A long and technical essay about the word 'the'. Embark at your peril.

Keith Donnellan in his 1966 paper 'Reference and Definite Descriptions' introduces the distinction between attributive and referential uses of definite descriptions. The two uses differ in how they contribute to the truth conditions of an utterance containing them.

The pragmatics of this distinction can not be discussed in isolation. Before I do so I need to explore the logical structures of the two uses, and afterwards I need to compare possible semantic origins for them.

1. Overview

With an attributive use the descriptive part is essential: the definite description denotes precisely the thing satisfying it. This accords with Russell's (1905) original analysis of definite descriptions, as an implicit conjunction of assertions. Strawson (1950) by contrast thought of definite descriptions as introducing references into utterances. Donnellan (1966) accommodates these two analyses by saying there are two uses.

In the referential use the satisfaction of the descriptive part is not essential for the reference to succeed. The purpose of the description is to introduce the reference, and that it does so is pragmatic. The semantics of the description cues the identification of the referent, but it is the referent itself that enters into the truth conditions, not its anterior semantics.

Normally we speak truthfully, because of our intention and our knowledge of the world, so normally the semantics of our description corresponds to the object described. That is, normally the referent of our description is just the object that is satisfied by it, so the attributive and referential uses coincide in result.

One example used by Donnellan is 'the man drinking a martini'. Normally the property of holding a martini glass containing liquid corresponds to the property of drinking a martini. So, seeing a man holding a martini glass containing liquid, it is natural to refer to him as 'the man drinking a martini'. A sceptical view of our perceptions would claim that a more accurate formulation would be something like 'the man who it seems to me is drinking a martini', but I regard this as an outdated and debunked metaphysical position, and don't want to pursue it. But we can consider a minimal pair of descriptions, one justified purely by facts of perception, the other by inference: namely 'the man with a martini glass' versus 'the man with a martini'. The latter can be a false description while the former is true. Nevertheless we would normally call someone with a martini glass 'the man with a martini'. Later I'll discuss how relevance theory justifies this identification.

So when we make a statement about the former, using the latter locution, we are running a risk that circumstances are abnormal. A Russellian analysis of the definite description used in these circumstances would have that the false description makes false the statement of which it is a part, because the statement includes it as part of a conjunction: 'The man with a martini is a spy' unpacks as 'there is exactly one man with a martini' and 'whoever has a martini is a spy'. By contrast, Donnellan's analysis of this statement as containing a referential use unpacks as 'there is a man M with a martini' and 'M is a spy'.

So Donnellan's referential conjunction contains one part identifying an object, and another part making a statement about that object. (The assertoric nature of a statement is not significant here: the same logic applies to 'Do you think the man with a martini is a spy?' and 'Stop the man with a martini!'.) One key case in Donnellan's original distinction is that of misdescription: if unbeknownst to us the referent is a teetotaller with a martini glass containing only water. In this case the identifying description fails to be true, but the other predication does not contain the identification criterion.

What Donnellan has controversially claimed is that the whole statement can still be true in this case. If he's only drinking water, but he is indeed a spy, then 'The man drinking a martini is a spy' is functionally equivalent to something like 'that man is a spy', and is therefore true or false according to whether that man is indeed a spy.

Others have the intuition that this is bizarre: that of course the statement is false. The main riposte to Donnellan's argument is that of Kripke (1977), who distinguishes between what the utterance says and what it communicates. Although what it says is false, it can pragmatically be used to convey a truth. Kripke claims it is parallel to a bare misidentification, such as thinking that Smith is Jones, and therefore using the name 'Jones' to refer to Smith.

The essence of Donnellan's distinction is that the definite description 'the man drinking a martini' contributes not its full semantic content, but the referent obtained by resolving that content: that man.

This referential use is a matter of intention. It is not inherent in the use of the word 'the', nor in the semantics of an extended locution of a noun phrase beginning with 'the'. Later I'll discuss the degree of polysemy of 'the', and whether there is a syntactic difference between the two kinds of definite description, but for now I'll adopt a neutral position that there is a form of words that can be used in two ways. Which use is adopted depends on the speaker's intention, in how they want to describe a situation.

The attributive use of this same example can be illustrated with a meeting of a temperance society. The secretary reports to the president a rumour that a member has been seen surreptitiously drinking a martini. The president can now speak about 'the man drinking a martini' without knowing who it is: he might say 'The man drinking a martini will be expelled'. This falls under a Russellian analysis: there is exactly one person drinking a martini, and the president makes his statement about whoever that person is.

Although this is a sort of quantified use, as if it means 'whoever is drinking a martini will be expelled', it's actually stronger: it's 'whichever man is drinking a martini will be expelled'. The former could be a statement of a temperance society rule: anyone caught drinking will be expelled. This instantiates the second (universal) conjunct of the Russellian analysis. The first (existential) conjunct is satisfied when the president has knowledge that there actually is a man drinking a martini amongst them. Grammatically and semantically, this is a definite man, which is why the definite article is used. What it is not is a specific man: the president can't be specific about who this definite individual is. But the referential use, in contrast, denotes a specific man.

The issue is not really about whether there is a distinct referential use of 'the', but about where access to it lies in the gradation between semantics and pragmatics. It is clear that we can identify an intended object pragmatically, even if it's misdescribed, and apply the intended predicate to that object. Donnellan says there are two semantically different uses: the definite article (or its embedding in a definite description, if that makes a difference) intrinsically has two uses, and so is in some sense polysemous. (Later I shall suggest that a purely syntactic difference might be enough to generate different semantic properties, without requiring polysemy of 'the': the parallel is with defining and non-defining relative clauses.) Under Donnellan's analysis, one of these uses, the referential, is an inferential process resulting in the semantic placement of a referent into the utterance; and it is this enriched proposition that is interpreted for truth.

Against him, under Kripke's analysis the semantics is that of the literally interpreted description, and the inference to the referent takes place on the pragmatic side, once the literal truth of the whole utterance has been assessed. The truth conditions are what radically differ between the two approaches.

2. Logic of conjoined truth conditions

I've analysed the attributive and referential uses as conjunctions of two propositions, given respectively in (1) and (2).

(1) there is exactly one man with a martini, and whoever has a martini is a spy
(2) there is a man M with a martini, and M is a spy

The logic of falsification is different under the two analyses. If the man is a teetotaller but a spy, the referential claim falls neatly into two parts, one false and the other true. The main assertion that M is a spy still obtains. It is on the first part that opinions differ. Some would say that it is another assertion, that there is a man M with a martini, and is therefore truth-conditional. If it's not a martini, then he's not a man with a martini, so 'there is a man M with a martini' is false, so (2) is false. On the other hand we might want to regard it as a false description but not a false assertion, so that it does not have truth conditions of its own. If we read the existential 'there is' as a deictic 'there is', it simply picks out the object M, but does not predicate of it. For this to work we need to regard the deixis as a speech act different from predication: we're commanding 'behold the man' rather than stating 'the man is over there'. The statement is prey to being falsified, whereas ostension is not.

This problem of whether its existential conjunct is falsifiable is the nub of the discordant intuitions about the logical status of the referential use. We could agree that there are clearly distinct referential and attributive uses of 'the', corresponding roughly to 'that' and 'whichever', but disagree whether they allow the falsification of the description to be transmitted to its matrix. This issue of whether the definite description is ostension or predication is what decides whether truth conditions are differentially affected. So it remains to be shown that if ostension is denied, the denotatum satisfies the referential truth conditions if and only if it satisfies the attributive ones.

The referential/attributive distinction is not the specific/non-specific distinction. Both (1) and (2) have existential first conjuncts, but their second conjuncts are very different. The universal quantification in (1) is what makes it non-specific: whoever satisfies the protasis ('has a martini') satisfies the apodosis ('is a spy'). There is no equivalent in (2), which instead has a singular assertion about the previously specified object. This structural difference clearly distinguishes them regardless of whether the existential conjunct is ostensive or predicative. But it can be either in either situation.

We have so far considered an individual we can see. In asserting referentially that he is a spy, part of our evidence might be that we can see him drinking a martini, and we know James Bond drinks martinis, therefore we conclude that he is James Bond, and thus a spy. But suppose our evidence is an intercepted message that a spy will signal his presence to his contact by drinking a martini. We have no idea who the spy is, but look around until we spot someone drinking a martini. Even before that, assuming our intelligence is trustworthy, we can say attributively 'the man drinking a martini is a spy'. Before we have spotted anyone, we are saying it non-specifically about whichever man; afterwards, when we can see an individual, we are saying it specifically about that man. This seems to be a straightforward transition from an attributive to a referential use.

But notice that nothing in this scenario requires the spy to be drinking a martini: only to be holding a martini glass casually enough to be identifiable to his contact. The crucial identification procedure is fallible indifferently between the specific and non-specific uses. So in fact the whoever sense here is still referential; though perhaps that's no longer the most accurate term for it, as the referent is unidentified. The transition is in fact from non-specific to specific.

Contrast this with a truly attributive use: suppose we poison all the martinis. Then we know that whoever picks one up and drinks it will collapse insensible. The property of drinking the martini is essential to the identification, and this does not have to be a whoever sense. If we see a specific person take a martini we know is poisoned, and watch them drink it, we know they will be affected. We might not have been able to identify beforehand who would take it, but once we see someone do it, we know them specifically. So this is an attributive transition between non-specific and specific.

Note also that there is something logically odd about the attributive conjunction in this situation. Suppose that the man referred to is not in fact drinking a martini. Then one of the conjuncts should be false: either there is not exactly one man drinking a martini, or it is not true that whoever is drinking a martini is a spy. But not only is it not clear without further evidence which of these is false, but it is possible that both are still true, if unbeknownst to us there is still one and only man drinking a martini, who is indeed a spy. If there is such a man out of sight, the definite description ought to apply to him. But even if there is no such man, discovering the falsity of 'that man is drinking a martini and is a spy' does not tell us how many men are in fact drinking martinis, nor who if any is a spy. This suggests to me that the attributive analysis is not the right analysis for the referential use: falsification should help us to revise our assumptions.

I have argued that the debate about the referential use turns on disagreement on the nature of the first conjunct of the analysis (2). For Donnellan's intuition of truth conditions to hold we need an ostensive act (2a) rather than a predicative act (2b).

(2a) behold a man M with a martini, and M is a spy
(2b) over there is a man M with a martini, and M is a spy

Formulation (2b) contributes to the truth conditions of the whole in a semantically compositional manner. It states a putative fact, and assigns a reference to the person satisfying the factual conditions, so if he is not drinking a martini then the fact fails to be true, and presumably the reference M fails to be assigned. Then the other conjunct, whose truth depends on this reference, is unsatisfiable, and the whole is false, as it would be under the attributive interpretation.

I said (2b) presumably fails to assign reference. There are two differences between (2a) and (2b): in the latter there is a falsifiable assertion, and the reference is defeasible, whereas the former is intended to be simply an unconditional assignment of reference, and therefore indefeasible. It identifies the referent using the putative property ('with a martini'), but its purpose is not to assert the property, only to fix the referent. We could in theory consider an intermediate possibility, like (2a) but assertoric, but where the reference succeeds even if the assertion fails. But this is too much like a decomposition into (2c), a conjunction of an ostension and an assertion.

(2c) over there is a man M, and M has a martini, and M is a spy

But intuitively, 'The man with a martini is a spy' is not the same claim as 'That man has a martini and is a spy'. This latter shifts material from background to foreground, or from topic to focus, which I believe makes a semantic difference. The intuition that accords with Donnellan is that the background part does not make a falsifiable assertion in its contribution to truth conditions. It presents a reference. So I am going to pursue (2b) and ask how it contributes this reference even when the description of the reference is in error.

Since 'the man with a martini' compositionally implies that the man has a martini, it does not contribute its reference compositionally, that is in the automatic encoding of semantics. There is an inference from the description 'the man with a martini' to that man, whatever his properties are.

3. Pragmatics: relevance, focus, vagueness, and intention

Everyone agrees that 'The man with a martini is a spy' does communicate the proposition 'The man M is a spy', where M is the man believed to be drinking the martini. Pragmatically the reference succeeds. Where Kripke and those like him would disagree is in whether the original statement is true. For them, the pragmatics is in, having been given a false statement, inferring the speaker's intention in making the false statement, and finding the true proposition that led the speaker to make the statement.

Relevance Theory makes the path easy. The hearer follows the speaker's gaze, sees exactly one man holding a martini glass, and draws on general knowledge that martinis and only martinis are served in martini glasses. This makes that man highly salient, and he is immediately identifiable as the referent. There is normally no question of pursuing unlikely possibilities, and it is not relevant if he is not in fact drinking a martini. That fact does not enter into the speaker's description nor into the hearer's interpretation of it, under normal circumstances.

I don't want to pettifog for the sake of it, so I will just sketch what normal circumstances are, and take most of them for granted hereafter, excepting only when I need to vary one to make a point. So far I've been assuming that a man holding a martini glass is pragmatically taken to be a man holding a martini. I haven't even mentioned the stereotypical circumstances required for general knowledge to make this inference: that the man is at a party, or a customer in a bar. We would not make this assumption of a barman pouring a martini, or a Sotheby's assistant exhibiting an antique martini glass.

Normal circumstances also include speakers knowing the words they are using. If the hearer's gaze lights on no-one with any kind of drink at all, except one man with a brandy balloon, they might assess their knowledge of the speaker's vocabulary and worldliness, to decide whether the man with a brandy balloon is an accessible referent.

Normally hearers model the speaker's intention: if the hearer sees Jones holding a martini glass but knows that Jones is a teetotaller, they must then assess their knowledge about the speaker. How well do they think the speaker knows Jones? If in the hearer's estimation the speaker could reasonably infer from Jones's holding a martini glass that Jones is drinking a martini, then the hearer accepts Jones as the accessible referent. Note that this is pragmatic, independent of how the hearer would semantically interpret the statement. If Kripke was the hearer he would deny what was said but know what was meant.

We also have general knowledge that actors while performing do not usually drink real alcohol, but this is different. The man whose part includes drinking from a martini glass is straightforwardly identifiable with the locution 'the man with a martini', just as the actor playing the sheriff can be called 'the sheriff'. This is not in any useful sense a misdescription, but a quite regular transference. At a bar, the bar staff could refer to a martini drinker as 'the martini'. Such things are conventional, part of the semantics.

So far the circumstances I have discussed have made the identification of the putative referent more or less completely right or completely wrong. But we can always think of ranges of circumstances where it is impossible to say definitively whether a description fits. I have been using 'the man with a martini' rather than Donnellan's original 'the man drinking a martini' because this eliminates vagueness in 'drinking'. What if he's just taking one sip to be polite; what if he's just finished it; what if he's going to pour the ghastly thing into a pot-plant as soon as he can do so unobserved? What if he's a transvestite; a transsexual; a seventeen-year-old boy? What if his concoction has a splash of tequila in it, so that no purist would call it a martini?

Some of these can be labelled as misapprehensions: a woman disguised as a man; pretending to drink for politeness; mineral water in a martini glass. Others are vagueness: drinking is a protracted process, not all of which involves touching the liquid. Boys are not wholly different from men. The language any speaker uses in anything they say has some intrinsic vagueness, and it is typically not part of their intentions to assert anything clearer. To predicate that someone is drinking a martini is not usually to predicate that they have the glass up to their lips at the moment of speaking, nor to insist on the exact result of the bartender's craft. Something close enough to count as a martini for the purpose at hand is what is being referred to, and similarly for the other components of any utterance: there is never a straightforward literal proposition devoid of context and intention.

We have techniques for removing vagueness: focus intonation, or metaphysical qualifiers like 'real'. To say something isn't a martini, or isn't a real martini, is in some contexts to rule out the concoction with added tequila; and in some contexts to rule out the water that actors are drinking. But if the action of the play involves spiking or swapping drinks, we as spectators can say that one character has the real martini and the other only has water. We can also still refer to characters by their drinks, with focal intonation: the fictional spy is the man with the martini, not the man with the brandy. As Austin (1962) has shown, we can only distinguish the real from the unreal relative to a context.

Récanati (1989) talks about a diagonal proposition, being the skeletal part of a meaning that is required to be true in any world independently of context, but his example 'I am French' is just possibly one of the few sentences bare enough that the diagonal part makes sense on its own. Real-world examples don't: we can't pin down any term such as 'martini' to a precise eternal meaning, independent of context.

The linguistic illusion that something either is or isn't a martini is perhaps fostered by a pragmatic process of stereotyping: asked to think of something that is not a martini, we think of something that very clearly isn't, such as pure water. Thus we overlook much more debatable cases. These distinctions have been called 'near misses' (Lycan 2000), but these are particularly clear misses, and real near misses can be arbitrarily close. There is a halo of meaning around every word (and about every construction, such as the intermittent progressive 'is drinking'), which means we simply can't mean precisely the eternal word meaning by using the word.

So using a word is a pragmatic act of getting it to play a role in the meaning of an utterance. This applies to 'the man drinking a martini' whether or not he can be precisely described as a man, and drinking, and with a martini; all that is essential for the reference to succeed is that he can be pragmatically described that way. This also applies to attributive uses: if we spike the martini with poison, then whoever drinks the martini will collapse, and if we mask the taste with tequila, this will still be true even if someone wants to claim it's not literally a martini any more. This is also true even if the glass in fact had contained only water, unbeknownst to us.

The supposed literal truth of whether the man actually has a martini is not essential to the identification process, whether that process is attributive or referential: both uses are subject to the same infelicities, and the context is what enables the identification to be made pragmatically. The expression 'man with a martini' used attributively means something like a process 'think of someone who appears to be a man and who has what appears to be a martini', and used referentially the process additionally involves finding a specific person. The essence of the referential use is perhaps that the speaker has already identified the referent, and intends to communicate this identification. Identifiability is a strongly salient factor: Powell (2001) argues that where de re and de dicto interpretations are both readily available, the de re reference, is more relevant because its contextual effects imply those of the descriptive attribution it satisfies.

Intention is the distinguishing factor in cases where there is no question of misdescription, but two senses under which a referent can be described: attributive and referential. We might want to poison the spy because it's our job to get rid of all and any spies and poison is the usual method, or because we're jealous of his seduction of women and success at the gaming table. To intend to poison the spy's martini can be an attributive or a referential intention. And, as always, the poison works in plain water, regardless of the intentional context.

4. Semantics: lexical, syntactic, or inferential

The effect of the two uses is to give two meanings to the word 'the': one a singular deictic similar to 'that', the other a quantifier like 'whichever'. It can be asked whether the word has these meanings intrinsically or gains them by pragmatic effects. The distinction can be lexical, being polysemy or ambiguity; or it can be in differing constructions; or it can be inferential, either as explicature or as implicature.

It has been said that lexical ambiguity is an undesirable analysis, it being more parsimonious to have a single lexical meaning and derive the utterance meanings pragmatically. Myself I can't see this. The more functional a word, the more likely it is to have multiple shades of meaning idiosyncratic to the language. The English definite article can be universal ('The llama lives in Peru'), introductory ('The clocks were striking thirteen'), anaphoric ('There was a llama... The llama...'), and all sorts of other stray uses ('two shillings the lot', 'the more the merrier') in addition to the two under discussion. I would be most reluctant to try to find a common meaning underlying all these. I doubt there is any clear line between ambiguity and polysemy: if there are ambiguities in the use of 'the' they are systematic, not phonetically coincidental like 'bank', so we would expect the same ambiguities to be found in other languages.

In German the idiom of 'the more the merrier' is translated not with definite articles but with je... desto, but it would not be surprising if they had used the same idiom English does. English does not overtly capture the distinction between referential and attributive, it being left to the context; and this covertness is common to many, perhaps most languages. But some do make an overt distinction of a very similar nature. In Spanish an animate direct object is marked with the preposition a if it is specific, but not if it's non-specific, and this is true for both definite and indefinite objects: Quiero ver (a) una bailerina 'I want to see a ballerina'. So the fact that non-distinction in grammar is widespread does not imply that it's a pragmatic artefact. Not all available grammatical distinctions are encoded.

It is illegitimate to assume that if the word 'the' has the same meaning (whatever that is) in two sentences then there is no semantic difference between its uses in the two, and that therefore the difference in use must be pragmatic. This ignores the central role syntax plays. Attributive and referential 'the' are two distinct semantic functions, and there are two possible semantic sources of this functionality: one is a polysemy in the lexicon, and the other is syntactic encoding of a semantic form. Donnellan says his intuition is that neither is needed, but both Neale (1990) and Kripke argue that his position actually involves a semantic claim.

A parallel here is with defining and non-defining relative clauses. I don't need to assume any polysemy of 'who' for the difference between 'pensioners who get free travel' and 'pensioners, who get free travel' to be semantic, not pragmatic. The syntactic difference between them is in the bar level at which the modifier is attached, one within the scope of the determiner and one outside it. Relative clauses are not only a syntactic but a semantic parallel. In the attributive use the qualifier is essential: there is a structure similar to 'whichever man is drinking a martini'; but the referential use has something more like 'that man, who is drinking a martini'. While there is no corresponding prosodic difference between the two uses of 'the', a scope difference at Logical Form is a possibility. As this essay is supposed to be more on the pragmatic side, I won't pursue possible analyses of the distinction in terms of syntactic structures, though it has potential, because it is parsimonious.

An argument by parsimony requires that ambiguous or polysemic copies of 'the' require more cost, perhaps in storage or access or computation, than a single underlying copy translatable into multiple modes. This in turn requires that the pragmatic translation should be done by processes independently motivated, that is which do other translations too; for otherwise the process of interpreting 'the' as universal, anaphoric, or referential would need to be stored and unpacked just like a lexical feature, with no saving.

Récanati addresses this by proposing a feature REF, present on inherently referential words such as deictics and 'I'. Its absence from 'the' does not prevent 'the' from being referential, but makes the property facultative: the context determines whether 'the' is used referentially or not. He demonstrates it is unparsimonious to have an opposing feature for the attributive use as well. The choice of which to interpret as is motivated inferentially, but as part of the explicit enrichment, not pragmatically.

5. Conclusion

I have mentioned various ways of locating the referential/attributive distinction, but not endorsed a particular explanation. My intuition is that the speaker is making a logical distinction at the syntactic level, but since it is not prosodically marked, the hearer has to reconstruct the intention pragmatically from context. It would be difficult to find a single utterance in a single context that is genuinely ambiguous between the two uses.

References

Austin, J., 1962, Sense and Sensibilia, Oxford
Donnellan, K., 1966, 'Reference and Definite Descriptions', in Philosophical Review 75; reprinted in Ostertag (1998)
Kripke, S., 1977, 'Speaker's Reference and Semantic Reference', in P. French et al. (eds), Contemporary Perspectives in the Philosophy of Language, University of Minnesota Press; reprinted in Ostertag (1998)
Lycan, W., 2000, Philosophy of Language, Routledge
Neale, S., 1990, Descriptions, MIT Press; chapter 3 'Context and Communication' reprinted in Ostertag (1998)
Ostertag, G. (ed.), 1998, Definite Descriptions: A Reader, MIT Press
Powell, G., 2001, 'The referential-attributive distinction: a cognitive account', in Pragmatics and Cognition 9: 69-98
Récanati, F., 1989, 'Referential/Attributive: A Contextualist Proposal', in Philosophical Studies 56: 217-249
Russell, B., 1905, 'On Denoting', in Mind 14; reprinted in Ostertag (1998)
Strawson, P., 1950, 'On Referring', in Mind 59; reprinted in Ostertag (1998)

Originally submitted for a Philosophy of Language course at UCL. Don't worry, it sounds like gibberish when I read it too. Please don't point out any logical errors: it's way too late for that.