J E R R O L D J. K A T Z
THE THEORY
OF SEMANTIC
REPRESENTATION 1
1. This paper contrastively expounds the theory of semantic representation (henceforth TSR). It sets out the principles of this theory by contrasting them with those of other current philosophical approaches to the logical structure of natural languages. Such approaches are of two kinds. On the one hand, there are "reductive approaches'. These dispense with the notion of sense or meaning, building their theory of the logical structure of natural languages on some philosophically preferred notion. Reductive approaches differ on the basis of the notion chosen as philosophically preferable. The main reductive approaches are the meaning-as-use approach (associated with philosophers like Grice, Strawson, and Searle), the Tarskian approach (associated with Davidson, Wallace, and Harman), the possible worlds approach (associated with Montague, Hintikka, and D. Lewis), and the causal approach (associated with Kripke, Putnam, and Donnellan). I will contrast TSR with each of these approaches. On the other hand, there are approaches based on the notion of sense or meaning. These 'intensional approaches' differ on the basis of how they conceive of senses. The main intensional approaches are those of Frege, Carnap, and TSR. I will contrast TSR with Frege's and Carnap's approaches, too. One motive for contrasting TSR with the other intensional approaches is to make clear that TSR is no more the same approach as Frege's or Carnap's than one reductive approach is the same as another. Two widespread misconceptions about TSR are that it is simply a linguisticized version of Frege's theory of sense, z and that semantic representations in TSR are really equivalent to Carnapian meaning postulates. The first misconception is probably due to the association of TSR with these other intensional approaches. The issue of reductionism, particularly in the form of the conflict between Quine and Carnap, has long been the major issue in the philosophy of language, and TSR entered recently on the intensionalist side, apparently just to provide Frege and Carnap with support from linguistics. The second misconception is probably due to the extensionalist orientation of philosophy of logic and language. This orientation made it difficult for intensionalists themselves Erkenntnis 13 (1978) 63-109. All Rights Reserved Copyright 9 1978 by D. Reidel Publishing Company, Dordrecht, Holland
(34
J E R R O L D J. K A T Z
to see what might be wrong with trying to handle all logical structure in natural language in predicate calculi with appropriate meaning postulates because it made them think in exclusively extensionalist terms about conditions of adequacy. But, even apart from the issues internal to intensionalism, it is important for the controversy between intensionalism and reductionism that TSR be seen as a new and different version of intensionalism. What makes TSR distinct from Frege's and Carnap's approaches overcomes difficulties in them, so that unless this distinctness is recognized intensionalism will continue to be represented in the controversy by its weaker versions. Collectively, the contrasts that I will draw are designed to present a reasonably comprehensive picture of TSR in terms of the stands TSR takes on the major issues facing all approaches to logical structure in natural languages. These stands, I hope to show, are consequences of the systematic elaboration of new ideas about the traditional notion of sense or meaning. I also hope to show that the fundamental ideas of the meaning-as-use approach and the causal approach do not have to lead to stands on these issues that conflict with those TSR takes, and that, when TSR's stands are compared with those of genuinely conflicting approaches, TSR takes the most sensible stands on the widest range of issues. 2. TSR is part of linguistic theory and part of logical theory. This is because it is a theory of senses and it adopts the strict identity of the linguist's senses and the logician's concepts and propositions, a From the linguist's viewpoint, the senses of a sentence constitute its meaning in the language, while, from the (intensionalist) logician's, the senses of a sentence constitute the aspects of its grammatical structure that determine the propositions it expresses or its logical form. TSR claims that these are two viewpoints on the same thing. This claim enables TSR to claim further that the synonymy relation provides identity conditions for propositions, that the semantic properties and relations of sentences provide grammatical evidence for hypotheses about the structure of propositions, and that the logical relations between propositions can be used to provide non-grammatical conditions on hypotheses about the structure of senses. TSR thus takes senses to be legitimate objects of scientific study. The
T H E T H E O R Y OF S E M A N T I C R E P R E S E N T A T I O N
65
science of language, linguistics, has, among others, the aim of describing the structure of senses and their relation to sentences in natural languages. Therefore, TSR is thoroughly anti-Quinean - t h o u g h it does not necessarily involve the reification of senses. ~ The reason why Quin,e's approach has been omitted in this paper, even though it is a more articulated reductivist approach than those considered, is both because his objections to senses are ones that I have responded to on a number of previous occasions 5 and because his philosophically preferred notions, namely, those of Skinnerian reinforcement theory, are generally acknowledged, since Chomsky's criticism of this theory, to be inadequate as a basis for the semantics of natural languages. Briefly, my response to Quine was that his objections rest on the now refuted taxonomic theory of grammar. These objections take their strongest form in the thesis of the indeterminacy of translation (i.e., there is no synonymy relation to provide a condition of identity for propositions). But this thesis turns out to hold only in the taxonomic theory where there is no semantic level of representation because such a level violates its anti-mentalistic methodology. Indeterminacy of translation is just an artifact of taxonomic theory's inability to impose tight enough constraints on the representation of sentences to secure translation. 6 In the description of TSR up to this point, there is no indication of a divergence from Frege's approach. Both approaches have a common anti-Quinean case. Divergence appears, and it is the most significant difference between TSR and Frege's theory, when we ask how each conceives of senses. Frege says: "I would like to call the s e n s e of a sign, wherein the mode of presentation [of the reference] is contained". 7 The same reference is presented in 'the evening star' and 'the morning star' but their senses are different, the former presents the planet as a star appearing in the morning and the latter as one appearing in the evening. The difference in sense lies in how what they refer to is to be picked out. Thus, although Frege's theory is widely cited as the prime example of a theory of sense, it conceives of senses derivatively. Senses are thought of as instructions for determining the referent of an expression or the truth value of a sentence. This seems clear from the way Frege makes his case for the existence of senses in terms of their role in explaining the informativeness of identity statements and in accounting for the reference of expressions in opaque contexts. 8 It also seems as if Frege took this position
66
J E R R O L D J. K A T Z
on sense in part as a reaction to philosophers of logic, like Husserl, whom he thought had got sidetracked into a preoccupation with sense and thereby forgot about truth. TSR, on the other hand, conceives of senses as primary and susceptible of independent theoretical characterization. TSR sets up such a characterization as follows. It takes senses to be what ordinary speakers of a language refer to when they say such things as that 'brother' and 'male sibling' have a sense in common, or when they argue with one another about whether a word like 'liquidate' has more than one sense, or make fun of someone for using 'free gift' because 'gift' already has the sense of 'free'. Since senses in this sense are objects of study in the linguistic description of natural languages, TSR takes the form of a theory about the structure of these objects which is built up on the basis of such judgments from speakers. The justification for introducing senses into the theory of language, then, is quite different from Frege's. Frege construes the sense a speaker attaches to an expression from the criterion he or she uses to pick out its referent, and so he justifies positing senses on the basis of evidence about such language-to-world relations. TSR takes the attitude that positing senses is the way to construct the best theory of synonymy, ambiguity, redundancy, and other properties and relations of the same kind. 9 Hence, TSR's justification derives from properties and relations internal to the language. (TSR's account of the relation between sense and reference, too, is based on such internal properties and relations. 1~ This difference in the way senses are conceived has significant consequences. One is that the difficulties found with Frege's approach in connection with proper names does not arise on TSR. On Frege's approach, if speakers of a language pick out the referent of a word on the basis of an instruction for applying the word to objects in the world, this instruction is ipso facto the sense of the word in the language or in the idiolects of the speakers who employ the particular instruction. Since speakers pick out the referents of proper nouns on the basis of an instruction to look for some identifying property, the Fregean approach is committed to the claim that proper nouns have a sense like common nouns. Frege, of course, noted that, unlike common nouns such as 'brother' or 'chair', instructions may differ widely from speaker to speaker about what the identifying property, and hence the sense, for a particula.r proper noun is. He observed that the sense of a name like 'Aristotle'
THE THEORY
OF S E M A N T I C
REPRESENTATION
67
9 might, for instance, be taken to be the following: the pupil of Plato and teacher of Alexander the Great. Anybody who does this will attach another sense to the sentence 'Aristotle was born in Stagira' than will a man who takes the sense of the name: the teacher of Alexander the Great who was born in Stagira.11
Frege's views on sense were adopted by a number of recent philosophers. C. I. Lewis used them in connection with an attempt to explicate the concept of necessity, and this version of them was subsequently criticized in papers by Donnellan and Putnam which became the link between Wittgenstein's criticisms of Frege and Kripke's influential criticism of current Fregean semantics. TM Wittgenstein pointed out that on Frege's view a sentence with a proper name as its subject can mean various things, but that when someone asserts 'Moses did not exist' he or she is not always ready to substitute some o n e of these things for the name. Wittgenstein comments: I shall perhaps say: by 'Moses' I understand the man who did what the Bible relates of Moses, or at any rate, a good deal of it. But how much? Have I decided how much must be proved false for me to give up my proposition as false? Has the name 'Moses' got a fixed and unequivocal use for me in possible cases? xa As the recent philosophical literature tells us, this move led to Searle's sacrifice of a sharp sense-reference distinction. 14 Searle's account of proper names consists in the answers 'Enough', 'Yes', and ' N o ' , respectively, to Wittgenstein's three questions. But this sacrifice only delays the defeat of the Fregean view one move: it comes with Kripke's G r d e l Schmidt-mate. 1~ According to TSR, Frege's is a losing position from the start. Searle found it easy and natural to jettison the sense-reference distinction because a full-blooded notion of sense played so small a role in Frege's theory of logical structure. The weakness in Frege's position, according to TSR, was that it launched an attack on the problem'of reference without developing, independently, a sophisticated theory of meaning. With no such theory to fall back on, the Fregean position is trapped by the fact that it is possible that we use ' G r d e l ' referring to G6del even though Schmidt is really the man who discovered the incompleteness of arithmetic. TSR plays a different variation of the intensionalist game. It first develops a theory of meaning, putting off the problem of reference until it can be viewed from the perspective of such a theory. As we indicated, TSR takes the same theoretical attitude toward senses that any science takes
68
JERROLD
J. K A T Z
toward their object of study: it takes senses to be whatever we must postulate in order to predict and explain the semantic properties and relations of sentences in natural languages. 16 Prediction and explanation, as used in this connection, are also to be explicated within the framework of the developing theory of meaning. The development of the theory begins with initially clear cases of semantic properties and relations like 'meaningfulness', 'sameness of meaning', 'multiplicity of meaning', and 'redundancy of meaning', proceeding to determine the full range of prediction and explanation as other grammatical properties and relations can be shown either to fall under the principles of the developing theory or to be excluded by them. 17 Guided by the familiar canons of scientific methodology, this case of theory construction can be expected to eventually zero in on the truth about its object of study. TSR thus introduces an entirely new factor into the controversy between intensionalist and reductivist approaches. It will be recalled that the force of Donnellan and Putnam's criticisms of C. I. Lewis's Fregean position is that purely referential criteria for applying words provides no way to separate information that is part of the meaning of a word from all sorts of information that is not but for contingent (or logical) reasons could serve just as well for applying it. 18 The new factor is that development of a theory of meaning provides a means of separating these two kinds of information. The development of such a theory will characterize hypotheses that predict an expression's synonymy relations, whether it is meaningful, its degree of ambiguity, and so on. Thus, it will be possible to say that a feature F is part of the meaning of an expression just in case the simplest such hypothesis takes F to be part of the meaning of the expression. What separates semantic information from other information used in applying words is that only semantic information plays a role in the system of meanings internal to the language (so only this information provides a basis for predicting semantic properties and relations). The series of moves that leads to Searle's sacrifice and then Kripke's G6del-Schmidt-mate does not occur in the TSR variation. There is a great deal of evidence that proper nouns (in natural languages) do not have semantic properties and relations (except, of course, for the property of meaninglessness). ~9 A proper noun like 'Mary' is the name of indefinitely many people, cats, dogs, shops, planes, and other artifacts, but it would be absurd to call 'Mary' ambiguous. Similarly, 'Mark Twain' and 'Samuel
T H E T H E O R Y OF S E M A N T I C R E P R E S E N T A T I O N
69
Clemens' name the same person but it is equally absurd to call them synonymous. With the means TSR provides for determining the semantic information in a sense, this evidence becomes the basis for an argument that the simplest explanation of this absence of semantic properties and relations is that proper names lack a sense. Another point at which TSR diverges significantly from Frege's approach is in the very different attitude it takes toward natural languages and toward their relation to theories of logical structure. TSR does not share Frege's critical attitude toward natural languages, and consequently, it does not share his desire to construct a logically perfect language, or conceptual notation, 2~ to serve as a better tool for science. The attitude of TSR toward natural languages is more like Wittgenstein's,2t though without Wittgenstein's doctrine about meaning and use or his suspicion of theory. Frege and his followers 22 take features of language like ambiguity and the failure of grammatically well-formed expressions to have a referent to be 'imperfections' ("So long as the reference remains the same such variations of sense may be tolerated, although t h e y . . , ought not occur in a perfect language" 23). TSR takes such features to be phenomena for a theory of language to explain instead of flaws in a language that motivate the attempt to invent a flawless one. TSR claims that such an attempt is hopeless because such things as ambiguity and referential failure are essential to anything that can function as a genuine language. Referential failure is a consequence of the fact that the formal sentence generating rules cannot have available the complete knowledge of the world necessary to make sure that every grammatically well-formed expression designates an object. Since we can have no such complete knowledge, we would have to use the very incomplete knowledge available to us to limit the operation of sentence generating rules. If we thus limit the operation of these rules to the point where we guarantee no risk of their producing grammatically well-formed sentences that fail to designate, we also guarantee that there will be little chance of our acquiring further knowledge, since the constrained rules will not produce sentences expressing hypotheses about what we do not as yet know. Sentence generating rules with the power to assign correct descriptions of syntactic structure automatically assign non-equivalent phrase markers to some strings of morphemes. Thus, ambiguities arising from the
70
J E R R O L D J. K A T Z
coincidence of two or more syntactic structures in the same string, as in 'old men and women', are a reflex of structural facts such as that nouns like 'men' and 'women' and expressions formed from them like 'men and women' belong to the same recursive category of nouns and hence take the same class of modifiers. Sentence generating rules without the potential to produce ambiguity would also lack the potential to produce the diverse range of syntactic structures necessary for the free expression of a speaker's thoughts that is essential to language. 24 A logically perfect language would be a perfectly illogical language. The search for the utopia of a logically perfect language was based on the failure to make two important distinctions. One was the methodological distinction Chomsky has drawn between (a theory of) a language and (a theory of) how speakers use a language, z5 The other was the semantic distinction between languages and theories. 26 Once we make these distinctions, the motivation for the search on the part of Frege and his followers, namely, their concern lest the unwary reasoner be misled by a referential failure or an ambiguity, makes no sense. Then, the possibility of fallacies stemming from a plunge into a truth value gap or a wrong turn at an ambiguity can no more be considered a reason for claiming that such gaps and ambiguities are imperfections in a language than the threat to hikers from gorges and forks in the road can be considered a reason for claiming that such features are imperfections in the terrain. The careless reasoner, like the unwary hiker, bears responsibility: reasoning and hiking are matters of performance. There is no more a prospect of a perfectly safe language than there is a prospect of perfectly safe terrain. But, given the distinction between languages and theories, there is a prospect of a correct theory of a language, just as there is a prospect of a correct geographical description of a terrain. In such a theory, presuppositions about reference and ambiguity have the status of features of the grammatical terrain like rhyme, and well-formedness. Their occurrence is to be predicted and their nature explained. Hence, whereas Frege sought a logically perfect language, a conceptual notation for expressing the propositions of science, TSR seeks an empirically optimal theory, a representation scheme and a set of principles for explaining the logical structure of propositions in terms of the grammatical structure of the sentences. Frege's example of a conceptual notation served as a model for the
THE T H E O R Y OF S E M A N T I C R E P R E S E N T A T I O N
71
artificial languages that Carnap and other logical empiricists constructed as 'rational reconstructions' or 'explications'. As a consequence, programs of rational reconstruction have inherited the mistaken assumptions of the model and are plagued with the problem of there being no satisfactory constraints in the construction of artificial languages. A logical positivist is free to set one up to vindicate the thesis that metaphysics is nonsense but, then, the metaphysician is free to set one up to vindicate metaphysics. The reason no constraints are tight enough to decide between competing rational reconstructions is that explication is free to depart from the meaning of a term in natural langizage in order to perfect it. But the perfection of concepts is no better off than the perfection of languages. To a logical empiricist, translatability into an empiricist language looks like a better concept of 'being cognitively meaningful', though to a rationalist expressibility in the a priori system of semantic primitives will seem better. 27 TSR abandons the entire enterprise of rational reconstruction modeled on Frege's notion of a logically perfect language because of its inherent circularity: the enterprise is supposed to settle philosophical questions (like the meaningfulness of metaphysics) but it rests on a form of analysis (explication) that selects preferred analysans (e.g., translability into an empiricist language 2a) on the basis of the explicator's (i.e., the logical empiricist's) own notion of perfection. In place of this enterprise, TSR puts the construction of an empirically adequate theory of meaning in natural language. This does away with the circularity of rational reconstructions, and as a consequence, supplies the missing constraints on semantic analyses. The constraints absent in rational reconstruction are present in a linguistic theory of meaning in the form of conditions on the empirical and systematic adequacy of semantic analyses of sentences. The predictions of the semantic properties and relations of sentences that follow from the representation scheme must conform with fluent speaker's judgments about the sentences; the set of principles must explain the nature of such properties and relations across natural languages; the representation scheme must be the simplest possible; the scheme and the set of principles must fit in with the rest of grammatical theory and with the theory of logical deduction; 29 and so on. A semantic theory of natural language in this sense provides a basis for evaluating (as far as semantic analysis can) competing philosophical positions (e.g.,
72
JERROLD
J. K A T Z
empiricism and rationalism) that is neutral between the positions under evaluation. 3. In the fifties and early sixties, Carnap became the principal proponent of the Fregean approach (though not as close a disciple of Frege's as Church). Carnap's most important contribution to this approach was his extension of Frege's account of quantificational structure to cover the contribution of the entire vocabulary of a language to the logical structure of its sentences. This extension is his theory of meaning postulates. 3~ This theory adapts the postulational account of implication in predicate logic to the treatment of implications based on the internal logical structure of the nouns, verbs, etc. that comprise the terms and predicates of a language. Thus, the standard first order predicate logic account of implications like that between (1) and (2) became the model to which the formalization of implications like that between (1) and (3) are assimilated. (1) (2) (3)
Jack is a bachelor Jack is a bachelor or a fool Jack is male
The implementation of this extension was extremely simple and direct. Standard first order predicate calculi were augmented in two ways: their vocabulary was supplemented with sets of individual constants and predicate constants, and their postulates were supplemented with a set of meaning postulates. These meaning postulates were set up in analogy to the logical postulates of the calculus, and aside from the new individual and predicate constants, meaning postulates used only the logical apparatus of the familiar predicate calculi. In these augmented calculi, implications resting on the so-called extra-logical vocabulary as well as those resting on the logical vocabulary come out as logical truths. From the Fregean viewpoint, the virtue of Carnap's theory of meaning postulates is that it preserves Frege's conception of logical structure and implication while bringing the entire extralogical vocabulary of a language under its scope. Frege had achieved the formulation of a univocal notion of logical structure and implication by means of what can only be described as an arbitrary stipulation. He collapses the distinction between analytic implication (in the sense of Kant's concept of analyticity) and logical implication (in the sense of first order predicate logic) by taking the notion
THE T H E O R Y OF S E M A N T I C R E P R E S E N T A T I O N
73
of implications of the former kind to be simply a special case of his own concept of logical truth, namely, truths provable exclusively on the basis of the laws of logic, without assumptions from 'the sphere of some special science'. 31 Frege makes this substitution without comment and without justification, and to make matters worse, he calls his own concept 'analyticity', thereby confusing the situation down to the present. Carnap's theory of meaning postulates puts on the 'finishing touches' by developing a formal account of analytic implication which is the same as that of logical implication. From the viewpoint of TSR, this virtue of Carnap's theory is its main vice. Thus, another important difference between Frege's and Carnap's approach and TSR is that TSR explicitly contradicts this aspect of Frege's conception of logical structure and implication. TSR claims that, in natural language, the logical structure of sentences should be represented on the basis of an essentially Kantian concept of analyticity (suitably revised and extended32), while the deductive relations between sentences which outstrip the apparatus for representing analytic implication should be represented on the basis of apparatus for representing what is essentially the Fregean concept of logical truth. (This, of course, allows for cases of logical truth which require to mark them as logical truths the apparatus for representing analytic implication, e.g., "John is a bachelor or John is not an adult unmarried human male".) Carnap's approach attempts to make the theory of deduction serve also as a theory of definition; TSR constructs a new theory which explains definition independently of deduction. Another way of contrasting Carnap's approach with TSR is to note that 'meaning postulate' is a misnomer because rather than describing the meaning of expressions such a postulate states a relation between the extensions of expressions. The theory of deduction becomes a theory of definition because the meaning postulates allegedly specify the definitional constraints on the admissible models, and so the notions of analytic implication and synonymy are collapsed into the broader notions of L-implication and L-equivalence, which is specified in terms of validity in all assignments of extensions in which the relations expressed by both the logical and meaning postulates hold. In contrast, semantic representations describe the meaning of expressions, thereby providing an independent account of definition. This difference makes it possible to formulate
74
J E R R O L D J. K A T Z
arguments against Carnap's approach that exploit the features of extensional relations that makes the apparatus for representing them particularly unsuited to representing senses. One such argument arises in connection with expressions like 'the least convergent series', 'the round square', 'the consistent and complete first order formalization of arithmetic', etc. Such expressions are meaningful. Therefore, we shall require some account in terms of meaning postulates. On an adequate such account, 'the least convergent series', 'the round square', 'the consistent and complete first order formalization of arithmetic', etc. come out as synonymous because such expressions are provably L-equivalent (which is the strongest form of definition equivalence in this approach). But this consequence is obviously false, aa TSR avoids such absurdities because, in having apparatus for describing senses, it can distinguish the senses of such expressions in terms of their compositional meaning, and the condition for synonymy can be sameness of compositional meaning. A compositional notion of sense requires that a theory of sense mesh with a theory of the syntactic structure of sentences. This is necessary because otherwise there is neither an account of the senses that combine to form the sense of a syntactically complex expression or sentence nor an account of the syntactic relations that indicate which elementary senses combine. Thus, from the start, TSR has been set out as part of the theory of generative grammar, in Chomsky's sense. The choice of Chomsky's theory of grammar, although obviously not the only possibility, was dictated by the belief that only a transformational theory reveals enough syntactic structure for a full description of semantic structure. As part of the theory of generative grammar, TSR takes the form of a theory of the representation of sentences at the semantic level of grammars. Such a theory specifies three things:
(h) (B) (c)
What senses, if any, the sentences of a natural language have, and what the logical structure of senses is. Definitions of semantic properties and relations. What the compositional mechanism is by virtue of which the meanings of sentences are a function of the meanings of their parts.
THE THEORY OF SEMANTIC R E P R E S E N T A T I O N
75
The specification of(C) is a projection rule that establishes a correspondence between representations at the semantic level and representations at the syntactic and phonological levels. The specification of (B) takes the form of generalizations about the formal structure of semantic representations, correlating particular formal structures with particular semantic properties and relations. The specification of (A) is a formal vocabulary and a syntax for semantic representations. The fundamental feature of these representations is their analytic or decompositional nature. This is a consequence of the fact that they are required to predict redundancies like 'left-handed southpaws' and analyticities like 'Nightmares are dreams'. Since this means that we must design the semantic representation of the head construction so that the semantic representation of the attributive appears as a proper part, such semantic representations analyze or decompose the sense of syntactic atoms like 'southpaws' or 'nightmares'. Thus, TSR makes a definitional/non-definitional distinction that is far narrower than the distinction allowed by the conceptual apparatus in a theory of deduction. Informally, we may make the point as follows: TSR can coun,~ the property of being male as a definitional property of'bachelor' without also having to count the property of being either a male or a fool as such a property because not every logically necessary property is ascribed to its sense by its semantic representation under the projection function (only those analytically part of it, in the narrow Kantian sense). In the theory of deduction, doing double duty as a theory of meaning, both such properties count as definitional (if either does) - in contrast, for example, to counting the property of being a fool as a definitional property of 'bachelor'. Both (2) and (3) are consequences of (1) in the customary Carnapian systems. Each is inferable from (1) using the general logical truths formalized therein, without assuming the sort of proposition that is needed to prove (4). (4)
Jack is a fool.
The critical issue is whether there is any compelling reason to go beyond the simpler Carnapian systems to handle sense in natural language. The proponent of TSR thinks there is the same reason to go beyond such predicate calculi as there was to go beyond sentential calculi. 8~ To see that TSR is an advance over predicate calculi of the same order as the
76
JERROLD
J. K A T Z
improvement of the latter over sentential calculi, let us imagine the counterpart of Carnap, someone we will call 'Panrac', who, in the early days of predicate logic, wishes to hang onto the conception of logical structure in sentential calculi. Panrac is faced with a band of Young Turks who insist that their more complex system with newfangled apparatus of quantifiers, individual variables, and predicate variables is nonetheless preferable to sentential calculi. Their claim is that such complications are necessary to handle not only the implications that rest on the inter-sentence relations in sentential co-ordination but implications that rest on the intra-sentence relations between quantifiers and variables. The Young Turks produce cases of inferences that they claim are essentially different from those that rest on such inter-sentential relations and cannot be handled in sentential calculi, for example, the inference from (5) to (6).
(5) (6)
All horses are animals All heads of horses are heads of animals
Let us suppose Panrac to be more of a diehard and more resourceful than any actual defender of sentential calculus would have been. In particular, we may suppose that he hits on the idea of using meaning postulates to show that the cases like (5) and (6) produced as counter-examples are, in fact, not beyond the scope of sentential calculi. As Carnap did in the case of predicate calculus, Panrac augments standard sentential calculi by adding a set of sentential constants $1, $2 . . . . to their vocabulary and by adding meaning postulates like (7) to the axioms. 85 $1, $2. . . . (7)
$5 -+ $6
can be thought of as belonging to the same category as the symbols 'T' and 'F' in some versions of sentential calculus, but $1, $2. . . . stand for particular sentences of the natural language (say, 'S~' stands for the ith sentence generated in some designated grammar of the language). Panracian meaning postulates are taken as expressing necessary connections between the extensions of sentences in analogy to the way that Carnapian meaning postulates express necessary connections between the extensions of predicates. Panrac thus confronts the Young Turks with as much justification for his position as Frege and Carnap have for theirs. Panrac can handle
THE THEORY
OF S E M A N T I C
REPRESENTATION
77
examples like the inference from (5) to (6) using an augmented sentential calculus containing (7), where '$5' stands for (5) and 'Sd stands for (6). He, too, can claim that his theory has the virtue of preserving the familiar conception of logical structure and inference while bringing the entire range of quantificational phenomena under its scope. Let us look at what makes Panrac's justification fail, since, given the symmetry of the cases, this will also be what makes Carnap's fail. A Panracian account of an inference like (8) or (9) (8) (9)
Panrac, if anyone, can save sentential calculus theory, but even he can't do this No one can save sentential calculus theory
states the brute fact of a valid implication, but it is adhoc because it sheds no light whatever on the source of its validity. It fails to explain the intrasentential connections responsible for its validity: the relation between the quantifier "anyone" and the proper and pronominal names in (8). This, moreover, is a failure that is characteristic of the way that Panrac's theory treats str inferences, since the theory, in principle, makes no distinction between validity in virtue of the form of inter-sentential connection and validity in virtue of intra-sentential quantificational relations. Parallel to this failure to make an explanatory distinction is Carnap's failure to distinguish between validity in virtue of intra-constant logical structure and validity in virtue of inter-constant logical structure. This is a consequence of the fact that Carnap's approach, in collapsing understanding and inference, treats the logical structure internal to individual and predicate constants in the same way that it treats the logical structure external to them, namely, in terms of deduction with inference licensing laws. 36 Hence, just as Panrac will treat inferences like that from (5) to (6) or from (8) to (9) as if they were inferences like (10) and (11) to (12), (10) (1 I) (12)
If it has rained, the streets are wet It has rained The streets are wet
Carnap will treat inferences like that from (1) to (3) as if they were inferences like (13) to (14). (13) (14)
Sue is smarter than Moe and Moe is smarter than Lem Sue is smarter than Lem
78
J E R R O L D J. K A T Z
Carnap marks the validity of both kinds of inference in the same way. In both, the conclusion is shown to follow from the premiss and an appropriate meaning postulate, (15) in the former case and (16) in the latter, (15) (16)
(x)(bachelorx ~ malex) (x)(y)(z)((smarter thanx,~ & smarter than~.~) ~ thanx.~)
smarter
by an application of modus ponens. But this misses the difference between these inferences as much as Panrac's account of cases like (8) to (9) and cases like (I0) and (11) to (12) misses the difference between them. In the case of analytic implications like (1) to (3), the logical form of the conclusion is included in the logical form of the premiss because of the intraconstant logical structure of the predicates in them. Thus, (3) follows from (1) without a special postulate like (15), just as (6) follows from (5) without a special postulate like (7), and without an application of a law of logic like modus ponens. All that is required is a definition under (B) of semantic entailment (validity by virtue of inclusion of logical forms), a7 Such a definition expresses a condition on the intra-constant logical structure of a premiss and its conclusion, as reflected in their semantic representations, which picks out cases where the logical form of the latter is included in that of the former. It is obvious from inspection that the inference from (13) to (14) is not of this sort. There is nothing in the logical form of the premiss (13) that corresponds to the conclusion (14). The premiss (13) is a conjunction whose first conjunct compares the intelligence of Sue and Moe and whose second, and only other conjunct, compares the intelligence of Moe and Lem. No part of the logical form of (13) compares the intelligence of Sue and Lem. In fact, the individual constants 'Sue' and 'Lem' do not appear as coarguments of any predicate in (13). To reach the point where there is a statement in which these constants appear as co-arguments of a predicate, it is necessary to apply a substitutional principle to (16). But even upon such substitution, the conclusion (14) does not follow because it is still necessary to detach the consequent of the statement resulting from the substitution and (13). So, here, there are clearly essential appeals to laws of logic to establish a logical connection not found in the intra-constant logical structure. The criticism of Carnap's approach is, then, that it fails to explain the
T H E T H E O R Y OF S E M A N T I C R E P R E S E N T A T I O N
79
source of the validity in analytic implications. This and Panrac's failure to explain the source of the validity of quantificational implications correspond to the failure Chomsky noted in connection with phrase structure grammars. 38 Chomsky pointed out that, although these grammars were adequate in what he called 'weak generative capacity', i.e., they could generate the appropriate set of sentences, they were not adequate in what he called 'strong generative capacity', i.e., they could not generate the proper syntactic descriptions for sentences. At the semantic level, weak and strong generative capacity concern, respectively, the ability of the system to mark each valid inference as such and the ability of the system to mark their validity on the basis of descriptions of their premisses,and conclusion that properly explains the source of their validity. Carnap's approach advocates theories of the semantics of natural languages that are weakly but not strongly adequate. Since TSR advocates separate theories of semantic representation and logical deduction, in which the former concerns the relation of analytic implication and other semantic properties and relations while the latter concerns valid inference, there is no barrier to the theories that TSR advocates being strongly adequate. 39
4. We turn now to the reductive approaches. The first to be considered is the possible worlds approach. The philosophically preferred notion for those who take this approach is that of the extension of an expression or sentence in a possible world. The attractiveness of this notion has, I think, two sources. One is the belief that recent advances in modal logic have cleared the notions of possibility and necessity of the charges made against them by Quine, 4~ Goodman, 41 and others. The other source is the appeal of referential semantics in some form to those who view logic too exclusively in model-theoretic terms. The belief that the philosophical reputation of possibility and necessity can be rescued in so easy a way is mistaken. Quine, Goodman, and the others who attacked meaning and analyticity took the notion of possibility to belong to the same family of concepts. Actually, Goodman adopted a stronger position, claiming that the question of when two predicates apply to the same possible entities is the same as the question of when two predicates have the same meaning. This stronger position is false, since two predicates like 'is the number two' and 'is the even prime' can differ
80
JERROLD J. KATZ
in meaning without one applying to a possible that the other does not apply to. Be this as it may, the basic point of their criticism, that the notion of possibility belongs to the family of meaning terms is correct. We might express this by saying that the notion of possibility is a 'poor relation': possibility (necessity, etc.) depend on the other notions in this family but they do not depend on it. The argument for this is straightforward. The notion of a possible world is that of a state of affairs described in a consistent expression. Now, consistency is partly determined by meaning (e.g., 'a mortal who will live forever,' 'a married bachelor', etc.), and partly determined by considerations beyond meaning (e.g., 'a complete and consistent first order formalization of arithmetic'). Therefore, although the notion of possibility depends on that of meaning, it depends on other things as well. This being so, it is misguided for possible world theorists to promote their theory as superseding the theory of meaning. If, as Hintikka claims, the "...whole concept of meaning (as distinguished from reference) is very unclear and usually hard to fathom", 42 then, without doubt, the whole concept of possibility and necessity is very unclear and usually hard to fathom. The other source of the attractiveness of the notion of possible worlds to possible worlds theorists is that this notion provides a way of saving referential semantics once the facts force abandonment of the simple referential theory of meaning. This theory identifies the meaning of an expression with what it refers to, e.g., the meaning of 'Jimmy Carter' is the president of the United States at present, the meaning of 'politician' is the collection of politicians, and the meaning of 'Jimmy Carter is a politician' is the True. In this theory, semantic properties and relations are defined as illustrated in 07) and (18). (17) (18)
An expression E has a meaning (is meaningful) just in case E refers to something (has a non-null extension). Two expressions E~ and E~ have the same meaning (are synonymous) just in ease E~ and Ej refer to the same thing (have the same extension).
Abandonment of this theory is forced by meaningful expressions like 'ghost', 'the golden mountain', and 'a projectile traveling faster than the speed of light'. Since they fail to refer, they are counter-examples to (17). Counter-examples to (18) are non-synonymous expressions like "creature
THE THEORY OF SEMANTIC REPRESENTATION
81
with a heart' and 'creature with a kidney' which have the same extension. Given an initial commitment to referential semantics, it will seem completely obvious that the cause of the trouble is the simple referential theory's restriction of extension to the actual world. Possible worlds like that described in Hamlet contain ghosts. Thus, to save referential semantics, it will seem obvious that the step to take to overcome these counterexamples is to identify meaning with extension in possible worlds, or what amounts to much the same thing, to say that the meaning of an expression is a function from possible worlds onto the expression's extension in them. Thus, (17) and (18) are replaced by (19) and (20) (19)
(20)
The expression E has a meaning (is meaningful) just in case E refers to something in some possible world (there is a possible world in which E has a non-null extension). Two expressions E, and Ej have the same meaning (are synonymous) just in case they refer to the same thing (have the same extension in every possible world).
to which none of the counter-examples to (17) and (18) apply. But, as we said above, semantic properties and relations do not rest on the notion of possibility as (19) and (20) claim. The problem is that only part of this notion has to do with semantic structure, other parts have to do with logical and mathematical structure. Consequently, such structures can preclude a meaningful expression from referring to something in any possible world and make non-synonymous expressions refer to the same thing in every possible world~ Counter-examples of this sort to (19) are 'a consistent division by zero', 'the largest natural number', and 'something not in any possible world'; counter-examples of this sort to (20) are 'the even prime' and 'the number two' or 'two plus two is four' and 'a consistent first order formalization of arithmetic is incomplete'. Some possible world theorists have departed from their advertised position that referential considerations of the sort embodied in (19) and (20) are all that is required to define the semantic properties and relations necessary to a theory of natural languages. ~8 The step they have taken is to take the syntactic structure of expressions into account in defining semantic properties and relations. In this they are following the idea inherent in Carnap's proposal about intensional isomorphism. 44 Such a step, although a step beyond purely referential considerations, can hardly
82
JERROLD
J'. K A T Z
be in the right direction, since the counter-examples (such as those cited in the previous paragraph) do not arise because of a neglect of syntactic structure. The attractiveness of this step comes from three things: first, it offers alternatives to (19) and (20), second, given the wide range of options in syntactic description, it provides considerable room for the possible world theorist to maneuver, and third, it is fully acceptable to those who for Quinean or other reasons have no wish to go so far as to embrace meanings. We examine this step in order to show that it is necessary to go beyond it by introducing meanings. Carnap proposed the relation he called intensional isomorphism as an explication of the intuitive notion of synonymy and as a criterion for identity of belief. He characterized the relation relative to an applied predicate calculus (roughly) as in (21). (21)
Two expressions E~ and Ej are intensionally isomorphic relative to the predicate calculus C just in case E~ and Ej have the same syntactic structure under the formation rules of C, each individual constant in E~ and Ej corresponds to an L-equivalent individual constant in the other, and each predicate constant in E~ and Ej corresponds to an L-equivalent predicate constant in the other.
Recent logicians, principally Montague (with the notion of an analysis tree) and D. Lewis (with his notion of a s~mantically interpretedphrase marker) have extended Carnap's proposal by sophisticating the conception of syntactic structure. 45 Both replace the exceedingly simple, clearly inadequate, formation rules of the artificial languages used to state predicate logic with the more complex syntactic rules of a serious theory of natural languages. Montague tries to construct his own theory, while D. Lewis takes Chomskyian transformational systems as the optimal English grammar. But such extensions, even if we grant that they employ the most powerful theory of English syntax, do no more than permit cases like 'The cow kicked the stool over' and 'The stool was kicked over by the cow' to count as intensionally isomorphic. They fail to touch the heart of the problem because they only concern syntax. The real problem is semantic. Church anticipated it some years ago. 46 He pointed out that natural languages contain indefinitely many synonymous expressions that
THE T H E O R Y OF S E M A N T I C R E P R E S E N T A T I O N
83
are primitives at the level of syntax, such as 'perhaps' and 'maybe'. (In fact, all the point requires is that at least one member of such synonymy sets be syntactically primitive.) Given such cases, the examples cited above as counter-examples to (20) refute Carnap's proposal and any extension, no matter how sophisticated its account of syntax. L-equivalence is too coarse a grid to separate out these genuine synonyms from the infinite range of examples like those cited above. 47 Consequently, these proposals also fail as criteria for identity of belief, since, people who believe that two and two are four do not ipso facto believe that consistent first order formalizations of arithmetic are incomplete, though people who have the belief expressed by 'Perhaps war will come' ipso facto have the belief expressed by 'Maybe war will come'. 48 TSR is predicated on the claim that referential considerations and syntactic considerations, no matter how powerful, do not suffice to define the semantic properties and relations of sentences in natural languages and thus to provide the further concepts needed to understand the logical structure of such languages. TSR's approach of postulating senses and constructing a theory of them that specifies (A), (B), and (C) makes the theory of reference depend on the ~theory of meaning, although, as we shall see below, not in the simplistic manner of Frege's approach. 5. The next contrasts require us to say more about the choice between different hypotheses concerning the meaning of a sentence. Since, as we indicated earlier, the best hypothesis is the one that predicts the widest range of semantic properties and relations with the simplest means, we must now say enough about semantic representations for the reader to see how they predict and how they can be compared in something like the standard way in linguistics of counting symbols. 4~ This means that we have to say enough to make their normal form explicit. Above we described the way in which semantic representations that can predict redundancy and analyticity will analyze or decompose senses into their component concepts. The logically independent components of a sense are represented by semantic markers; a semantic representation is a set of semantic markers. The logical dependencies among component concepts are represented by the formal relations in the structure of semantic markers. These are labelled trees, so that simplicity measures familiar in syntax will be applicable here.
84
JERROLD J. KATZ
The formal character and empirical interpretation of semantic representation using semantic markers has been discussed extensively elsewhere,s~ The main points are these. The formal structure of the tree is to be isomorphic to the conceptual structure of the sense it represents. The complexity of the branching represents the complexity of the conceptual structure. The root or topmost node bears the label that determines the base concept, that is, the one that determines the semantic category of the whole sense. A minimal branch connects a node with one it immediately dominates. Each such branch represents the application of a predicate function, that is, an operation that converts a less specific predicate (or individual) constant into a more specific one of the same semantic category. Consider the semantic marker (22) (22)
/ [NP, ~(Activity ? ) S ] )
(Physical)
I
(Movement)
I I
(Purpose)
((Catch)tNP,S])(xV;,
(Rate)
((Exceeds average using same mode of transportation)
[NP, s] 1 ?) /
which represents the sense of the verb 'chase'. The base ~concept is that of an activity and the branch connecting the node labeled with '(Physical)' to the root node converts this activity concept into the concept of a physical activity. This distinguishes the meaning of 'chase' from that of 'think', 'remember', 'plan', and other mental activity verbs. The occurrence of variables, called 'categorized variables', represents the fact that the sense of 'chase' is a predicate concept and the presence of exactly two distinct categorized variables represents the fact that this predicate concept is two placed. (22) describes the structure of a sense, but in and of itself it says nothing
THE THEORY OF S E M A N T I C R E P R E S E N T A T I O N
85
about the English verb "chase'. To have what (22) says about the structure of a sense be a claim about the semantic structure of this English verb, we must have a way of saying that the sense it represents is a sense of 'chase'. In a theory of the semantic level of a grammar, this means having a way of assigning (22) to the syntactic and phonological representation of 'chase'. For syntactic atoms of the language, the morphemes, such an assignment can be accomplished case by case because there are only finitely many of these constituents. A complete assignment is a dictionary. It is a list of pairs, each consisting of an entry from the lexicon of the syntactic component, e.g., (chase, [+V, + _ N P . . . . ]), paired with a semantic representation for every sense of the lexical item. Since there are infinitely many constituents and sentences, the assignment of semantic representations to them must be made by a rule. Moreover, since the rule must specify (C), it must be a projection function that forms the semantic representation for a syntactically complex construction from the semantic representations of its component morphemes and their syntactic relations. 51 In the present form of TSR, the projection function does this by two operations. One is the substitution of a semantic
IF] representation R for an occurrence of a categorized variable X just in
<> case the constituent to which R has been assigned has the grammatical function [F] in the sentence and R satisfies the selection restriction ( ~ ; the other is a transformation of semantic representations in the scope of operators (e.g., negation). 52 Given the assignment of semantic representations to all constituents and sentences, the notion of predicting semantic properties and relations becomes well-defined once we specify (B). Current work in TSR has provided definitions for almost all of the familiar semantic properties and relations. In place of the possible worlds approach's definitions (19) and (20), TSR provides (B1) and (B2). (B1) (B2)
The expression E has a meaning (is meaningful) just in case E is assigned at least one semantic representation. Two expressions E~ and Ej have the same sense (are synonymous on a sense) just in case they are assigned the same semantic representation.
86
J E R R O L D J. K A T Z
(Note that none of the counter-examples t,~ (19) will be a counter-example to (B1) because each expression will have a compositional meaning, and that none of the counter-examples to (20) will be a counter-example to (B2) because the compositional meanings of each member of the pairs will be different.) (Bo) provides a further synonymy definition. (Bs)
(B~) (Bs)
Two expressions E, and E~ have the same meaning (are fully synonymous) just in case they are assigned the same set of semantic representations. The expression E has no meaning (is semantically anomalous) just in case E is assigned no semantic representation. The expression E is ambiguous just in case E is assigned two or more semantic representations.
(B4) and (Bs) are obvious next steps. The further semantic properties and relations of semantic inclusion, analyticity, and semantic entailment, though requiring more explanation about semantic markers than given above, can be defined in terms of the structure of trees like (22). The property of presupposition, in roughly Frege's sense, sa can also be defined, using an appropriate notation for referential occurrences of variables. ~4 The fact that the semantic representations assigned to all constituents are all of the same form, namely, sets of trees like (22), means that, in principle, there is no problem in applying formal simplicity considerations of the familiar sort to these representations. Hence, we have a reasonably clear criterion for distinguishing semantic information from extralinguistic information: (23)
The information I is semantic information about the linguistic construction C just in case the optimal semantic representation for C represents C's meaning as containing I. A semantic representation R is optimal for a linguistic construction C just in case, on the basis of correct definitions of the semantic properties and relations, R predicts each semantic property and relation of C and there is no simpler semantic representation of C that also predicts them.
6. The causal approach claims that we refer on the basis of empirical knowledge about the conditions under which the uses of words have the
THE T H E O R Y OF S E M A N T I C R E P R E S E N T A T I O N
87
appropriate causal connection to the baptismal ceremony in which the referent became its bearer. This will be an uncontested general account of language use only if the classical account, on which we refer on the basis of knowledge of meanings of words, can be eliminated first. Thus, the question of whether the 'counter-examples' brought up by Donnellan and Putnam eliminate the classical account is the principal one at issue in the controversy between TSR and the causal approach. The importance of this criterion, we reiterate, is that, by making no appeal to referential relations, that is, relations between linguistic constructions and their referents, but confining itself to features of the internal grammatical structure of sentences, the criterion avoids the difficulties that Donnellan and Putnam found with C. I. Lewis's account of meaning, and which, in one form or another, have plagued Fregean accounts that rest on notions like Lewis's notion of applicability, as We showed above that the difficulty about proper names is avoided because TSR does not make the claim about proper names that Fregean theories do. We now will show that the allegedly parallel difficulty about common nouns isalso avoided. Donnellan observed that it is indeterminate on C. I. Lewis's account whether animality is a semantic feature of the word 'cat', that is, whether (24) is analytic, or whether it is simply an invariant feature of cats. (24)
Cats are animals
Lewis's criterion, because it is based on the referential notion of applicability, provides no way to make non-arbitrary decisions in hypothetical cases, and hence, no way to decide who is right when one person says 'cat' applies to something that is not an animal and another denies this. To know who is right, the person who takes animality as a contingent property of cats or the person who takes it as a necessary property (by virtue of what "cat" means and how words refer), we cannot ask about the analyticity of (24) or about change of meaning. Questions about analyticity presuppose the answer to these questions of definition, and questions about change of meaning presuppose a non-arbitrary way of deciding what the meaning of a word is at the initial and terminal states of the process. 56 Putnam sharpens the criticism by imagining that the things 'cat' has been applied to all along are not animals but robots cleverly designed to look and act like animals, and then arguing that,
88
JERROLD J. KATZ
insofar as the criteria speakers use in applying 'cat' license applications to robots in this hypothetical case, (24) cannot be analytic, since in this case (24) would be false. 57 On our theory, as contrasted with Lewis's, it is not indeterminate whether animality is a semantic feature of 'cat' or simply a contingent feature of cats. Given (23), the question of whether the information that a cat is an animal is semantic is separate from the question of whether a referential relation holds between the word 'cat' and certain objects. This is because its answer depends on speaker's intuitions about the semantic properties and relations of linguistic expressions in which 'cat' occurs. Suppose that we have adequate evidence to show that an optimal semantic representation of 'cat' represents its sense as containing the concept of animality (e.g., the explanation of the ambiguity of 'cat' might require animality in the sense under consideration to contrast with the 'backbiting woman' sense of 'cat'). (23) then permits us to conclude that animality is part of the meaning of 'cat', that (24) is analytic, and that a change of meaning would occur if the same sort of evidence about a later stage of the language were to show that at that stage the optimal semantic representation of 'cat' represents its sense as lacking the concept of animality. Putnam's hypothetical case does not count against these Conclusions because it has to do with application. These conclusions are not obtained from a conception of semantics on which meaning is determined by the application of words, and they are thus not about such referential relations. These conclusions are about grammatical relations among words internal to the language. The conclusion that (24) is analytic is as much about a grammatical relation as the statement that (24) is well-formed. But not only do applications of 'cat' to robots not contradict these conclusions about the meaning of 'cat', these applications are exactly what one would expect on these conclusions. If (24) is analytic and if speakers falsely think (as Putnam imagines) that the meowing feline-like objects around them are animals, then uses of 'cat' ought to refer to the objects falsely thought to be feline animals28 There is, in principle, no reason why the causal approach and TSR should be opposed. The anti-intensionalism of the former appears to be nothing more than anti-Fregeanism, and hence nothing that ought to divide these approaches once the differences between Fregean theories like Lewis's and TSR is recognized. Nothing of substance from the rich
THE THEORY OF SEMANTIC R E P R E S E N T A T I O N
89
body of insights about reference in the work of Kripke, Donnellan, Putnam, etc. is lost by their accepting the role of meaning in reference. The distinction between semantic information and factual information employed in the use of language simply divides the domains of the two approaches. TSR also distinguishes the separate roles of these two kinds of information in reference. 59 TSR distinguishes between type reference and token reference: the type referent of a word is an object conforming to its meaning; the token referent is what it actually refers to on a use of the word. On Putnam's hypothetical case, there is no type referent of 'cat' but numerous robot spy devices have been among its token referents. TSR claims that token reference works like heuristic programming. The principal function of stereotypes and other encyclopedia information is to expedite token reference. Practically speaking, token reference cannot be based on semantic information directly. We cannot use the meaning of 'doctor' in English to pick out a doctor because, in most cases, there is no way to check whether the intended referent has earned the degree of doctor q/" medicine. Encyclopedia information provides the identification strategies with easily verifiable properties like 'wears a white hospital jacket with prominently displayed stethoscope'. But not any easily verifiable property will do. It will not do, for example, to employ the property 'wears hippie clothes' in an identification strategy for doctors. The properties have to correlate highly enough with the set of definitional properties. Thus, the function of semantic information in token reference is both to provide the anchor for identification strategies and to serve as a failproof procedure in cases where trade of a small risk of error for a large saving in time and effort backfires, e.g., picking out an actor on location in a hospital. 6~ 7. The conflict between TSR and the Tarskian approach is about the truth of a theory such as this in which meanings have an indispensable function in reference. As developed by its principal proponent, Davidson, this approach is explicitly reductivist in replacing the intensionalist paradigm for describing the logical structure of sentences's means p' by the Tarskian paradigm ' " s " is true - p,.61 The choice of the Tarskian paradigm as the replacement and Davidson's adaptation of the paradigm to natural language are designed to eliminate meanings at the outset so
90
JERROLD .l. KATZ
that the speaker's factual beliefs by themselves play the role in Davidson's approach that meaning plays in the intensionalist approach. 62 I have argued that an appeal to Quinean strictures against meanings and against semantic properties and relations like synonymy constitute Davidson's only justification for replacing the intensionalist paradigm; I also argued that the replacement of the intensionalist paradigm by an extensionalist one weakens the constraints on the description of logical structure to such an extent that a variety of absurd accounts of the logical structure of a sentence are indistinguishable from acceptable ones. 6a I won't take up these criticisms here: suffice to say that not only are the counter-examples to (20) counter-examples to Davidson's paradigm of analysis but so are materially equivalent sentences like 'Snow is white' and 'Grass is green'9 Here I want to examine Davidson's recent defense of his attempt to make the speaker's factual beliefs do the work of both dictionary and encyclopedia in reference. Davidson's account of how we identify something in the use of words, "the intelligibility of such identifications must depend on a background of largely unmentioned and unquestioned true beliefs", 64 is recently defended as follows: 9 h o w c l e a r a r e we that the ancients - some ancients - believed that the earth w a s flat? This earth.'? Well, this earth of ours is part of the solar system, a system partly identified by the fact that it is a gaggle of large, cool, solid bodies circling around a very large, hot star. If someone believes none of this about the earth, is it certain that it is the earth that he is thinking about? An answer is n o t called for. The point is made if this kind of consideration of related beliefs can shake one's confidence that the a n c i e n t s b e l i e v e d the earth w a s flat. es
But Davidson is mistaken in thinking that his point is made if the question can shake one's confidence that the ancients believed the earth to be flat. Shaken confidence can equally be diagnosed as a sign of nothing more than a disposition to Davidsonian views about dictionaries and encyclopedias. Someone who thinks dictionary information can be sharply distinguished from encyclopedia information (say, on the basis of (23)) and entertains the lexical hypothesis that 'earth' means 'the physical object part of whose surface is the ground' runs no risk of having their confidence in their beliefs about the ancients shaken. It is perfectly clear to such a person that an ancient who was thinking about the physical object whose surface included the ground on which he or she stood was thinking about
THE THEORY OF SEMANTIC R E P R E S E N T A T I O N
91
the earth, even though he or she did not believe the earth to be part of the solar system, etc., etc. Once the methodological point is clarified, the example can be used as a counter-example to Davidson's account. Davidson has a way of coming up first with the counter-examples to his theory but of presenting them in a context that makes them seem favorable to it. It is true that the ancients referred to the same earth, sun, and moon as we do with our vastly different beliefs (even if not all of us would think it certain), but Davidson must deny this. People with such vastly different beliefs cannot have the same thing in mind. Thus, to change the case, a racist and a member of the NAACP cannot disagree about Blacks. On Davidson's account, they would be making claims about different things. In fact, since the racist's beliefs about Blacks are false, we could not even say that the racist's beliefs are beliefs about Blacks or that the racist believes false things about Blacks9 Hitler didn't have anything against the Jews, he just didn't know what he was talking about. The assorted 'bizarre' consequences of the Tarskian approach are put up withbecause philosophers have been convinced by Quine that nothing is available by way of the semantic relations. It is hard to imagine a philosopher who thought that logical analysis might be put under stronger constraints than material equivalence still accepting the consequence that an analysis of 'Snow is white' which is good as any other is the condition that grass is green, or the consequence that a racist's beliefs are not about Blacks so there is no disagreement with NAACP members. Accordingly, the Tarskian approach is undermined if, as I have claimed, Quine's arguments against semantic relations depends on discredited assumptions from Bloomfieldian taxonomic grammar. It is supposed by some that, even if Quine's arguments fall though, Davidson has little to fear because he can rest his move from ' " s " means p' to ' " s " is true - p' on Harman's arguments against semantic properties and relations. Harman differs from Quine in siding with Chomsky against Bloomfield and the structuralists. Therefore, so the supposition runs, Davidson has Harman to fall back on. 66 According to Harman's argument, positing semantic properties and relations in order to describe the logical structure of sentences 9 is a mistake9 The relevant notions of equivalence and of implication are the ordinary o n e s . . . It adds nothing to one's understanding [of language] if one can
92
J E R R O L D J. K A T Z
distinguish 'analytic' equivalence and implications from 'synthetic' ones. In fact, most people cannot do so. Only those who have been 'indoctrinated' can.e7 It is irrelevant whether most people can distinguish analytic and synthetic relations. It matters no more than whether most people can distinguish indirect objects from objects of prepositional phrases or distinguish quantificational and truth-functional validity. Not indoctrination but mastery of the appropriate system underlies the ability to make the distinctions. The question at issue, then, is whether there is evidence, o f the same sort as used to ground the ordinary notions of equivalence and implication, that shows that a theory of the logical structure of natural language also requires the notions of synonymy and analytic implication. On this question, there is overwhelming evidence for semantic properties and relations. It comes not only from the bizarre consequences that arise when these further properties and relations are unavailable but from clear cases of sentences and constituents that exhibit semantic properties and relations, e.g., cases like the analyticity of (25) or the corresponding redundancy of (26). (25) (26)
Nightmares are dreams a nightmare that is a dream
It is not easy to see why inferences like (1) to (2) or (8) to (9) are supposed to be obviously necessary inferences while inferences like that from 'Norman had a nightmare' to 'Norman had a dream' are not. I have always been puzzled as to why Harman thinks that synonymy is not a transitive relation and that analytic implication is better understood as non-demonstrable inference, 6a but doesn't think that logical equivalence is not transitive and logical implication merely non-demonstrable. Overwhelming evidence can fail to overwhelm. Faced with clear cases like (25)-(26), Harman stonewalls it. 69 He argues (from his theory) that what we respond to in such cases is nothing more than the result of a computation over beliefs, namely, that it is extremely likely that every nightmare is a dream. So it looks now as if we have reached an 'ultimate disagreement'. This appearance is confirmed when we look at the arguments and counter-arguments over the status of semantic anomaly and analyticity. Harman claims ...there is no such thing as semantic anomaly and no such distinction as the analytic-synthetic distinction. To believe otherwise is to suffer from a lack of imagination. 7~
THE T H E O R Y OF S E M A N T I C R E P R E S E N T A T I O N
93
Harman's typical proof for this claim is the assurance that he and other more imaginative folk can conceive of such things as happily married bachelors, southpaws that are righthanded, nightmares that are not dreams, and irrational numbers that go berserk and beat people to death in dark alleys. Our counter-claim is that it is not lack of imagination on our part that keeps us from seeing the errors of our ways, but too active an imagination on Harman's that makes him think he can conceive the inconceivable. Our view is that happily married bachelors, righthanded southpaws, and the like are as unimaginable as something that both has and does not have some property, and that our knowledge of their impossibility is likewise based on an intuition of logical structure. But, though it surely looks like an ultimate disagreement, fortunately, this is not so. Harman recognizes ambiguity as a genuine grammatical phenomenon. Here, he does not claim that intensionalists have manufactured a grammatical phenomenon because their sluggish imaginations couldn't do better, but that the phenomenon of ambiguity can be explained without invoking the notion of sense or meaning. This acknowledgement that the t~henomenon is grammatical makes the issue a straightforward one of whose explanation of the phenomenon is best. If Harman's argument that ambiguity can be explained without appeal to the notion of sense or meaning fails for essential reasons, the issue is settled in favor ofintensionalism. Harman divides ambiguities into two classes, syntactic ones, typified by (27), (27)
Visiting philosophers can be unpleasant
where the ambiguity arises from the coincidence of different derivations in the same surface structure, and lexical ones, typified by 'grass', where items having no constituent structure bear more than one sense. Harman sees the explanation of syntactic ambiguities as accomplished satisfactorily by the assignment of syntactic descriptions that represent ambiguous sentences as having one or another syntactic structure. But it is too facile to claim that all syntactic ambiguities can be fully explained on the model of cases like (27), and even too facile to claim that in cases like (27) the explanation is purely syntactic and phonological. Sentences like (28) and (29) (28) (29)
John wrote a letter about his experiences It was done by an automated processing device
94
JERROLD J. KATZ
are assigned different syntactic derivations, exactly in the manner of (27), but neither is ambiguous. These examples show that meaning relations can nullify the potential of differences in constituent structure to produce ambiguity. Thus, ambiguity cannot be defined at the syntactic level: syntactic definitions will fail to cover cases like (28) and (29) and will leave unexplained why cases like (27) are different from them. H a r m a n ' s explanation of lexical ambiguity must be quoted in full: we must assume that he [the speaker] distinguishes a word used in one sense from the same word used in another sense. But we can do that without assuming that he makes the distinction by assigning different readings to the word. We may mark the distinction in sentences by a device as simple as a subscript. The inferences and paraphrases that are then permissible depend in general on the subscript selected. One reason why no more is needed than a subscript is that the relevant sorts of inference and paraphrase are those possible by virtue of background information9 Such background information itself must be 'stored' as sentences under certain structural descriptions including subscripted words. 71 9
What makes a sequence of symbols a semantic representation ('reading' is another name for a semantic representation) is not its complexity. Senses might have turned out to be very simple in structure, in which case their representations would be sirnple too. What makes a sequence of symbols a semantic representation is that its formal structure captures the sense distinctions speakers make in connection with the constituent to which it is assigned. With this in mind, even H a r m a n ' s devilishly clever euphemism 'subscript' can no longer hide the fact that such special subscripts are rudimentary semantic representations. N o t only do they "distinguish a word used in one sense from the same word used in another sense", they provide an appropriate formal basis for defining the semantic properties and relations of synonymy, meaningfulness, semantic anomaly, etc. as well as ambiguity (assuming, as H a r m a n must allow, that each n-ways ambiguous lexical'item receives n distinct subscripts or appears in the lexicon n times with a distinct subscript each time)9 The definitions, which are restricted to the lexical level because H a r m a n speaks only about assigning subscripts to words, are (30)-(33). (30) (31)
Two lexical items are synonymous if, and only if, they receive the same subscript A lexical item is meaningful if, and only if, it receives at least one subscript
T H E T H E O R Y OF S E M A N T I C R E P R E S E N T A T I O N
(32) (33)
95
A lexical item is semantically anomalous if, and only if, it receives no subscript. A lexical item is ambiguous if, and only if, it receives two or more distinct subscripts.
For example, with the subscripting (34) and (35) (34) (35)
poh, potj grassy, grass~
allowed us on Harman's scheme, we cannot only mark the ambiguity of 'pot' and 'grass' but their synonymy (on a sense). This argument can't be gotten around without giving up Harman's original claim that the speaker's distinctions between a word used in one sense and the same word used in another can be accounted for without resorting to semantic representations. If Harman were to claim that the notion of sense underlying the definitions (30)-(33) is to be understood either in terms of Quine's conception of stimulus meaning or, more plausibly, in terms of some conception of inference-set (e.g., the sense of a sentence might be taken as the set of all and only the sentences inferable from the appropriate subset of the speaker's beliefs), the resulting match between predicted sense distinctions and speaker's actual sense distinctions would be too wildly off for one to accept Harman's putative reduction of the theory of meaning. With the inference-set conception of a sense, the prediction in the case of racists and members of the NAACP is that the word 'negro' is ambiguous in the manner of 'pot' or 'grass' but the fact is that neither group makes a sense distinction in the use of'negro' corresponding to their beliefs about negroes. We can imagine an argument between a senior citizen and a young person about the legality of growing grass dissolving because it is pointed out that 'grass' has both the sense of 'ordinary lawn grass' and 'marijuana' but we can't imagine an argument between a racist and a member of the NAACP dissolving because it is pointed out that 'negro' has both the sense 'member of the black race of mankind' and 'member of a race biologically inferior to the Caucasian race of mankind in intelligence, etc.'. It is absurd to suppose that such an argument could be settled on the grounds that the racist and the NAACP member were using 'negro' to refer to different things. Furthermore, it is not even possible to provide a non-circular specifica-
96
J E R R O L D J. K A T Z
tion of the appropriate subset of the speaker's beliefs required for a well-defined inference-set conception of sense. Such a specification would have to antecedently distinguish beliefs (which, for Harman, are sentences) on the basis of their ambiguous constituents, since otherwise the consequence set for a particular set of the speaker's beliefs will be the incoherent product of fallacies of ambiguity. But such an antecedent differentiation of sentences presupposes that we already have a well-defined inference-set conception of sense, since it is just this that is needed for distinguishing the ambiguities in sentences. The circle cannot be broken so long as the same class of sentences are both the bearers of the ambiguities and the objects used to disambiguate. The dilemma Harman faces is this: either he takes the notion of a sense in a way that maps one-one onto the speaker's lexical distinctions or he takes it in a weaker way. In the former case, he has a full-blooded theory of senses, in no way different from TSR, except perhaps in sophistication. Harman's apparatus for making lexical ambiguities thus contradicts his philosophical claim that there is no need to set up a level of grammatical representation, beyond the syntactic and phonological levels, at which semantic properties and relations can be defined. In the latter case, Harman has no alternative account of the speaker's distinctions among lexical senses that can claim to be empirically realistic. Hence, here, too, he can make no claim to have dispensed with the need to set up a semantic level of representation in grammars. As such subscript apparatus is forced to become less rudimentary, Harman's philosophical claims become even more implausible. What forces the apparatus to become so is that in its rudimentary form it offers no account of ambiguity above the lexical level. There is, in the first place, no way to account for the contribution of lexical ambiguities to the ambiguities of phrases, clauses, and sentences. (Harman is curiously silent on this matter.) Secondly, there is, as examples like (28) and (29) show, no such thing as purely syntactic ambiguity. Since ambiguity is inherently semantic, even the attempt to mark so-called syntactic or constructional ambiguity cannot succeed without some account of the contribution of lexical senses to sentences. Harman's claim that "no more is needed than a device as simple as a subscript" loses all plausibility once the full scope of the problem of explaining the ambiguity of sentences is appreciated.
T H E T H E O R Y OF S E M A N T I C R E P R E S E N T A T I O N
97
So, even if Harman's case against a semantic level of representation were otherwise air-tight, he could not make good his claim to handle ambiguity unless his scheme of subscripting can be extended to the infinite range of sentences. To do this, it must be made less rudimentary at least by the addition of a mechanism for marking the degree of ambiguity of syntactically complex expressions and sentences by projecting subscripts for them from the subscripts assigned to their lexical items. But if this addition is made, the scheme automatically defines semantic properties like semantic anomaly and analyticity. Suppose we have such a mechanism. Since the number of actual senses of an average, say, twenty word sentence is far less than the hundreds possible on the basis of the number of senses of the lexical items in it, some rather severe form of selection will be required to filter out the possible but not actual senses. It will be theoretically possible that in some cases this form of selection spares nothing. Such cases will embody the extreme of such ambiguity reduction, a degree of ambiguity of zero. Degree of ambiguity of zero is, however, nothing other than semantic anomaly. It should be observed that, even if there is, as a matter of fact, no sentence whose degree of ambiguity is zero, this definition of semantic anomaly in terms of a degree of ambiguity of zero is still legitimate. In fact, it is more than legitimate. If a natural language contains no sentence that is zero ways ambiguous, we must be able to state this fact. It doesn't matter how we state it in the notation of subscripts, we will be able to define semantic anomaly, as indicated. Such a mechanism enables us to define redundancy and analyticity, too. Given the immediately preceding definition of semantic anomaly, we can also define redundancy, as in (36). Analyticity can be defined as in (37). (36)
(37)
A modifier head construction E is redundant if, and only if, it is not the case that the modifier is semantically anomalous and E is synonymous with the head of the construction by itself. A subject predicate sentence S is analytic if, and only if, the modifier head construction formed from S by taking S's subject as the head and S's predicate as the modifier is redundant.
98
JERROLD J. KATZ
(36) says that the modifier, the head, and the whole construction have subscripts and the subscripts of the head and the whole construction are the same. ~2 At the outset of this discussion, we said that attempts to handle ambiguity without going beyond the syntactic and phonological levels will fail for essential reasons. The nature of these reasons now emerges: the semantic properties and relations of a natural language are so tightly interconnected that semantic representations of linguistic constructions that are fully adequate for defining ambiguity cannot fail to be fully adequate for defining the other semantic properties and relations. 8. We come now to the last contrast, the contrast between TSR and the meaning-as-use approach. As in the case of the causal approach, we claim that there is no necessary opposition between these approaches. In the case of the meaning-as-use approach the opposition arises because, on the one hand, use-theorists have tried to make ideas about how to describe and explain the use of language apply to semantic questions in the domain of grammatical structure, and on the other, TSR insists on the sharp distinction, expressed in (23), between the speaker's knowledge of the semantic structure of sentences and the speaker's extralinguistic knowledge (on which use is also based). Thus, TSR's viewpoint is that such use-theorists have the relation between meaning and use backwards. Rather than defining meaning in terms of use, use has to be defined, partly, on the basis of an antecendent grammatical notion of meaning. Typical of formulations of the meaning-as-use thesis is Strawson's claim: To give the meaning of an expression.., is to give general directions for its use to refer to or mention particular objects or persons; to give the meaning of a sentence is to give general directions for its use in making true or false assertions, va Typical of the difficulties for this approach is the following. Strawson, arguing against Russell, says: The fact that the sentence ['The king of France is wise'] and the expression ['the king of France'], respectively, are significant just is the fact that the sentence could be used, in certain circumstances, to say something true or false, that the expression could be used, in certain circumstances, to mention a particular person; and to know their meaning is to know what sort of circumstances these are. 7~
T H E T H E O R Y OF S E M A N T I C R E P R E S E N T A T I O N
99
Thus, consistently with the basic idea of the approach, meaningfulness is explicated in terms of the possibility of being used in an appropriate way. This, however, is wrong as counter-examples such as (38) clearly show. (38)
The least convergent series is infinite
(38) is certainly meaningful, yet it cannot be used to say something true or false, and (38)'s subject is certainly meaningful, yet it cannot be used to mention anything. The fact that we know there are no circumstances under which it mentions something does not make the expression meaningless: its meaning is what enables us to prove that there are no such circumstances. This, of course, is only one example, but it nicely points up the fact that meaning is the prior notion. The more familiar objections to taking meaning as use are pairs of synonymous words and expressions whose members have different uses. For instance, there are pairs like: (39) (40) (41) (42)
bunny, rabbit shit, feces nigger, Black skunk, Alaska sable
What such counter-examples show is that use is controlled by such nonsemantic factors as appropriateness to children, sensitivity to obscenity, prejudice, and euphemistic rhetoric, none of which is p e r se semantic. Thus, TSR takes semantic representations to describe the contribution of the speaker's knowledge of the language to his or her use of the language, and so, to be part of a broader theory of the use of language that explains how non-semantic factors also influence utterances and their comprehension. TSR has sought to clarify this question of domain by sketching a framework for theorizing about language use. ~8 This framework distinguishes two notions of use, first, a notion in which use is a function of meaning, "use in the null-context', and second, a notion in which it is a function also of features of some particular context, that is, features not shared by all contexts, 'use in a non-null context'. We think of contexts arranged on a dimension, representing the degree to which contextual information determines aspects of the utterance meaning of a use of a sentence. At
100
JERROLD
J. K A T Z
one extreme are contexts in which contextual information completely determines the utterance meaning of a use of a sentence. Nothing in the meaning of the sentence (in the language) plays any role in its utterance meaning. The obvious example of such contexts are cases in which a code determines utterance meaning. At the other extreme is the null context. It is null in that no particular features of the context determine any aspect of utterance meaning. The utterance meaning of a use of a sentence in this context is e.xactly the meaning of the sentence in the language. In the range of cases between these two extremes are contexts in which utterance meaning is a function both of grammatical laws and relevant features of the context. With this distinction, we can characterize the independent domains of grammatical structure and language use more satisfactorily. We can take grammatical competence in Chomsky's sensC s to be the knowledge that qualifies a speaker of a language to use the language in the null context. The notion of the null context is, of course, an idealization, like the notion of a frictionless plane. Thus, the behavior of speakers in the null context is a function of the grammatical laws of the language just as the behavior of physical objects on a frictionless plane is a function of physical laws. This makes it appropriate to equate the ideal speaker's knowledge of the grammatical laws with the ability to use language in the null context. Further, we can hypothesize that speakers have a knowledge that qualifies them for using language in non-null contexts. We may call this pragmatic competence. This knowledge is what enables a speaker to exploit contextual information to produce or understand an utterance of a sentence to have an utterance meaning different from the meaning of the sentence in the language but nonetheless determinable by everyone with the same pragmatic competence and contextual information. The framework for theorizing about language is, briefly, this. A theory of pragmatic competence is the specification of a function of the form (43), (43)
P R A G (D(SO, I(C(t))) = R1,..., R~
where D(S~) is the full grammatical description of the sentence S~ in an optimal grammar of the language, I(C(t)) is the information about the context C relevant to the interpretation of the token t of S~ in C, and R1 . . . . . R~ are semantic representations, reflecting the utterance meaning
THE T H E O R Y OF SEMANTIC R E P R E S E N T A T I O N
101
of the token t of S~ in C. 77 There can be zero, one, or a number of such representations in the output of such a function, since a token of a sentence can be situationally anomalous, bear a unique sense, or be situationally ambiguous. Since the utterance meaning of a token of a sentence in a context can be conveyed by the sense of a sentence, the output of (43) is formulated in the same notation used to represent senses in grammars of natural languages. Having semantic representations do double duty in this way not only saves us the job of creating a new notation but, more importantly, expresses the claim that utterance meaning is not another kind of meaning. The claim is that there is a single notion of meaning common to grammar and pragmatics, and that what makes a meaning a sentence (grammatical) meaning or an utterance meaning is whether it is viewed as the meaning of a sentence in the language (the null context) or the meaning of a use of a sentence in a non-null context. A grammar and a pragmatic theory both express sound-meaning correlations. 78 A grammar correlates sound and meaning in the language, while a pragmatic theory correlates them in contexts of language use. The perlocutionary notion of meaning, developed principally by Grice but also by Searle and others, ~9 is an account of what it is for a person to mean something by, for instance, uttering a sentence (but also by performing non-linguistic acts). In our framework this notion is understood as a proposal about the pragmatic principles speakers use to relate sound and meaning in non-null contexts. The attempt on the part of perlocutionary theorists to explain meaning in terms of the effect on hearers that a speaker intends to produce by means of their recognition of his or her communicative intention is an attempt to specify the character of the principles for correlating sound and meaning in such contexts. Grice's maxims, etc. is part of the specification of the particular principles and of the notion of relevant contextual information. These maxims, etc., of course, do more than this. They also explain how speakers convey information that is not part of the meaning of their utterance in the context. A conversational implication to the effect that the speaker of the utterance 'My wife is either in the kitchen or the study' does not know which of these rooms his wife is in is to be explained, as Grice has shown, on the basis of shared principles of conversation about how much information utterances can contain. But, insofar as such ap-
102
.IERROLD $. K A T Z
paratus addresses just the question of utterance meaning, it is construed in our framework as part of the explanation of how the sound-meaning correlation in non-null contexts can systematically diverge from the logically prior correlation in the language. This raises the too little debated question of what the fundamental notion of meaning is. This is not the old question of whether explanations of meaning are explanations of how to use words to refer and make assertions, but it is related. Moreover, underlying both is, I think, the deep philosophical problem of what the source of meaning in Grice's sense of 'non-natural meaning' is. Philosophers like Grice, Searle, and their followers belong to a tradition that holds that its source is in human agency, in our uses of words and other meaningful signs. Wittgenstein put the position thus: Philosophers very often talk about investigating, analysing, the meaning of words. But let's not forget that a word hasn't got a meaning given to it, as it were, by a power independent of us, so that there could be a kind of scientific investigation into what the word really means. A word has the meaning someone has given it.s~ The other, more rationalist, tradition holds that the source is indeed something independent of us, namely, our innate conceptual apparatus. Such apparatus, so this tradition claims, is presupposed in any uses of words or other meaningful signs. TSR, as I've developed it, is clearly based on this rationalist tradition. The burden of the above discussion has been to show in what sense the grammatical notion of meaning is logically prior to the perlocutionary notion, namely, the grammar's specification of the sound-meaning correlation in the Language is a necessary part of the pragmatic theory's sound-meaning correlation for contexts, but not the other way around. I have not tried to present arguments here. Interesting arguments have been made by Ziff and Chomsky, sl and I have given some elsewhere, a2 There has as yet been no debate stemming from them, but some philosophers influenced by Grice have argued for the logical priority of the perlocutionary notion. 83 Schiffer, in an important contribution to the understanding of perlocutionary meaning, 64 puts forth two considerations favoring the view that perlocutionary meaning is the primary notion. His first consideration is that knowing what one would 'normally or ordinarily mean by uttering x' is a necessary condition for knowing the meaning of a whole-
THE THEORY OF S E M A N T I C R E P R E S E N T A T I O N
103
utterance type x. a5 Although one might quibble about the truth of this claim, let us accept it for the sake of argument, and reply that, even if true, the consideration offers no support for the priority of the pedocutionary meaning. This is because the notion of what one normally or ordinarily means by uttering an expression depends on the prior notion of grammatical meaning. This notion provides the norm to which the utterance meaning of a use of an expression must conform to count as what is normally or ordinarily meant by uttering it. a6 Schiffer's second consideration is that " . . . i t is possible for a person to mean something by uttering x even though x has no meaning", a7 This is obviously true, but not So obviously relevant. The claim that grammatical meaning is prior does not imply that the pedocutionary meaning of every sign-token is associated with that sign in the grammar's sound-meaning correlation. What the claim implies is that the meaning occurs in the system of grammatical meanings, and that the systematic use of expressions and sentences from the language to mean what they do in non-null contexts requires information about the antecedent correlation of sounds and meanings in the language. Schiffer's example - someone uttering 'grrr' to inform the audience of his or her anger - exploits a natural meaning connection, to use Grice's notion. The sign, however, does not occur in the stock of signs that the grammar correlates to meanings, so the example is no counter-example. The point is this. The mechanism of communication by uttering something with the intention to produce comprehension in an audience by means of their recognition of this intention will work as long as there is a connection between a sign and a meaning to start with. For the limited sort of communication illustrated in Schiffer's example, the connection can be a natural connection, such as that dog grrrs mean an angry dog, since it takes very little to see the onomatopoetic relation a speaker has in mind. But, for the unlimited sort of communication characteristic of natural languages, the connection has to transcend natural connection. It must be nothing less than the sophisticated system of connections expressed in the sound-meaning correlation of natural languages.
The Graduate Center The City University of New York
104
JERROLD
J. K A T Z
NOTES 1 The author wishes to thank Joe Bevando and Linda Wetzel for a helpful discussion of this paper. 2 G. Frege, "On Sense and Reference", in Translations from the Philosophical Writings of Gottlob Frege, eds. M. Black and P. Geach, (Oxford: Basil Blackwell, 1952), pp. 56-78. a See J. J. Katz, 'Logic and Language: An Examination of Recent Criticisms of Intensionalism', in Language, Mind, and Knowledge, Minnesota Studies in the Philosophy of Science, vol. VII, ed. K. Gunderson (Minneapolis: Univ. of Minnesota Press, 1975), pp. 36-40. Some interpretations of the semantic representations in TSR do reify senses. But such Platonist interpretations are not the only way to interpret this theory. It may also be interpreted as a theory of semantic competence. Furthermore, Platonist interpretations do not materialize abstract objects as sometimes suggested by its critics. Taking senses to be legitimate objects of scientific study allows for the possibility that further knowledge of them, gained in the construction of a theory of their behavior, may, as in the case of the discovery of the nature of the planets, be necessary to find out what sort of objects senses are. Our approach is anti-Quinean just in accepting theory construction as legitimate in semantics. See W. V. Quine, Word and Object, (Cambridge: M.I.T. Press, 1960), pp. 73-79. s In particular, J. J. Katz, Semantic Theory, (New York: Harper & Row, 1972), 'Logic and Language: An Examination of Recent Criticisms of Intensionalism', and J. J. Katz, 'Where Things Now Stand with the Analytic-Synthetic Distinction', Synthese 28 (1974), pp. 284-319. 6 I say "tight enough constraints" because it is reasonable to construe Quine's requirement that translation preserve the logical truths in the case of the logical particles and relations of stimulus meaning in the rest of the vocabulary as a (weak) constraint on the level of semantic representation. We may briefly explain the assertion in the text that indeterminacy of translation is an artifact of failing to impose strong enough constraints. The important point to note is that claims of Quine's like "We could equate a native expression with any of the disparate English terms 'rabbit', 'rabbit stage', 'undetached rabbit part', etc., and still, by compensatorily juggling the translation of numerical identity and associated particles, preserve conformity to stimulus meanings of occasion sentences" (Word and Object, p. 54) are true relative to his own minimal constraints but false relative to stronger ones like our requirement that semantic representations predict all the semantic properties and relations of the expressions to which they are assigned. Thus, with such stronger constraints, we can justify a unique translation of the native expression, say as 'rabbit' on the grounds that bilingual natives say their expression is synonymous with 'rabbit'. (See the further discussion of this point in 'Where Things Now Stand with the Analytic-Synthetic Distinction', pp. 289-291.) Quineans will, of course, feel that this is an attempt to 'pull a fast one'. They are likely to query, 'How can you propose to use evidence from such natives about synonymy when the issue of indeterminacy is an issue about whether such a relation exists?' The answer is that, if there were a circularity, it would be Quine's. Quine, after all, has claimed to establish the thesis of indeterminacy, so he has to show that we cannot employ such stronger constraints in this way. Quine, however, has presented arguments against the possi-
THE THEORY OF S E M A N T I C R E P R E S E N T A T I O N
105
bility of such further constraints. The trouble is that these arguments depend on the taxonomic theory of grammar, in particular, its conception of grammatical representation as limited to what can be discovered in a constructive, step by step fashion, beginning with speech sounds, where each new piece of grammatical apparatus is introduced solely on the basis of distributional evidence and cannot be defined using apparatus not yet introduced. Quine's arguments fall through when this theory is replaced by Chomsky's conception of a grammar as an abstract system with no such restrictions on its levels of representation. I have discussed Quine's reliance on taxonomy and the effect of Chomsky's revolution in linguistics on Quine's philosophical work in 'Logic and Language: An Examination of Recent Criticisms of Intcnsionalism' and 'Where Things Now Stand with the Analytic-Synthetic Distinction'. 7 Frege's famous telescope analogy, on which a sense is compared to a real image of the object viewed (to which the referent is compared), p. 60. a Ibid., pp. 65f. The notion of sameness is that of common explanatory pattern. Fodor explicates the notion further in saying, " . . . a science has to discover what it is about; it does so by discovering that the laws and concepts it produced in order to explain one set of phenomena can be fruitfully applied to phenomena of other sorts as well". J. A. Fodor, Psychological Explanation, (New York: Random House, 1968), p. 10. See my application of this idea to semantics in J. J. Katz, 'The Real Status of Semantic Representations', Linguistic Inquiry 8 (1977), pp. 570-574. 10 This relation is discussed in a number of recent works, principally in J. J. Katz, Propositional Structure and Illocutionary Force (New York: T. Y. Crowell, 1977), chapter 1, and in J. J. Katz, 'A Proper Theory of Names', Philosophical Studies 31, no. 1, pp. 35f; and Katz, J. J. 'The Neoclassical Theory of Reference', in Contemporary Perspectives in the Philosophy of Language, P. A. French, T. E. Uehling, and H. K. Wettstein, eds., The University of Minnesota Press, Minneapolis, (in press). n 'Sense and Reference', p. 58. 12 K. Donellan, 'Necessity and Criteria', reprinted in Readings in the Philosophy of Language, eds. J. Rosenberg and R. Travis (Englewood Cliffs: Prentice-Hall, 1971), pp. 42-52, and H. Putnam, 'It Ain't Necessarily So', also reprinted in Readings in the Philosophy of Language, pp. 52-62. la L. Wittgenstein, Philosophical Investigations, Basil B1ackwell, Oxford, 1953, pp. 36o-38 ~. 14 j. Searle, 'Proper Names', in Readings in the Philosophy of Language, pp. 212-218; see my discussion in 'A Proper Theory of Names'. 15 S. Kripke, 'Naming and Necessity', in Semantics of Natural Language, eds. D. Davidson and G. Harman, D. Reidel Publishing Co., Dordrecht, Holland, 1972, pp. 253-355. 16 We expect that the laws of grammar that explain semantic properties and relations, once discovered, will help to determine the nature of meanings. The claim is that at the outset there is no more need to commit ourselves to an account of the nature of meanings than the ancient astronomers had to commit themselves to an account of the nature of the planets prior to their learning the laws that govern planetary motions. 17 This is discussed in J. J. Katz, 'The Real Status of Semantic Representations', pp. 570-574.
ta 'Necessity and Criteria' and 'It Ain't Necessarily So'.
106
J E R R O L D J. K A T Z
1, The principal philosophical source is J. S. Mill, A System of Logic, Vol. I, 5th Edition, Parker Son, and Bourn, West Strand, London, 1862; See also, Semantic Theory, chapter 5, and ' A Proper Theory of Names,' pp. 11-13. 2o G. Frege, Conceptual Notation and Related Articles, ed. T. W. Bynum, The Clarendon Press, Oxford, 1972. 21 L. Wittgenstein, The Blue and Brown Books, Basil Blackwell, Oxford, 1958, p. 28. 22 Here I include most of the logical empiricists, logicians like Tarski, and others who see rational reconstruction not as theory construction but as 'reforming ordinary language'. Cf. The Linguistic Turn, ed. R. Rorty, University of Chicago Press, Chicago, 1967, pp. 193-205. 2a 'On Sense and Reference', p. 58. 24 See the discussion of effability in 'Effability and Translation' in The Theory of Translation: Linguistic and Philosophical Approaches, eds. M. Guenthner-Reutter and F. Guenthner, Duckworth, London, 1978. 25 N. Chomsky, Aspects of the Theory of Syntax, M.I.T. Press, Cambridge, pp. 3-9. 26 The reason that this is a semantic distinction is that it rests on analyticity: what makes something a theory is that it has synthetic sentences of a language that are set down as postulates. 27 See my paper 'The Rationalist Criterion of Cognitive Significance', in preparation. 2a C. G. Hempel, 'Problems and Changes in the Empiricist Criterion of Meaning', in Readings in the Philosophy of Language, pp. 420-436. 29 Frege's conceptual notation might fit into such a theory at this point. 30 R. Carnap, Meaning and Necessity, 2nd. Edition, University of Chicago Press, Chicago, 1965, pp. 222-229. 31 G. Frege, The Philosophy of Arithmetic, trans. J. L. Austin, Basil Blackwell, Oxford, 1953, p. 4 ~ a2 See the discussion in Semantic Theory, pp. 171-182. aa This point shows, as we shall see below, that Carnap's approach is subject to essentially the same criticisms as the possible worlds approach. 34 See the more complete form of the argument in J. J. Katz, 'The Advantage of Semantic Theory over Predicate Calculus in the Representation of Logical Form in Natural Language', New Directions in Semantics, The Monist 60 (1977), pp. 380--405. as It might be thought that the cases are disanalogous because there is no upper limit on the number of sentential constants and sentential meaning postulates that is required, but, of course, the number of predicate and individual constants, and so the number of meaning postulates for them, is also infinite. The objection fails to take account of the fact that the same recursive grammatical rules that generate sentences generate complex phrases functioning as predicate and individual constants. The problem is thus the same in both cases, namely, the absence of projection machinery. 3e 'The Advantage of Semantic Theory over Predicate Calculus in the Representation of Logical Form in Natural Language' argues that this is an inherent feature of predicate calculi. (Carnap's contribution was to realize the potentiality in predicate calculi to treat logical relations within the extralogical vocabulary on the model of their treatment in the logical vocabulary.) a7 Semantic Theory, pp. 183-192. as Aspects of the Theory of Syntax, pp. 60-62. ao A theory of logical deduction must contain a theory of semantic representation, since these representations explicate logical form and hence the features of sentences
THE THEORY OF S E M A N T I C R E P R E S E N T A T I O N
107
on the basis of which laws of logic apply, but the latter is separate from the former. Hypotheses about intra-constant logical structure are justifiable, independently of the theory of deduction, by their role as the basis for predicting semantic properties and relations. to W. V. O. Quine, 'Two Dogmas of Empiricism', in Readings in the Philosophy of Language, pp. 63-74. 41 N. Goodman, Problems and Projects, The Bobbs-Merill Company, New York, 1972, pp. 221-238. See also my reply in Semantic Theory, pp. 234--243. Goodman has, to my knowledge, never responded to this criticism. 92 j. Hintikka, Models for Modalities, Humanities Press, New York, 1969, pp. 87-88. Hintikka goes on to claim, "However it is understood, it seems to me in any case completely hopeless to try to divorce the idea of the meaning of a sentence from the idea of the information that the sentence can convey to a hearer or reader, should someone truthfully address it to him. Now what is this information ? Clearly, it is just information to the effect that the sentence is true, that the world is such as to meet the truth-conditions of the sentence" (p. 88). It should be observed that this is one of the most amazing uses of the word 'just' to occur in the literature of philosophy. Yet Hintikka gives no indication of how he hopes to acquire the omniscience required to construct such a theory of sentential information. Nor does he explain to us why it is "completely hopeless" to try to characterize a somewhat narrower notion of meaning. 4a F. Katz and J. J. Katz, 'Is Necessity the Mother of Intension ?' The Philosophical Review 86 (1977), pp. 86-96. 44 Meaning and Necessity, pp. 56-59. 45 R. Md'ntague, Formal Philosophy, R. H. Thomson, ed., Yale University Press, New Haven, 1974, pp. 204f.; D. Lewis, 'General Semantics', Semantics of Natural Language, pp. 169-218. t5 A. Church, 'Intensional Isomorphism and the Identity of Belief', in Philosophical Studies 5 (1954), pp. 65-73; also, 'On Carnap's Analysis of Statements of Assertion and Belief', Analysis 10 (1950), pp. 97-99. 47 The failure is admitted by Lewis and others who have pursued this velsion of referential semantics, see 'General Semantics', pp. 182-183. 9e Semantic Theory, pp. 261-292. 49 N. Chomsky, Aspects of the Theory of Syntax, pp. 37-47. 55 Propositional Structure and lllocutionary Force, pp. 59-87. 51 Semantic Theory, pp. 113-116. 59 Ibid., pp. 47-55, and pp. 157-171. 5a 'On Sense and Reference', pp. 157-171. 9t Propositional Structure and lllocutionary Force, pp. 88-117. 55 C. I. Lewis, An Analysis of Knowledge and Valuation, The Open Court Publishing, La Salle, Indiana, 1946, pp. 134-135: " A sense m e a n i n g . . , is a schema; a rule or prescribed routine and an imagined result of applying it which will determine applicability of the expression in question." 56 'Necessity and Criteria', pp. 48-52. 57 'It Ain't Necessarily So', pp. 52-54, and Putnam applies this criticism to TSR in H. Putnam, 'Is Semantics Possible?', Metaphilosophy 1 (1970), pp. 189-201. Detailed replies to Putnam's criticism of TSR are found in my 'Logic and Language: An Examination of Recent Criticisms of Intensionalism', 'A Proper Theory of Names', pp. 9-26, and 'The Neoclassical Theory of Reference'.
108
J E R R O L D J. K A T Z
58 ' A Proper Theory of Names', pp. 34-73. 59 Ibid., pp. 34-73. e0 The fail-safe procedures, namely, the dictionary, introduce a standard of correctness in such cases of reference. Thus, the rationale for claiming that there is both a dictionary and an encyclopedia, apart from internal grammatical considerations like (23), is that the referential use of language requires two independent bodies of information, one to expedite the application of words to things so that it can occur within the constraints of time and opportunity of verification, and one to provide a standard of rightness and wrongness that remains relatively fixed with changes in bel'ef and circumstance. el D. Davidson, 'Truth and Meaning', in Readings in the Philosophy of Language, pp. 480-465, in particular, the discussion on p. 455. 62 Ibid., p. 455. In "sweep[ing] away the obscure 'means that'", Davidson eliminates meaning conditions in favor of extensional conditions. 88 "Logic and Language: An Examination of Recent Criticisms of Intensionalism", pp. 63-76. 84 D. Davidson, 'Thought and Talk', in Mind and Language, ed. S. Guttenplan, Clarendon Press, Oxford, 1975, p. 20. 65 Ibid., p. 21. e6 G. Harman, 'Language, Thought, and Communication', in Language, Mind, and Knowledge, pp. 270-298. e7 'Language, Thought, and Communication', p. 291. 68 G. Harman, Thought, Princeton University Press, Princeton, 1973, pp. 84-111. 69 G. Harman, 'Katz's Cradle', Synthese 32 (1976), pp. 387-394. 7o 'Language, Thought, and Communication', p. 295. 71 Ibid., pp. 292-293. 72 It might seem as if these criticisms of Harman are avoidable if the subscripts are taken not to distinguish particular senses but simply to express the number of distinct senses of a lexical item. On this view, assigning subscripts is like assigning a number expressing the degree of ambiguity of an item. But this is not what Harman means. This use of subscripts would not distinguish the different senses in which a word can be used, which is what Harman introduces subscripts for, and would not permit sense selection over the sentences constituting the background information, which leaves open the question of inference and paraphrase. But the view under consideration does not actually save the account of ambiguity. For it is still not possible to mark the degree of ambiguity above the lexical level. Consider cases like (A) and (B). (A) naked men and women (B) naked nudists and nudes If subscripts are counters, there is no productive way to determine difference in degree of ambiguity in cases like (A) and (B), since, as a consequence of the inclusion of the sense of the modifier in the sense of the second noun in (B), (B) does not have a sense comparable to the 'men who are wearing no bodily covering and women' sense of (A). 73 p. F. Strawson, 'On Referring', Readings in the Philosophy of Language, p. 181. 74 Ibid., p. 184. 75 Propositional Structure and lllocutionary Force, chapter 1. 76 Aspects of the Theory of Syntax, pp. 3-15.
THE THEORY OF S E M A N T I C R E P R E S E N T A T I O N
109
77 For a more detailed discussion, see Propositional Structure andlllocutionaryForee, pp. 16t". 78 One of the principal advantages of this proposal for relating meaning and use is that it permits us to accommodate the facts of use without sacrificing the formalization of logical structure, that is, the formal representation of propositions and hence the formal explanation of valid argumentation in terms of formal statements of the relations between propositions. This consideration provides a further argument for our claim in the text that sentence meaning and utterance meaning have the same representation. If this were not the case, it would not be clear at all how we could explain the arguments people make in their uses of language on the basis of a formal theory of the logical structure of sentences and their implication relations. ~9 p. Grice, 'Meaning', The Philosophical Review 66 (1957), pp. 377-388, and related papers, including, in particular, P. Grice, 'Conversation and Logic', Syntax and Semantics: Speech Acts, eds. P. Cole and J. Morgan, Academic Press, New York, 1975, pp. 41-58; J. Searle, Speech Acts, Cambridge University Press, Cambridge, 1969. a0 L. Wittgenstein, The Blue and Brown Books, Basil Blackwell, Oxford, 1958, pp. 27-28. el p. Ziff, 'On Grice's Account of Meaning', Analysis 28 (1967), pp. 1-8, and N. Chomsky, Reflections on Language, Pantheon Books, New York, 1975, ch. 2. a2 Propositional Structure and lllocutionary Force can be viewed as an extended argument for the thesis. 8a For example, in connection with the question of performativeness, see K. Bach, 'Performatives are Statements Too', Philosophical Studies 28 (1975), pp. 229-236, and my reply in Propositional Structure and lllocutionary Force, pp. 175-177. 84 S. Schiffer, Meaning, Oxford University Press, Oxford, 1972.
s5 Ibid., p. 7. 88 The notions of normal and ordinary, as I am interpreting them, both involve reference to a norm in the sense indicated in the text. (Abnormal uses and out-ofthe-ordinary ones depart from the norm.) If these terms were interpreted in some weaker way, say, having just to do with regularity, I would want to deny the truth of Schiffer's claim.
e7 Meaning, p. 7. Manuscript received 3 January 1978
