2000, and in fact criticised by Carston Reference Carston2002). This might be viewed as a mere formal detail; however, it directly reflects the reasoning that underlies Carston's analysis. This reasoning is definitively moderate contextualist: Carston proposes a certain kind of encoded semantics for the express purpose of accounting for context-dependent effects in intuitions of truth-conditional meaning. This is quite simply the Mistaken Assumption in action.

There clearly are significant differences between numerals and other expressions. Furthermore, contrasts like that between (5) and (6) are in principle interesting and worthy of explanation. But in RT the issue should never be framed in terms of a categorical distinction between items that affect truth conditions and those that don't.Footnote ¹⁸ There are many inherent and fairly uncontroversial properties of numerals, both at the semantic level and in likely conditions of use, which can be argued to account for contrasts like (5) versus (6).Footnote ¹⁹ Furthermore, such contrasts are demonstrably not categorical: it is easy enough to construct examples that show ‘truth-conditionality’ with non-numerals and ‘mere implicature’ with numerals (as I show in Wedgwood Reference Wedgwood2005: 163–167). As such, it is notable that the moderate contextualist element in Carston's analysis – the concern that semantics should account directly for observations of Propositionality_MC – is also the point that leads to empirical inaccuracy.

It is also notable that Horn (Reference Horn1992) is able to use Carston's proposals to minimise damage to the neo-Gricean position: if there is something special about numerals that allows them alone to ‘affect truth conditions’ via inference, they can be sectioned off as an exceptional case, leaving the story otherwise unchanged.Footnote ²⁰ Thus we have an illustration of how slippage into moderate contextualist reasoning is to the detriment of RT arguments both in terms of their internal coherence and their external impact.

It is important to note that this instance of covert moderate contextualism is (to my knowledge) an isolated blip in Carston's work and that such things are, as they should be, rather rare in RT work as a whole. Nevertheless, it serves as a reminder of the importance of remaining suitably ‘radical’. This is significant even at the level of presentation. Moderate contextualist (in particular Gricean) assumptions are deeply ingrained in linguistic semantics and this may all too easily influence how linguists' claims are understood, whether by other linguists or by philosophers (as demonstrated by C&L's misconceptions about RT). In this context, it is important for the relevance theorist to be quite clear about the implications of the pervasiveness of contextual influence on interpretation, and here it can be a real help to be armed with some sharp philosophical tools. In this way, C&L's notions of the radical/moderate divide and the Mistaken Assumption can help relevance theorists to clarify just where certain dangers lie.