SORITES, ISSN 1135-1349

Issue #13 -- October 2001. Pp. 52-71

Truth Conditions Without Interpretation

Copyright © by SORITES and John Collins

Truth Conditions Without Interpretation

John Collins

1: Introduction

Donald Davidson has long argued that the concept of meaning should be seen through the lens of a Tarski-style truth theory format. A theory answering to such a format will be successful as a theory of meaning for a language L if it offers an informative interpretation of L-speakers, where such an interpretation affords content attribution to both speech acts and mental states. My brief here is to separate these two components; the optic provided by the Tarski format will be commended but the stated criterion of success will be rejected. Thus, I shall argue that Tarski's format, once suitably re-jigged, is indeed apt to depict a speaker's core semantic competence (§2). This stance will be seen to run counter to a prevailing deflationism about truth, which sees the concept as (in some sense) explanatorily antecedent to that of meaning. Davidson tends to challenge deflationism on the grounds that it gets truth wrong precisely because, under its construal, the concept is no longer serviceable in a theory of meaning. This, I think, simply initiates a stand-off. My tack will be more circumlocutionary. I shall suggest that the Tarski format offers an optimal account of compositionality, while it is genuinely moot whether the kind of non-truth involving accounts of meaning the deflationist is obliged to deal with can so much as accommodate compositionality. Such an approach, of course, is non-demonstrative (and none the worse for that); still, on the assumption that a theory of meaning must account for compositionality, we have a prima facie case for truth being the core concept of semantic competence (§3). This argument, for sure, is Davidsonian in spirit, where I more fundamentally depart from Davidson is on his understanding of the aim of a theory of meaning.

It will be argued that there is no a priori link between truth and translation or interpretation of the kind Davidson attempts to derive from Tarski's work. Instead, it will be claimed that a truth theory's theorems do their work precisely because they are trivial. This stance will be seen to be the natural consequence of a cognitive conception of a theory of meaning which stands against Davidson's interpretive social view of language (§4). This cognitive approach views a theory of meaning as a component of a naturalistic framework for the investigation of human cognitive capacity. Now it might seem that we have extirpated ourselves from a stand-off about truth -- deflationism vs. the truth theoretic conception of meaning -- only to sink ourselves into a another about meaning -- interpretation vs. naturalism. It will be argued, though, that a theory of meaning cannot be screened off from empirical constraints deriving from psychology and linguistics, for the very characteristics of a truth theory which make it initially attractive to Davidson and others (its treatment of compositionality, unboundedness, etc.) are themselves such constraints. A recognition of this parity decisively militates for the cognitive conception over the interpretive social one (§5).

2: From Tarski to Truth

Why are truth conditions fit to articulate meanings? Because the disquotability of `is true' shows that the conditions under which a sentence is true are the very conditions which express the content of the sentence. Schematically:

(TM) X means that p X is true  p,

where the substitutions of `X' are quotation-names of sentences and those for `p' are the sentences stripped of their quotation marks. If we are in a position to determine each instance of the antecedent in the schema, then, trivially, truth conditions are expressive of meaning.Foot note 4_1 Such a procedure, however, hardly counts as the explication of meaning at the hands of truth. Meaning here is a constant relative to truth. Davidson's (1984) seminal idea is that we can reverse this relation by placing the right constraints on a theory which issues in the instances of the consequent of (TM):

(T) X is true iff p,

such that the substitutions of `p' interpret the named sentences which substitute `X' without the trivialising benefit of an appeal to notions of meaning, translation, etc. In other words, a truth theory may serve as a theory of meaning because the procedure by which we would determine the truth conditions of a speaker's utterances is one which would take us from a description of utterances to what the utterances say, to a position where we can attribute content to the speaker and so understand him as a rational creature with the same basic beliefs and desires as ourselves. I shall argue that this conception of a theory of meaning as essentially a theory of interpretation is mistaken. Showing this, however, does not involve dispensing with the idea that a speaker's semantic competence amounts to knowledge of truth conditions. I shall turn to Davidson's thoughts on interpretation in §4. Before that I shall discharge my first burden by offering what I think is a strong argument for the truth conditional conception independent of Davidson's understanding of interpretation.

Davidson's (1984) central idea is that Tarski's (1956) format for the explicit introduction of truth predicates may double as a format for truth theories that are expressive of meaning. Tarski's achievement was to show that, for formal languages that do not contain their own truth predicates, a predicate Tr may be defined that is co-extensional with our intuitive notion of truth over the given objectlanguage. That co-extensionality holds is not part of the definition, nor, of course, is it proved elsewhere; rather, it is enshrined in a material adequacy condition on the definition, Convention T, which constrains the metatheory, of which the definition is a part, to entail an equivalence answering to (T), for each sentence of the objectlanguage, with Tr in place of `is true'. The instances of (T) are theorems of the metatheory: their left-hand sides are metalinguistic names (structural descriptions) of objectlanguage sentences, and their right-hand sides are metalinguistic translations of the sentences described to their left.

The real interest Tarski has for a theory of meaning is his method of recursively defining a relation -- satisfaction -- from which a truth predicate may be explicitly introduced as a special case. We can think of satisfaction as a generalised reference function SAT that maps infinite sequences of objects from the universe of discourse onto each wff of the object-language L such that SAT(φ), for any wff φ L, is computable from the values of SAT taking the primitive wffs that feature in φ as arguments. SAT defined for the primitive expressions of L constitutes the base of the definition. In this way the definition of SAT mirrors the recursive definition of `wff of L' which is also part of the metatheory. SAT, of course, can be explicitly defined, given some restrictions on the objectlanguage, either set theoretically or via second-order quantification. Being so defined, Tr + metatheory is a conservative extension of the metatheory, i.e., the deductive strength of the metatheory is unchanged. The metatheory in general contains conceptual resources no greater than those expressed in L, save for the set theory, logic, and theory of syntax of L required for the definition of SAT. The eliminability of Tr at the hands of such resources shows that the metatheory is free from paradox, or at least the definition introduces no paradox.

Equally, however, eschewing explicit definitions, and so departing from Tarski's aims, we can give the clauses of the definition an axiomatic status. The recursion is still effected just as it is when we let `0 + a = a' and `a + suc(b) = suc(a + b)' tell us the values for `+' for any argument pair <a, b> without bothering to define explicitly the set of pairs <<a, b>, c> thereof determined. Even if we lacked the techniques to carry the definition through, the postulates would still obviously give us the right results. Further, we can replace the notion of SAT in the axioms with that of semantic value expressed by the primitive 2-place predicate `v(x, y)'.

Unlike SAT, v assigns objects to expressions that are specifically appropriate to them. Singular terms are assigned individual objects and n-placed predicates are assigned n-tuples. These axioms have the same form as SAT clauses; in particular, the conditions for x to be a semantic value of y are expressed by conceptual resources no greater than those of L as translated into the metalanguage. (As with Tarski's explicit definitions, the rest of the metalanguage is simply to expedite the construction.) Thus, where `NN' and `F(x1,..., xn)' are proxy for singular terms and n-place predicates respectively, the axioms will have the following form:

(A1) v(σ, `NN')  σ = NN

(A2) v(<σ1,..., σn>, `F(x1,..., xn)')  F(σ1,..., σn)

Since our relation v pairs the appropriate object with an expression, when it comes to sentences (as defined by our recursive definition of wff) we go the whole hog and introduce a truth predicate as a primitive symbol. Again, following the syntax of L, true, being the value of a sentence S, can be recursively computed from the axioms such that, with details elided, (T1) holds:

(T1) v(true, `NN is F')  (σ)(σ = NN & σ is F)  (NN = NN & NN is F)  NN is F.

The derivations are as cumbersome as they simple. Assuming axioms for each expression of L (including quantifiers and connectives), we may make a similar derivation for each of the infinite sentences of L. Our theory, therefore, satisfies a modified from of Convention T, as is patent from the above example.

3: Truth and Compositionality

Davidson (1990a, 1997a) holds that that the kind of re-jigging exhibited above is possible shows that the Tarskian definitional format reveals truth to be constitutively linked to the notions of reference and satisfaction. Such is what Davidson means by the structure of truth: the truth relevant combination of primitive expressions that enter into the determination of truth conditions. For Davidson, such revelation of compositionality is non-negotiable for a theory of meaning; that a truth theory, therefore, turns the trick of specifying meaning through derivations that reveal compositional structure is not to be easily relinquished. Now Davidson believes that this spells insuperable difficulties for the common deflationary conception of truth.

Deflationism is a mixed bag; for our purposes, let the notion refer to the thesis that there is no more to an adequate, complete account of truth than the instances of (T), or some generalisation thereof. If this thesis is right, then it seems that truth cannot play the role Davidson envisages, for the instances of (T) can be viewed as constitutive of an agent's understanding of truth only if the agent understands the sentences which go into the schema. Otherwise, while an agent might well recognise the instances (if homophonic) as true, he would not understand them to the extent that he could not `disquote'. In short, for deflationism, TM is basic, not (T) itself, and so some well behaved notion of meaning or content is assumed antecedent to truth conditions (Horwich (1990, 1998), Field (1994), and Soames (1999)).

As Horwich (1999) rightly complains against Davidson (1997a), it is pointless to challenge the deflationary reading of truth on the grounds that it makes the truth unfit to serve as the core concept of a theory of meaning. It is a tenant of deflationism that this is indeed a sound consequence of the thesis (e.g., Field (1994) and Horwich (1998)). The burden thereby acquired, though, is to account for meaning in a non-truth-involving way. Davidson thinks the attempt forlorn. One should not think that this so, however, because Tarski has shown us that truth is set up to play an interpretive role within a theory of meaning. This Tarski did not do (see §4). Equally, one should not think that Tarski showed that truth essentially goes beyond disquotation due to its conceptual entwinement with compositional structure in the shape of satisfaction.Foot note 4_2 Davidson, I think, tends to peddle these two thoughts.Foot note 4_3 He has at his disposal, however, a much stronger argument against deflationism that, to my knowledge, he has never made explicit. The value of this argument is that it does not depend upon anything Tarski is supposed to have demonstrated, nor upon any judgement about what is essential to the concept truth. It does, however, trade upon the ingenuity of Tarski's notion of satisfaction. Explicitly, though, the argument depends merely on the idea that a theory of meaning must account for the compositionality of complex `meanings'.

Let us assume that any theory of meaning must show how the meaning of a complex expression is composed from the meaning of its parts and their structural arrangement. This desideratum may be questioned (e.g., Schiffer (1987) and Jackendoff (1997)). Let me be clear: a theory of meaning is not, in the first instance, an account or analysis of the property of having meaning. No serious investigation is interested in properties as such; rather, the aim is always to explain the phenomena at hand, and properties are identified in terms of their ineliminably featuring in such explanations. A theory of meaning, therefore, is not beholden to preserve the integrity of our intuitive concept. Compositionality, though, is a datum of most of the constructions with which we are competent. Whether or not, then, a given language does, in general, have a compositional semantics, the extent to which it is compositional is the extent to which an account of meaning for the language is constrained to specify what speaker/hearers know such that the exhibited structured competence is explained. I say this prescriptively, but its value will emerge as we go along.

We know that a truth theory of the kind sketched in §2 specifies knowledge which, if possessed, would explain compositionality. We do not know, as indicated, how to generate truth conditions for all constructions; e.g., perhaps subjunctives, non-intersective modifications, modal adverbs, etc. Still, a truth theory cleaves to compositionality because if we know how to assign semantic values to lexical primitives and we know how to specify the semantic value of a construction type C, then we can show that the truth conditions of a token of C are compositionally determined through a derivation of the kind given in §2. This is far from a trivial achievement, as we shall see. Ceteris paribus, therefore, it is a sound hypothesis that truth is indeed the core concept of an adequate account of meaning, i.e., speakers know a truth theory for their languages. Now deflationism explicates truth on the basis of an antecedently determinate notion of meaning; patently, truth conditions cannot enter into this notion. Still, whatever notion is provided, it must be such as to account for compositionality; this is non-negotiable. In effect, then, deflationism is betting on a non-truth theoretic explanation of compositionality, which would show that ceteris is not paribus in favour of the truth theory. But is there such a thing as a non-truth theoretic explanation of compositionality?

There are no shortage of theories that seek to account for meaning apart from truth conditions (semantic values more generally): use theories, conceptual role theories, paradigm theories, verificationist-recognitional theories, etc. Jerry Fodor and others have for a number of years been arguing that none of these alternatives, notwithstanding any other virtues they might possess, come close to accounting for compositionality.Foot note 4_4 The core problem is easily explained. Consider [NP ADJ N'] constructions where the modification is intersective, such as in pet fish, male nurse, red ball, etc. Compositionality tells us that the meaning of, say, pet fish, must be a function of the meaning of pet and fish, and that, moreover, this meaning must determine that pet fish are those things which are both pets and fish. Such, after all, is what it means to be an intersective modification; schematically:

(1) v(x, [NP ADJ N'])  x {ADJ &intersection; N'}.

The problem with the alternatives is that they do not appear to have the required consequences. The properties they cite as constitutive of primitive meanings (lexical or conceptual) are often variant over changes in context. For example, a paradigm pet is not a fish, it is, say, a dog, but the concept of dog does not enter into the meaning of pet fish (similarly for recognitional theories: I can recognise pets and fish when I see them, but it does not follow that I can recognise a pet fish). Likewise, I might be canonically committed to assert that fish are good to eat, but my taste surely does not stretch to pet fish. Suffice it to say that the debate continues. The important moral for my purposes is that it is genuinely moot whether anything other than a truth theory can explain even the most elementary constructions. Such is why I said that its achievement in this regard is far from trivial. Compare, for instance, the problems raised above with the limpidity of the derivation a truth theory allows for:

(2) v(x, `pet fish')  (v(x, `pet') & v(x, `fish'))  (x is a pet & x is a fish)  is a pet fish.

The important feature of the concept of semantic value is its uniformity, i.e., the value of any expression e is the value which e contributes to the value of any complex in which it occurs. In this sense we may say that the concept of semantic value (or satisfaction) is inherently structured: the condition under which a complex expression has the value V just is the condition that holds for the composition of the values of its parts. Thus, to give the semantic value of an expression e is to specify that property of e which contributes to the truth conditions of the sentences in which it occurs. In the other direction, the truth conditions for any sentence may be depicted as systematically arising from the contributions of its parts. This is why a truth theory offers an optimal explanation of compositionality. As far as we can tell, only truth theoretic notions (or satisfaction) have this inherent structure; use, conceptual role, and the rest appear to lack it.

It is tempting to conclude that the concept of truth is essential to a theory of meaning, with, consequently, deflationism indirectly refuted. This is too strong; there is no need to gainsay the future ingenuity of philosophers and linguists. What is clear, however, is that truth has an optimality vis-à-vis compositionality; that is, a truth theory offers us conditions that do compose. Horwich (1998) has recently challenged this conclusion; he claims that compositionality is trivial because it is consistent with whatever we take primitive meanings to be; a truth theory does indeed cleave to compositionality, but so does (more or less) every other theory of which one can conceive. No value, therefore, is accrued by the truth theory's success.

Horwich's idea is this. Consider a sequence of words <w1, w2,...wn> that comprise a sentence S. Let <p1, p2,> be the corresponding sequence of the meaning properties of the words of S, where these are whatever we like. Assume that S has the structure [S [NP N] [VP [V] [NP ADJ N']]] (e.g., Mary likes male nurses), where this expresses some combinatorial principle P with the meaning property pn+1. Now Horwich contends that there is nothing to compositionality other than the idea that to understand a complex expression is just to understand its elements and the way they are combined, an idea which is fully captured by a construction property. In the case of S, this amount to (3):

(3) S means S = S results from the application of P, with meaning constituting property pn+1, to <w1, w2,...wn>, with the meanings <p1, p2,>.

Effectively, a construction property (as given on the right flank) is simply a statement of the value of the function expressed by the structure of a given sentence taking the sequence of primitive meanings of the sentence's words as argument:

(4) pn+1(<p1, p2,>) = S

The important thing to note here, is that we have let the meanings be anything we please, and we have still managed to specify a construction property. Trivially, we can do this for any complex. Since the construction property constitutes what the complex means, we have thus showed how compositionality can be satisfied without appeal to any particular theory as to the nature of meanings. A truth theory, therefore, is no more privileged than a use theory, say, in satisfying compositionality. Thus, deflationism can, ceteris paribus, appeal to a use theory, say, to account for meaning and then give a deflationary characterisation of truth upon this basis, which is just what Horwich proposes.

Horwich takes it to be a virtue of his approach that it shuns any uniform relation, such as satisfaction, that may relate parts and whole via their realising the same property or being a relatum of the same relation. By Horwich's lights, the meaning constituting property of male nurse, say, is not the same kind of property that constitutes the meanings of male and nurse; for the first meaning here is given by a construction property, which is available independently of whatever is understood to constitute the primitive meanings. This rejection of uniformity certainly allows Horwich to claim that, while the meaning of nurse might well be some set of paradigm nurse features (including, presumably, being female), the meaning of male nurse need not be a paradigm at all. This goes no way, however, to show that uniformity is not required to capture compositionality.Foot note 4_5 It does not even answer the complaints raised above, for how the primitive meanings are expressed in the complex meanings remains obscure. What, on Horwich's account, do paradigm nurses, say, have to do with the meaning of male nurse? It seems that Horwich is saying that one could understand male nurse without having a clue about paradigm nurses, for what constitutes the meaning of the complex does not express what constitutes the meaning of its parts, the parts just need to have some meaning or other. The problem here is, I think, a species of a problem Russell faced.

The unity problem bedevilled Russell from the time of Principles (1903) up to the period of logical atomism (1918). Simply put, Russell's problem was how to understand relations (transitive verbs, say) as simultaneously constituents of propositions and relations between the other elements of the proposition, or, in Russell's (1903, §54) terms, how to distinguish between «a relation in itself and a relation actually relating».Foot note 4_6 For example, enumerating the meanings of John, loves, and Mary will not issue in a structure that will articulate that John loves Mary; that is, the enumeration will not constitute what a speaker means who utters the sentence, for the enumeration does not relate John to Mary in any way. It will not do to add a further relation -- a `relating relation' -- to the enumeration; we already have that relation: if we need a further relation, then we would need yet another one to express its `relatingness', ad infinitum. What appears to be required is a constituting structure outside of the proposition, but this threatens to destroy the very notion of knowledge of meaning, i.e., to conflate the act of saying something with what is said. Russell did not resolve this dilemma. If, however, we eschew enumerations in the first place and depict a speaker's understanding as being constituted by knowledge of truth conditions, then the problem disappears, for to know truth conditions is ipso facto to know the semantic contributions the parts make to the whole. We do not have to specify this contribution as a further element to the knowledge since the operative notion of semantic value is uniform across constituents and host complexes: to know the conditions under which the parts have their semantic values just is to know the conditions under which the whole is true.Foot note 4_7 Russell's problem appears to arise only when one attempts to specify lexical meanings independently of their structurual options or, perhaps better, argument structure or logical form. As Russell himself put it: «when analysis has destroyed the unity, no enumeration of constituents will restore the proposition» (Russell (1903, §54)).

Now Horwich's theory precisely destroys unity in that individual words are depicted as having meaning constituting properties independently of the kind of arrangements they can enter into with other words. Horwich's appeal to construction properties and schemata arrives too late, as Russell would have understood. A construction property essentially tells us that if we put meanings together like this, then we get this other meaning. Well, sure; but that is a statement about meanings in themselves, as it were, it is not a statement that depicts meanings `relatingly' in their joint articulation.Foot note 4_8 The logical form of the sentence -- how the constituents are related one to another to engender the complex meaning -- remains outside of the content a speaker articulates when using the sentence. But the problem of compositionality is precisely to account for the form of meaning: we want to know what a sentence means in terms of its parts; we do not need to be told that it means whatever it does because it is made up out of these parts rather than those. That is utterly trivial.

It bears emphasis that compositionality on the truth theoretic understanding is not explained because SAT (or its cognates) maps n-tuples onto lexical items; rather, it is because a truth theory states the conditions under which the assignment holds in complexes, where the content of such a statement may be attributed to a speaker in such a way as to explain the speaker's performance. The inherent structure revealingness of satisfaction would be lost if we thought of a truth theory as simply a mapping of objects onto words, for no such mapping can constitute what a speaker knows who knows a language. The equivalence relations in (2), for example, do not state that pet contributes the set of pets to the meaning of pet fish. This would be of no help at all. Rather, (2) states the conditions under which pet would be satisfied in such a way that those conditions are re-articulated in the conditions for the satisfaction of pet fish. On the other hand, it just makes no sense whatsoever to attribute construction properties to a speaker. To repeat the moral from above: as far as we know, only truth theoretic notions admit to such an explanation, and only such an explanation appears fit to tell us what the meaning of a complex is on the basis of the meaning of its parts.

If these thoughts are anywhere near correct, the deflationist will have the most difficult task to find a non-truth-involving theory of meaning that explains compositionality. And since compositionality is a non-negotiable constraint on meaning, the deflationist will not be in a position to explain meaning per se. If we want a theory of meaning, therefore, we have to acknowledge, it seems, that truth will be a core concept of it, which is just to deny that truth can be explained by an antecedent notion of meaning, which in turn is one way of saying, albeit indirectly, that deflationism about truth is false. As I remarked earlier, I cannot find this argument stated explicitly in Davidson's work. Nevertheless, it is clearly a `Davidsonian' argument and should be where the truth theorist makes his stand against deflationism. This result has a direct bearing, I think, on how we should understood the relation between truth and translation. To this I now turn.

4: Truth and Translation

From his initial statement of the idea that a truth theory might serve as a theory of meaning to his latest writings on truth (e.g., Davidson (2000)), Davidson has claimed that truth is an apt concept to capture meaning because Tarski showed that it essentially involves a notion of translation. The initial idea was that where Tarski keeps translation constant (in homophonic definitions such constancy is simply morphological identity) to get at truth, one may reverse the relation by assuming that a stable attitude towards truth can be identified in speakers which may pivot a translation/interpretation of that which they hold true. This initial idea has undergone many changes in the intervening years, changes I shall not discuss. What I want to focus on is the core idea that truth and translation are conceptually entwined, such being a chief reason why a truth theory is the correct approach to meaning. I think this is a superfluous idea; what is important, I submit, is that a truth theory states knowledge which may constitute a speaker's linguistic competence. Whether a speaker possesses such knowledge is, ultimately, to do with the computational structure of his brain rather than how well we can get on with him; the form of interpretation Davidson takes to be constitutive is, I shall suggest, at best marginal data as to such structure.

Davidson (1990a) puts his point by saying that what is missing from Tarski is the general conditions for the application of truth predicates; if we could specify such conditions, then we would perforce have a translation of the languages to which the predicates apply. Here is how Davidson (2000, p. 70) puts the point:

[Truth is «interesting and important» because] of its connection with meaning. This is the connection of which Tarski makes use, for translation succeeds only if it preserves truth, and the traditional aim of translation is to preserve meaning.

The idea is this: if a theory cleaves to Tarski's Convention T, then the truth conditions for each sentence of can be furnished by respective translations of each L-sentence; translation thus provides the conditions for the application of a truth predicate to L. Translation, however, also preserves meaning. If, then, we can give the conditions for the application of a truth predicate for L, we will have concomitantly interpreted L. An argumentative motif of those who follow Davidson can be best understood, I think, in light of this methodology.


(5) `Snow is white' is true(in-E) iff snow is white.

(6) `Der Schnee ist weiss' is true(in-G) iff snow is white.

(5) appears trivial to us because we know the objectlanguage (English). In other words, the conditions for the application of a truth predicate to our own language are antecedently known to us qua speakers of the language. Yet (6) reveals that the general case is not trivial: laying down the conditions for the application of a truth predicate for a language we do not know provides us with an informative interpretation. This means that the general notion of translation specified in Convention T does indeed equip truth to play a role in the empirical interpretation of languages. Davidson (1999, p. 85) takes this possibility to refute the idea that truth just is disquotation, for disquotation depends upon a particular instance of translation (homophonic), and so can say nothing about the general case. (Davidson correctly points out that the issue relates specifically to idiolects: if one didn't understand `snow is white', then being an English speaker would not make (5) trivial for you.). McDowell (2000) gives the name extended disquotation to the kind of role the truth predicate plays in (6). This is unfortunate. Field (1994) uses the same phrase for a notion that is designed to show that truth just is disquotation after all. A happier nomenclature I have used elsewhere is cognitive opacity (Author 1). The notion is best explained in terms of heterophonicity, although it truly applies to idiolects, not languages as such (what Chomsky (1986) would call E-languages). Consider a monolingual English speaker, Smith, who is given the sentence

(7) `Der Schnee ist weiss' is true(in-G).

Smith can use (7) to express a judgement, say, on Gottlob's assertion (Smith agrees with everything Gottlob says). Smith, however, cannot disquote, for he does not know the truth conditions of Gottlob's assertion. There is more to truth, therefore, than disquotation. This final step, of course, is far too quick, although Davidson is happy to make it. The reason, I think, is that he views truth, retaining my jargon, as cognitively transparent; more precisely, to characterise truth for a language is to make the speakers cognitively transparent to the interpreter, it is to reveal the details of their intentional profile. This is achieved through the essential connection of truth with interpretation (translation) which provides for (6). Disquotationism, on the other hand, is bound to particular languages; it cannot tell us what truth predicates have in common without appealing to a substantial notion of translation, but just such a move is to give up on deflationism.Foot note 4_9

It was suggested above that Davidson, at best, initiates a stand-off with deflationism if he simply claims that truth is not trivial because of its role in a theory of meaning: the deflationist may simply reject such a role and seek to explicate meaning in a non-truth involving way. What does challenge the deflationist is the apparent of optimality of truth in an explanation of compositionality, as elaborated in §3. Do the considerations just rehearsed show otherwise? I think not.

Davidson contends that Tarski established a coupling of truth with translation; in one sense he did, but not so as to confute deflationism. A Tarskian explicit definition makes use of the syntax of the objectlanguage (what, in general, Tarski referred to as the morphology of the language) and the ontology therein expressed. Translation is required for the latter factor: the metalanguage needs to be able to express what the objectlanguage expresses. As Tarski (1956, p. 253) notes, this «has no advantage for the pursuit of the `pure' morphology of the language». The qualification of `pure' is important: Tarski still saw semantic concepts defined for the objectlanguage as being reducible to its morphology; witness Thesis A' (Ibid, p. 273). The reason for this is that the metatheory does not tell us what the objectlanguage sentences mean, the definition of a truth predicate is not informative about meaning or translation; a fortiori, a definition of a truth predicate couched in a metalanguage that did not contain a homophonic translation of the objectlanguage would not be informative either. For example, (6) is logically equivalent to (6') (for convenience, the mass noun snow is treated as a singular term):

(6') [(TG) [`Der Schnee ist weiss' TG & `Der Schnee is weiss' TG  `Der Schnee ist weiss' = `Der Schnee ist weiss' & `Der Schnee' = `Der Schnee' and `ist weiss' = `ist weiss' & (σ)(σ = snow & σ is white)]]  snow is white.

(6') makes patent that a Tarskian truth predicate is employable to tell us what is true given the way the world is, but not what is true given the way meanings are attached to sentences. Convention T can give the opposite impression. An instance of (T) appears to tell us about the meaning of the objectlanguage sentence, but when we substitute definiendum for definiens the appearance melts away. This reveals that, while the definition exploits a translation, the translation is not stated through the definition, it is exogenous. Put differently, Convention T is based on the (TM) platitude:

(TM) X means that p X is true  p

(TM) is common currency for all. The deflationist, for example, can take (TM) as indicating that, given a well-behaved translation (a meaning preserving relation), the content of a truth predicate is exhaustively specifiable in terms of that to which it applies (e.g., Soames (1999)). Such may be taken to be the moral of Convention T.

For all Tarski showed, the aptness of a truth theory to serve as a theory of meaning does not issue from a reflection on truth itself, for there is no conceptual entwinement of truth with an informative translation relation. This conclusion does not preclude the attempt to forge such a relation, but there is no a priori reason to expect success. What we do know for free is that truth conditions can specify content, so much is enshrined in (TM); but this is not what Davidson is after, for (TM) is not informative in the right way, i.e., it is not the application of a truth predicate that makes content transparent where otherwise it is not. It is here, I think, where the project of a theory of meaning, as conceived by Davidson and followed by many others, goes wrong. In simple terms, a theory of meaning need not be informative in the sense of offering a `translation' of a speaker's words.

The idea that a theory of meaning should be informative in the way translations are is not best understood as deriving from Tarski. It is better, rather, to understand the idea as derivative of the view that language is, in some sense, intrinsically social: language is behaviour. If this is so, then a translation constraint makes perfect sense, for understanding a language would amount to being able to interpret and be interpreted. This I think, is Davidson's reasoning.

Davidson's background assumptions are largely inherited from Quine (1960). In particular: what there is to meaning is that which translation makes available. Davidson (1990b, p. 78) takes it to be `obvious... that there can be no more to meaning than an adequately equipped person can learn and observe; the interpreter's point of view is therefore the revealing one to bring to the subject.'

A theory of meaning is not thereby a mere description of behaviour; it does offer a sense of speaker knowledge. Even so, the knowledge condition is counterfactual: if one were to have explicit knowledge of a theory of meaning for L, then one would be in a position to understand L-speakers (Davidson (1984, p. 125; 1990a, p. 312). We do not, then, advance beyond (or below, perhaps) behaviour. The theory describes behaviour in an informationally rich, structured, normative way, but «it does not add anything to... say that if the theory does correctly describe the competence of an interpreter, some mechanism in the interpreter must correspond to the theory» (Davidson (1986, p. 438)).

It is difficult to find an argument in Davidson's work for why the interpreter's point of view should prevail; why, that is, a theory of meaning should be constitutively insensitive to the `mechanism' that generates coherent language use. Davidson (1990a, p. 314) simply states that «what is to be explained is a social phenomenon... language is intrinsically social» [my emphasis]. But how is counterfactual knowledge supposed to be on explanatory duty? My worry here is not so much that of Schiffer (1987), who cannot see how a speaker might know what Davidson attributes to him; rather, it is that a theory of meaning is explanatory of a speaker's linguistic behaviour only if the speaker actually possesses the information recorded in the theory. Davidson is certainly correct if he means that the validity of a theory of meaning does not presuppose a particular instantiating mechanism; after all, information is one thing, its instantiation is another. But if that information is going to enter into an explanation of the speaker's performance, then the information does need to be instantiated (somehow) such that the speaker can employ it. Look at it this way: either the speaker (subconsciously) possesses the information our theory of him records or he doesn't. If he does, then its cognitive availability to the speaker will enter into the explanation of his linguistic behaviour and judgements, otherwise the information would be, absurdly, an idle cog. But if he does not possess the information our theory records, then he must have some other information, in which case we want a theory of that information.

A natural remedy to this quandary, once all species of behaviourism are foregone, is to follow Chomsky (e.g., 1980, 1986, 2000) and claim that the appropriate object for a theory of any aspect of linguistic competence is the internally represented information a subject exploits qua language user. In this light, a theory of meaning is a theory of a system of mental representations, the possession of which empirically accounts for the facts of semantic competence.Foot note 4_10 The general support for such a move is very familiar, and I shall not dwell on it. My aim, rather, independent of whether or not Davidson has a coherent conception of speaker knowledge, is to signal what happens to the putative central role of translation if we make this turn towards internalism and then in the next section offer a suggestion that so turning saves us from an insuperable tension.

On the prevailing view in linguistics, a truth theory is a description of the information represented by a semantically competent subject. It is, in other words, a theory of knowledge of meaning. This knowledge is not constituted by our being able to do anything, still less being disposed to do anything, whatever that might mean. It counts as knowledge just in the sense that it is information we possess and employ in our production and understanding of language. This means, of course, that anyone with the knowledge will, ceteris paribus, be able to do many things, such as translate another speaker. But, again, no such abilities amount to what it is to know a language. Meaning is not use. This being so, how are we to understand the apparent translation a truth theory makes available? Is it accidental?

Rather than say that a truth theory provides a translation or interpretation, we may say that it should issue in a statement of the conditions for sentences' evaluation as true or false (not necessarily disquotationally). A truth theory should do this precisely because it amounts to a test against the intuitive data that the theory records the information relevant to the understanding of the sentence in a way that explains how a sentence's meaning is composed from the meaning of its parts.Foot note 4_11 This was the big moral of §3. So understood, the articulation of content the theory affords will indeed, at least for many constructions, provide translations or interpretations, but these will not be informative in the sense asked for by Davidson's conception of success. Indeed, if they were so informative, the point of the exercise would be lost, for instances of (T) count as secure data because they are intuitively obvious; if they were in doubt, then our theories would not be corroborated by entailing them.

In effect, my suggestion here is the converse of Davidson's initial idea in `Truth and Meaning', where homophonic instances of (T) are taken to be non-corroborative of the theory that entails them because they are uniformative; heterophonic instances, on the other hand, do corroborate their entailing theories because they show that the theories issue in genuinely informative interpretations. Once, however, we drop the notion that a theory of meaning should constitutively double as a theory of interpretation, then the idea that we should be able to frame an English truth theory for German, say, is wholly otiose. Our ability to do such a thing does not enter into the justification of a truth theory as theory of meaning. If truth theories are theories of internally represented information, then neither an English speaker nor a German speaker, qua such speakers, represent an English truth theory for German, or vice versa; and so a truth theory for one language couched in the other will not be a reflection of any underlying competence. Maybe we can construct and test such theories as Davidson envisages, but nothing important about our semantic competence would be obviously revealed by the exercise. For instance, getting to know the truth conditions of utterances we do not antecedently understand will certainly give us a systematic grip on what the speakers are saying (e.g., Field (1978) and Lepore and Loewer (1987)); but this would be due to the co-operation of our theory of mind and our own antecedent semantic competence: one's gaining such a grip does not constitute either one's own understanding or that of one's interpretee.

The justification of a truth theory as a description of a speaker's semantic competence ultimately rests on evidence, not conceptual reflection. Does the information recorded enter into explanations of our judgements about meaning and our typical performance? Does the structure of the theory explain the apparent structure of semantic competence, such as systematicity, compositionality, unboundedness, etc.? Is the information such as to be computationally tractable? Does the semantics offer a smooth interpretation of syntactic structure at the level which interfaces with human conceptual resources in general (LF, say)? E.g., is the theory consistent with the existence of `empty categories'? Answers to these kinds of questions and many others will tell us if we are on the right track.

5: Language in Mind

From Davidson's remarks quoted above, it is clear that he considers the psychology or cognitive science of language to be one thing, and a theory of meaning to be quite another. This separation of the conceptual/normative from the empirical is now very common in a variety of forms. It might thus be countered that the above recommendations are a change of subject rather than a challenge to the interpretive conception of the job of a theory of meaning. This issue, suffice it to say, is far too wide ranging to be properly engaged with here. Still, on the assumption that empirical inquiry into the nature of cognition is not in any sense illegitimate, I think a general problem obtains for the imagined separation, and its instance with regard to linguistic meaning is especially vivid.

Davidson himself was one of the first to appreciate that a theory of meaning must satisfy constraints that are essentially empirical, such as explaining unboundedness, novelty and compositionality. These kinds of constraints now go largely without comment, but they are empirical in that they reflect real contingent features of our linguistic competence. Once such constraints are accepted, then, it is difficult to screen off one's theory from empirical findings. It is has proved convenient for those who a priori insist on the intrinsic sociality of language to characterise a naturalistic attitude to the mind -- what McDowell (1994) disparagingly refers to as «bald naturalism» -- as one which hankers after a reduction of meaning and related concepts to mechanical/computational states: what is not causally accountable for just ain't real. This characterisation presents naturalism as crude scientism, an ontological thesis that straightforwardly conflates is with ought, aetiology with constitution. It is certainly true that many philosophers have reductionist ambitions, but they are not required by naturalism; indeed, a lot of such positions, I should say, are more informed by long standing philosophical intuition than any scientific research. Naturalism is a methodological thesis that has no a priori commitment to what meaning, or anything else, for that matter, really is; rather, it is committed to the construction of empirical theories that attempt to explain the myriad of factors that enter into linguistic competence. There is very little mileage to be gained, therefore, from an attack on naturalism as if it were a thesis about what there is, for such benighted assaults go no way to secure the legitimacy of the initial move to screen-off one's philosophy from empirical findings. It is the very idea that one can determine substantive claims about the sociality of language from conceptual consideration alone that is in tension with naturalism, not the putative irreducibility of the mental.

I am not here begging the question in favour of a language organ (faculty) or any other naturalist project, the point is a general one: whatever the shape of our concepts about natural phenomena appears to demand may turn out to be wrong. In the present case, I have little doubt that the best description of our Lebenswelt involves linguistic inter-subjectivity as constitutive. I also think that mutual interpretation in a certain mode is something language enables rather than what it is; the ability to communicate is neither necessary nor sufficient for linguistic competence. In this regard, our Lebenswelt is misleading and the a priori constitutive theses are just false. One may abjure interest in anything other than the conceptual layout of what it is like to be human, but if so, then it ill behoves one to make substantive claims. The crucial point is that once empirical constraints are acknowledged (e.g., unboundedness), then it is illicit to commit to just those features of linguistic understanding that are consistent with one's pre-conceived a priori views to constrain one's theory Unfortunately, to acknowledge empirical constraints is one thing, to treat them seriously is another. Let us first look at Brandom as an example.

Brandom (2000, pp. 124-9) wants to show that the novelty and unboundedness of linguistic understanding is flush with his inferential-social model of thought. The explanation offered, however, is that the child is trained on a finite number of sentences and is then able to project to the «correct uses» of boundlessly many sentences through a segmentation of the members of the training set into substitution classes of singular terms and predicates. This Bloomfeldianism is patently motivated by Brandom's social view of linguistic understanding rather than a concern with novelty as a real phenomenon that is independent of any favoured theory. Unsurprisingly, it faces serious problems. First, the model is false. Any developmental text book will tell one that children are not trained on a finite number of sentences. Children begin with single words, go through a two word phrasal stage, and then onto full grammatical complexity (function words, inflections, etc.). The whole process is characterised by novelty and creativity. Second, there is no known projection function from a corpus of sentences to a grammar proper. Brandom's substitutional method fails even with such a crude partition as singular term and predicate. It is, on the one hand, too broad, e.g., ` dangerous' admits names (paradigm singular terms) as well as general nouns, gerunds, infinitives, quantifier phrases, wh-phrases, clauses, etc.; on the other hand, it is too narrow; e.g., `...washed herself' excludes all substitutends apart from third person, singular, feminine. The fact is that outside of artificial languages, substitution does not determine grammatical categoricalness; and there are, of course, `scrambled' languages, such as Warlpiri and Latin, with free order, where phrasal constituents may even be discontinuous. Third, what explains novelty is structure, e.g., that [NP NP PP] is recursive means that if one understands [X [PP [Y]]] (e.g., the boy behind the girl), then one will understand [X [PP [Y [PP [X]]]]], etc. If one does not have the grammatical concept of a preposition, under which a given preposition may freely take NP arguments, then being able to use a given PP will not ipso facto enable you to use any other PP iterated thereof. The best explanation we have of this is that grammatical selection features are given in the lexicon. Perhaps this is wrong, but if words are not distinguished by grammatical features, then novelty and boundlessness are not explained; and if words are so distinguished, then the projection Brandom has in mind is wrong: it is projection from the lexicon that is explanatory, not from a corpus of sentences. Similar complaints can be made against Davidson.

Davidson (1997b, p. 20) says that that linguistic capacities are largely innately endowed is not «philosophically significant», although he does think that it is probably true (so much the worse, perhaps, for such `significance'.) Davidson's point appears to be that such `facts' shed no light on the content of our thoughts and how we come by such content, the facts concern only syntax. The motivation here, however, derives from his unabashed «social engineering» view of language rather than any considered view of the empirical data, which, I should say, does not support his position.Foot note 4_12 Even so, does Davidson here offer a principled screen behind which an interpretive conception of meaning may remain unmolested by data or empirical theories?

It is fair to say that in modern discussion the view that language is characterised by novelty, systematicity, and unboundedness comes directly from Chomsky's early pioneering work (1957), although the characterisation goes back at least as far Descartes. The mere existence of such features, of course, does not reveal the falsity of the social view of language.Foot note 4_13 Indeed, for Chomsky in the late `50s psychological issues were very much in the background. It is also fair to say, however, that linguistics of the past thirty years or so has made it beyond informed doubt that linguistic understanding involves sensitivity to information that is not communicated or in any way encoded in the data (input) the hearer or learner is faced with. Nor is this information necessarily syntactic as opposed to semantic, as Davidson's dismissal of the significance of the internalism of syntax would suggest. In fact, the kind of distinction between syntax and semantics in which Davidson trades is problematic.

Syntax is no less conceptual than semantics. That a speaker can recognise that, say, it is the object of meet which is being questioned in Whom does Bill want to meet? (I help out here by using the unfortunately moribund accusative form), presumably requires the concept of transitive verb, and much else besides, just as, I should say, the recognition that Rover is a dog requires the concept of a dog. If there is a fundamental difference here, no-one has made clear what it is. Even if there is a principled separation, it just does not follow that the difference is that semantics is external (social) while syntax is internal (private). (See Chomsky (2000) and Segal (2000) for a wholly narrow view of content.) There is the tendency to think of syntax as meaningless symbols, as we think of `1's and `0's on a Turing machine tape that do not mean anything until we interpret the machine as, say, computing in base 2. Deep confusion arises when we think this way, for it encourages the thought that syntax and semantics are related as representation to represented, as if syntax is `inside' and semantics is `outside'. Patently, however, when we theorise about the syntactic competence of a subject, we do not think that syntax is a meaningless representational medium upon which the subject drapes meaning. If that were the case, a subject would not be in a position to make judgements about grammaticality; contrary to fact, competence would be an automatic mechanical response rather than a resource we can freely employ. Syntax, along with semantics, is represented in the mind, it is not the representational medium of the mind. Syntax and semantics are properly thought of as different levels of information that a competent subject has at his disposal qua competent.Foot note 4_14 Unsurprisingly, where syntax ends and semantics begins is not clear.

One may, of course, think that there is something special, conceptually revealing, about compositionality in distinction to empty categories and rest of the abstract structure revealed by current linguistics, although I know of no argument that even suggests that this is true; and even if it were true, we would still be owed an explanation of the phenomenon. If, then, one constrains a theory of meaning to be compositional simply because linguistic understanding appears to be compositional, by parity of reasoning, one must constrain one's theory of meaning to accommodate empty categories, etc. But if one is so consistent, then the kind of «social engineering» model Davidson favours is in real trouble. It might be, as Davidson (1997b, pp. 25-6; 2000, pp. 71-2) suggests, that we learn concepts through ostensively learning the reference of words, with ourselves forming a triangle with teacher and the world, where the former provides shared saliences and the latter correction. I think not. But it just makes no sense at all that we ostensively learn empty categories, for there is no `word' to learn. The point here is quite different from, say, the situation with inflectional properties. Such properties are phonetically marked, but they have no reference that may be ostended, they simply mark agreement features. Still, we can just imagine, I suppose, some extrapolation from a matrix of previously fixed words (no-one, of course, has a clue how this might be achieved). Empty categories, however, do refer, possessing pronominal and/or anaphoric features. This is crucial. Empty categories are not expletives such as it and there (in some constructions) in English that occur without making a semantic contribution. The presence of expletives in English is due, it seems, to a structural requirement for overt subjects. Empty categories, on the other hand essentially enter into the semantic evaluation of the constructions in which they appear. Empty categories, therefore, cannot be dismissed as mere «syntactic constraints». Let us briefly exhibit this difference with a grammatical chestnut.

In Bill is eager to please, the infinitive to please is transitive, but what are the arguments of the verb? No competent speaker has trouble telling us that Bill is not the object, for it is Bill who wants to be pleasing, even though, qua infinitive, the verb's subject is not marked; in finite paraphrase, Bill is eager that he (Bill) please someone or other. But compare, Bill is easy to please; here Bill must be the object of the verb, with the subject being empty (this becomes apparent with the paraphrase, It is easy to please Bill.) The difference is obvious but subtle still, for it depends on the fact that easy, unlike eager, does not categorise for an external (subject) argument; hence, the subject of the infinitive complement of easy is an empty category PRO, where the subject of eager is the raised subject Bill. In sum, a competent subject must understand the semantic values of empty categories, but no such values can be ostended, there is nothing to triangulate. Again, as far as we can see, the only place for empty categories to be is as represented in the minds of speakers/hearers. To acknowledge this is to admit that language is not «intrinsically social».

We have come some distance from the idea that a truth theory ought to embody a translation constraint; I hope, however, that the dialectic is clear. We can state a truth theory for a speaker (idiolect) in ways that satisfy constraints of novelty, compositionality, etc., and also offer intuitive tests of its correctness in the form of the demand that it entails articulations of the content of the speaker's sentences. None of this involves translation in the substantive sense of a truth theory being constrained to issue in informative translations of `foreign' sentences. The change of tack is a move from a social perspective, one under which a translation constraint makes sense, to a cognitive or internalist perspective, under which translation is otiose. I have attempted to show that staying with the former perspective lands the truth theoretic approach to meaning in an empirical dilemma out of which there is no apparent escape.

6: Concluding Remarks

Davidson's initial presentation of the truth theoretic approach showed the greatest sensitivity to empirical issues and could be profitably understood as continuous with Chomsky's work, as Davidson (1984, pp. 22/30) himself suggested. Davidson's clarity on the demand that a theory is required rather than consultation of dictionaries has justly proved to be seminal. Above all, Davidson's contention that a truth theory is what is needed to account for compositionality and unboundedness is a lasting achievement. As my title intimates, I seek a revision of Davidson, not a refutation. Once we get clear on the relationship between Tarski, a truth theory and deflationism, we can see Davidson as offering a profound insight into truth and meaning, one which is flush with an empirical approach to linguistic understanding.


John Collins