Donald Davidson: Truth, Meaning and Knowledge (Routledge Studies in Twentieth Century Philosophy)

Donald Davidson

Donald Davidson: Truth, meaning and knowledge discusses the philosophical views of Donald Davidson, one of the outstanding and most influential living philosophers. Since the 1960s Davidson’s views have provoked and inspired, and his intellectual dynamism has revolutionised the philosophy of mind, epistemology, the philosophy of language and semantics. Donald Davidson is known not only for his rigorous and often controversial thinking but also for his willingness to engage in dialogue and re-evaluate his ideas and paradigms. The present collection is testimony to his spirit of openness, and contains essays by leading academics on the major themes of Davidson’s philosophy, complete with individual responses by the philosopher himself. The essays address the main ideas of Davidson’s theories of language and epistemology with their implications for ontology and philosophy of mind. The book begins with polemical essays on the central problem of truth, one of the most controversial philosophical issues. The core of the controversy can be formulated in the question: is truth the goal of enquiry? This query has recently been raised by Richard Rorty, to whom Davidson has penned a personal reply. In the second part of the anthology, eminent contributors discuss Davidson’s semantic programme with regard to meaning, truth and interpretation, including, among others, the questions of learnability, understanding, holism and indeterminacy of meaning. Semantic questions and their applications are further presented in the third part, while the final part focuses on the question of knowledge and mind, discussed with reference to Davidson and the views of Willard V.Quine, John McDowell and Peter F.Strawson. The discussions with Davidson make this a unique collection which sheds light on a wide area of the philosopher’s thinking. Moreover, this book is an invaluable reference source thanks to its complete bibliography of Donald Davidson’s works. Urszula M.Żegleń is Professor of Philosophy at the Nicholas Copernicus University in Toruń, Poland.

Routledge Studies in Twentieth Century Philosophy

1. The Story of Analytic Philosophy Plot and Heroes Edited by Anat Biletzki and Anat Matar 2. Donald Davidson Truth, meaning and knowledge Edited by Urszula Żegleń

Donald Davidson Truth, meaning and knowledge

Edited by Urszula M.Żegleń

London and New York

First published 1999 by Routledge 11 New Fetter Lane, London EC4P 4EE This edition published in the Taylor & Francis e-Library, 2005. “To purchase your own copy of this or any of Taylor & Francis or Routledge’s collection of thousands of eBooks please go to www.eBookstore.tandf.co.uk.” Simultaneously published in the USA and Canada by Routledge 29 West 35th Street, New York, NY 10001 © 1999 Urszula M.Żegleń, selection and editorial matter; individual chapters, the contributors All rights reserved. No part of this book may be reprinted or reproduced or utilised in any form or by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying and recording, or in any information storage or retrieval system, without permission in writing from the publishers. British Library Cataloguing in Publication Data A catalogue record for this book is available from the British Library Library of Congress Cataloging in Publication Data Donald Davidson: truth, meaning and knowledge/ edited by Urszula M.Żegleń (Routledge studies in twentieth century philosophy; 2) Includes biographical references and indexes. 1. Davidson, Donald. 2. Truth. 3. Meaning (Philosophy). 4. Knowledge, Theory of. I. Żegleń, Urszula M. II. Series. B945.D384D66 1999 191–dc21 98–35259 ISBN 0-203-98460-9 Master e-book ISBN

ISBN 0-415-18904-7 (Print Edition)

Contents

Acknowledgements Introduction to reading Davidson KIRK LUDWIG AND URSZULA M.ŻEGLEŃ

vii 1

PART I Polemics

14

1

Is truth a goal of inquiry? Discussion with Rorty DONALD DAVIDSON

15

2

Davidson on deflationism PAUL HORWICH

18

PART II Meaning, truth and interpretation 3

4

5

Theories of meaning, truth and interpretation KIRK LUDWIG

24

REPLY TO KIRK LUDWIG DONALD DAVIDSON

43

How a truth theory can do duty as a theory of meaning GABRIEL SEGAL

45

REPLY TO GABRIEL SEGAL DONALD DAVIDSON

53

Radical interpretation and compositional structure PETER PAGIN

55

REPLY TO PETER PAGIN DONALD DAVIDSON

68

PART III Semantics and its applications 6

23

71

From semantics to ontology, via truth, reference and quantification STEPHEN NEALE

72

REPLY TO STEPHEN NEALE

82

vi

DONALD DAVIDSON 7

8

Semantics for quotation HERMAN CAPPELEN AND ERNIE LEPORE

85

REPLY TO HERMAN CAPPELEN AND ERNIE LEPORE DONALD DAVIDSON

96

Samesaying REINALDO ELUGARDO

97

REPLY TO REINALDO ELUGARDO DONALD DAVIDSON PART IV Knowledge and mind 9

10

11

114 116

McDowell on Quine, Davidson and epistemology ROGER F.GIBSON

117

REPLY TO ROGER F.GIBSON DONALD DAVIDSON

128

Davidson and the new sceptical problem ANITA AVRAMIDES

130

REPLY TO ANITA AVRAMIDES DONALD DAVIDSON

148

General comments DONALD DAVIDSON

151

Bibliography

155

Index

174

Acknowledgements

Donald Davidson’s philosophy has achieved its success not only by its new and inspiring ideas, but also thanks to its openness to criticism and discussion, which adds fresh insight to his thought, allows for revision of his views and inspires further inquiries. The current collection is just such a vivid example of doing philosophy. The preparation of this work was possible thanks first of all to Professor Donald Davidson, who acquainted himself with the essays and entered into a discussion of his views. Professor Davidson has also provided a complete bibliography of his writings, which will be very useful for anyone studying his philosophy. I think that I can express my special thanks to Professor Davidson on behalf of all potential readers of this book. I am also very grateful to Kirk Ludwig, who has been assisting me in my work on Davidson’s philosophy for a long time. He deserves special thanks for his contribution to this volume. My grateful thanks are due to my friend Tadeusz Szubka who helped me very much and in many ways in preparing this book. I owe thanks to the editors at Routledge, Adrian Driscoll and Tony Bruce, and to the anonymous referees of the manuscript. I also wish to thank my colleagues from Nicholas Copernicus University in Toruń (Poland), Marek Nasieniewski and Scott Thompson, and of course all the contributors, whose deliberations have determined the content of the book. Finally I should express my thanks to Ernie Lepore, who first encouraged me to undertake this project. Urszula M.Żegleń

Introduction to reading Davidson Kirk Ludwig and Urszula M.Żegleń

Davidson’s place in contemporary philosophy Donald Davidson has been one of the most influential of contemporary AngloAmerican philosophers since the appearances of ‘Actions, Reasons and Causes’ in 1963, and ‘Truth and Meaning’ in 1967, over thirty years ago. From these two starting points, one concerned with the nature of explanations of intentional behaviour, and the other with understanding what it is to speak a language, Davidson has developed a unified picture of the nature of rationality, language and thought, and their relation to the world, which is articulated through an investigation of the procedures of the radical interpreter—an interpreter of another’s speech whose epistemic position does not presuppose any detailed knowledge of a speaker’s language or psychological attitudes. He has pursued this central idea in chasing down a wide range of traditional philosophical problems with admirable skill, energy, ingenuity and even elegance. Davidson has revolutionised work in semantics, and made seminal contributions to epistemology and philosophical psychology; in the latter area especially with regard to his famous thesis of the anomalousness of the mental, which has important implications for basic methodological issues in the sciences of the mind. This volume brings together a number of essays addressing a variety of aspects of Davidson’s philosophical programme, and comments by Davidson on a number of the essays in the volume. This introduction will set out the framework of Davidson’s project, and canvass some of the most important contributions he has made to the philosophy of language, action, mind and epistemology, with an eye to providing the background necessary for nonspecialist readers to profit from the more searching essays that follow. We do not attempt to canvass the full range of Davidson’s contributions to contemporary discussion, but rather focus on those that bear most centrally on the following essays. We deal with Davidson’s work on formal semantics, the project of radical interpretation, and some central contributions to the philosophy of action, epistemology and the philosophy of mind. The final section provides an overview of the volume.

2 INTRODUCTION TO READING DAVIDSON

Truth and meaning Davidson’s most important contribution to the study of meaning, and one which is crucial to understanding the form in which he casts the project of radical interpretation, is the proposal that ‘truth theories can do duty as meaning theories’ (Davidson 1984a, p. xiv). The suggestion is first made in ‘Truth and Meaning’ at the end of a discussion of the difficulties in providing a compositional meaning theory for a natural language—a theory which aims to show how the meanings of complex expressions of the language depend on the meanings of their significant parts. Though we are finite beings, natural languages contain an infinite number of non-synonymous sentences. This means that our understanding of our languages must rest on the understanding of a finite number of semantic primitives and rules. Any adequate account of a natural language construction must show how it can be integrated into a complete compositional meaning theory for the language. A compositional meaning theory for a natural language L aims to provide a specification of the meaning of each of the sentences of L by appeal to assignments of meaning to each of a finite number of semantic primitives and rules. The language in which the theory is stated is the metalanguage, and the language the theory is about is the object language. Ideally, we would wish to have a formal theory which issued in theorems of the form (M), (M)s means in L that p where ‘s’ is replaced by a description of a sentence of L in terms of how it is constructed from primitive word units of L, and ‘p’ is replaced by a sentence of the metalanguage which is synonymous with the sentence mentioned on the lefthand side of (M). Knowledge of the M-sentence for a sentence of L suffices to understand it. The derivation of the M-sentence from an assignment of meanings to the significant parts of the sentence shows how its meaning depends on them. The difficulty is to see how to achieve this. In ‘Truth and Meaning’, Davidson argues that the most straightforward route to providing a compositional meaning theory is not open to us. A formal theory using the usual quantificational apparatus to prove theorems of the form (M) must treat ‘means in L that’ (or ‘means in L’) as a relational predicate; that is, it must treat the terms that replace ‘p’ or ‘that p’ as singular referring terms. He argues this has an unacceptable result, given two prima facie plausible principles. The first principle is that if a singular term in a complex singular term is replaced by another with the same referent, the complex singular term obtained thereby will have the same referent as the original. For example, assuming that definite descriptions are singular terms, if ‘Mark Twain’ co-refers with ‘Samuel Clemens’, then ‘the father of Mark Twain’ co-refers with ‘the father of Samuel Clemens’. The second is that logically equivalent singular terms have the same referent. Two singular terms can be taken to be logically equivalent if they co-refer on all interpretations of

KIRK LUDWIG AND URSZULA M.ŻEGLEŃ 3

non-logical terms in the language. If we treat sentences as singular terms, then by the second principle, Davidson holds that we must take logically equivalent sentences to have the same referent. Thus, for example, letting ‘S’ stand in for any sentence, (1) and (2) have the same referent: (1) S (2) {x: x=x}={x: x=x & S} Let ‘R’ stand in for any sentence alike in truth value to ‘S’. Then by our first principle, since ‘{x: x=x & S}’ co-refers with ‘{x: x=x & R}’ (if ‘S’ and ‘R’ are both true, then both terms refer to ‘{x: x=x}’, and if both are false, then to the empty set, (2) refers to the same thing as (3). By the same reasoning that showed that (1) and (3) co-refer, we can show that (3) co-refers with (4)). (3) {x: x=x}={x: x=x & R} (4) R We are led to the conclusion that ‘S’ and ‘R’ co-refer if they have the same truth value. But sentences alike in truth value may differ in meaning. Thus, treating sentences as referring terms will not ensure that true theorems of the form (M) use a sentence on the right that translates the one mentioned on the left, as required. The argument can be extended to terms of the form ‘that p’ on the assumption that terms of the form ‘that p’ and ‘that q’ are logically equivalent if ‘p’ and ‘q’ are logically equivalent. The argument then goes through as before with ‘that’ added before each of (1) through (4). This argument has been both influential and controversial.1 The importance of Davidson’s suggestion for how to get around quantifying over intensional entities does not depend on the success of his arguments against traditional approaches, and has virtues over those approaches even apart from technical difficulties they may encounter. Davidson makes novel use of a requirement that Tarski introduced on an extensionally adequate recursive definition of a truth predicate for a language (henceforth, ‘truth theory’). Tarski observed that a (consistent) recursive definition of a predicate ‘is T’ for a formal language which had as theorems all sentences of the form (T), (T) s is T iff p where ‘s’ is replaced by a structural description of an object language sentence, and ‘p’ is replaced by a metalanguage sentence that translates it, will be an extensionally adequate definition of a truth predicate for the language. The requirement that an extensionally adequate definition should have all such sentences, T-sentences, as theorems, is Tarski’s Convention T. That is, ‘is T’ will ha5ve in its extension all and only the true sentences of the language. This is so because of the relation between truth and meaning illustrated in schema (S).

4 INTRODUCTION TO READING DAVIDSON

(S) if s means in L that p, then s is true in L iff p If we know that what we put in the place of ‘p’ in (T) translates s, then we know the antecedent is satisfied, and so that s is true. Thus, Convention T requires ‘is T’ to have in its extension all and only true sentences of the object language. Davidson observed that if a theorem of a truth theory for L were a T-sentence, then ‘means in L that’ could be substituted for ‘is T iff’ in the theorem without change of truth value. Thus, if one had a truth theory which one knew met Convention T, and one could pick out formally the T-sentences, one would have a formal means of arriving at (M)-sentences for each sentence of the object language. Moreover, since a truth theory enables us to prove T-sentences on the basis of assignments of meaning giving satisfaction conditions to primitive elements of the language, it will have a claim to show thereby how the meanings of complex expressions depend on the meanings of their parts with no more ontological resources than are required for the theory of reference. Truth and interpretation The proposal to use a truth theory as a compositional meaning theory is complicated in ‘Truth and Meaning’ by Davidson’s desire to provide constraints on a truth theory, which themselves draw on no semantic concepts and which nevertheless suffice for the theory to meet an analogue of Convention T. The aim is to answer by way of a truth theory both the question how the meanings of complex expressions depend on those of their significant parts and the question what it is for the primitive expressions of the language to mean what they do. Davidson hoped an extensionally adequate theory would suffice, given that it would have to project correctly the truth conditions of context sensitive terms. This hope proved ill-founded (Foster 1976; Loar 1976). Davidson’s response was to develop more fully a theme touched on briefly in ‘Truth and Meaning’, namely, the idea that a truth theory for a natural language should be treated as an empirical theory, which must be confirmed for a speaker or linguistic community. The theme is developed in the description of the project of radical interpretation. A radical interpreter interprets another’s speech without prior knowledge of his subject’s language or detailed knowledge of his attitudes. Ultimately, the radical interpreter is responsible for confirming an interpretation theory by appeal only to facts about the speaker’s dispositions to respond to events in his environment. An interpretation theory provides interpretations of a speaker’s words and sentences, relativised as necessary to occasions of utterance, and assignments of contents to the speaker’s attitudes. The study of the project of radical interpretation aims to illuminate what it is for words to mean what they do by showing how one can confirm, and what must be true to confirm, an interpretation theory for a speaker relying on evidence that does not presuppose knowledge of what the theory is about. An examination of the

KIRK LUDWIG AND URSZULA M.ŻEGLEŃ 5

project of radical interpretation thus illuminates the concepts of the theory of meaning by relating them both to related concepts with which they are coordinated in a theory of interpretation for a speaker, particularly those drawn from a theory of agency, and the concepts deployed in a description of the evidential base for the theory. In the present volume, Davidson’s essay recommends the same approach to understanding the concept of truth, in lieu offering a definition or definition substitute. The concepts of meaning, truth and the psychological attitudes are too basic for us to expect illuminating definitions one by one in more basic terms. It is rather in understanding their interplay, particularly in the context of confirming a theory of truth for a speaker’s language, that we can expect deeper understanding. The standpoint of radical interpretation is motivated by the thought that mastery of a language is the acquisition of ‘a social art,’ as Quine puts it. This means that ‘The semantic features of language are public features. What no one can, in the nature of the case, figure out from the totality of the relevant evidence cannot be part of meaning’ (Davidson 1979c/1984a, p. 235). This provides ground for the claim that speakers are by nature radically interpretable. This in turn provides reason to think that the features of speakers and languages which a radical interpreter must find in his subjects are constitutive features. Many of Davidson’s most important theses about meaning and the propositional attitudes can be seen as consequences of adopting the standpoint of the radical interpreter as methodologically fundamental in investigating meaning and related matters. The use of a recursive truth theory as a compositional meaning theory motivates adopting it as the form in which an interpretation theory is cast. The investigation of the radical interpreter’s procedure then becomes that of how to confirm a truth theory for a speaker which meets an appropriate analogue of Convention T.Davidson helps himself to knowledge of a speaker’s hold-true attitudes (which we may assume the radical interpreter can identify on the basis of more primitive evidence). A radical interpreter can discover that a speaker holds true certain sentences in response differentially to various conditions obtaining or events occurring in his environment. A speaker’s hold-true attitudes are a product of what he believes about his environment, and what he believes his sentences mean. We have no direct knowledge of either of these. Davidson’s solution to this difficulty is to introduce the principle of charity. In the present context, this is the assumption that the speaker is mostly right about his environment. By assuming that the speaker is mostly right about his environment, and that, in identifying the conditions under which he holds true sentences on the basis of those beliefs, we have identified those conditions that have prompted the speaker’s beliefs, we can infer that the sentence we would use in our language to specify the conditions under which the speaker holds true a sentence also interprets it. This allows us to move from evidence in the form of lawlike correlations between a speaker’s hold-true attitudes to a provisional identification of T-sentences in a truth theory for his language. The interpreter constructs a truth theory which has as canonical theorems (intuitively, theorems

6 INTRODUCTION TO READING DAVIDSON

which draw only on the axioms of the theory) the provisionally identified Tsentences, and tests the theory against further evidence until a best overall fit has been achieved, that is, assignments of meanings and attitude contents are made holistically in the light of the evidence, in accordance with norms governing their interrelations. This provides the basis for Davidson’s proposal about what further non-semantically specified conditions a truth theory must meet in order to meet an analogue of Convention T, namely, that it must ‘optimally fit evidence about sentences held true by native speakers’ (Davidson 1973c/1984a, p. 139).2 Philosophy of action In the project of radical interpretation, the interpreter brings to bear a theoretical framework which employs a theory of agency as well as the framework for a compositional meaning theory provided by a recursive truth theory. The theory of agency constrains assignments of meanings to sentences, since they at the same time provide assignments of contents to the speaker’s propositional attitudes. Davidson’s work in the philosophy of action provides the theory of agency employed. We sketch here a few of the general features of that theory. Actions are bodily movements, particular datable events. ‘Actions, Reasons and Causes’ argues famously that action explanations (‘rationalisations’, in Davidson’s technical phrase), which consist of citing a primary reason (a pro attitude toward an action of a certain kind, and a belief that the action is of that kind) for the action, are causal explanations. The concept of an action itself is the concept of a bodily movement (construed broadly) that is caused (in the right way) by a primary reason that is the agent’s motivation for the action. The conceptual tie between action and rationalisation shows that every action must in a minimal sense be rational: there must be some primary reason in the light of which the action was seen by the agent as justified. An action is intentional under a description which reveals a motive of the agent in undertaking it. Every action is intentional under some description, and unintentional under others. The conceptual connection between actions described as such and reasons that motivate them had been seen as a reason to deny that action explanations were causal explanations. Causal relations are contingent, but the connection between something being an intentional Aing and the agent having a motive to A is not. However, the appearance of conflict is illusory. As Davidson pointed out, logical relations hold between descriptions of events, not between the events themselves. It would be absurd to deny that the cause of A caused A, even though the descriptions of these events are logically related. (This thought plays an important role also in Davidson’s argument for anomalous monism—see below.) There is, then, no reason to deny that rationalisations cite reasons as causes of actions. Furthermore, unless one treats reasons as causes of actions, there will be no way to explain how a pro attitude/belief pair, in the light of which an action is reasonable, may fail to explain it. While one may know that buying stamps will help to secure employment for postal workers, and have a pro attitude toward

KIRK LUDWIG AND URSZULA M.ŻEGLEŃ 7

actions promoting full employment, one may still buy stamps simply because one wishes to mail a letter. The connection between action, belief, pro attitudes and intention shows that the concepts of these are applied in coordination with one another (they come in ‘matched sets’, as Davidson aptly puts it). We could not imagine an agent which lacked beliefs, pro attitudes and intentions. To recognise something as an agent, we must recognise patterns in the attitudes we see the agent as having and the actions he undertakes which make him intelligible as an agent. To be an agent is ipso facto to be rational. This does not mean that agents must be perfectly rational. It means that what gives substance to the claim that something is an agent, and so has psychological attitudes, is that its behaviour can be displayed as the product of agency, and that requires seeing it as by and large describable in accordance with the canons of practical rationality and epistemic rationality. An agent’s preferences must be by and large transitive, for example, or we will not be able to see him as an agent. And an agent’s beliefs must be by and large consistent, or we will not be able to see him as a believer. These principles are constitutive of our subject matter. Nothing which cannot be described as by and large in conformity with the canons of practical and epistemic rationality can be coherently described as an agent. If an object cannot be described as an agent behaving in accordance with the norms of rationality, then we do not call it irrational, but non-rational. Non-agents are not capable of acting irrationally, because they are not capable of acting at all. The identification of an irrational act or attitude can take place only against a background of rationality. ‘Irrationality’, as Davidson puts it, ‘is not mere lack of reason, but a disease or perturbation of reason’ (Davidson 1982a/1985, p. 476).3 The attitudes are intrinsically holistic, according to Davidson, in at least two senses. First, they are the attitudes of an agent, and an agent must obey the norms of rationality for the most part to be identifiable as an agent. This requires that attitudes be attributed to agents in appropriate patterns. Second, to attribute an attitude with a particular content to an agent, one must attribute to the agent an unlimited number of attitudes with related contents. For example, we would not attribute to someone the belief that the president of Poland is an elected official unless we were also willing to attribute to him belief in a large number of related propositions, such as that Poland is a country, that it has a republican form of government, that in elections people vote for candidates, and so on. There is, Davidson says, no reason to insist on any particular related beliefs as necessary for having a given belief: ‘No particular list of further beliefs is required…but some appropriate set of related beliefs must be there’ (Davidson 1977a/1984a, p. 200). That the concepts of the attitudes are inseparable from the concept of an agent shows they are causal concepts. Our grasp of a casual concept requires us to think of particulars it subsumes as actually or potentially standing in causal relations to other things. The concept of sunburn, for example, is a causal concept, the concept of a burn caused by the exposure to the sun. We can call these

8 INTRODUCTION TO READING DAVIDSON

‘backward-looking causal concepts’. Similarly, the concept of murder is the concept of an event which causes a killing, and the concept of a soporific is the concept of something that induces sleep when taken internally. We can call these ‘forward-looking causal concepts’. Our discussion above shows the sense in which attitudes resemble forward-looking causal concepts, although the conditions under which they lead to action are more complex than the conditions under which a soporific leads to sleep, and involve essential reference to coordinated psychological states. The concept of action is a backward-looking causal concept. Epistemic psychological states are also clearly backward-looking causal concepts. To remember seeing an owl in the barn requires that one have a state whose etiology includes a certain kind of causal interaction with an owl in a barn. In the next section, we will consider one way in which Davidson has employed reflection on the procedure of the radical interpreter to argue that all psychological attitudes are backward-looking as well as forward-looking causal concepts. Epistemology The assumption that the standpoint of the radical interpreter is methodologically basic in investigating language and the propositional attitudes is the key to understanding Davidson’s conception of our epistemic position with respect to other minds, to things in our environments and to our own minds. The basic line of thought may be represented as follows. To hold that the standpoint of the radical interpreter of another speaker is basic in our understanding of language and the propositional attitudes is to hold that we would not be speakers or have propositional attitudes unless we were radically interpretable. From this it follows that what a radical interpreter must assume to succeed in interpretation is constitutive of his subject matter. Davidson argues that having knowledge in each of the domains mentioned above, other minds, our environments and our own minds, is a necessary condition on being interpretable. (Davidson 1975b; 1982a; 1983a; 1984c; 1987a; 1988d; 1989a; 1990e; 1990f/1991; 1991a; 1992b). First, as we observed above, on Davidson’s account, in order for the radical interpreter to bring his observational evidence, represented by lawlike biconditionals stating the conditions under which a speaker holds true sentences, to bear on his subject—the meaning of the speaker’s sentences and the contents of his attitudes—the interpreter must assume that the speaker’s beliefs about his environment are largely correct. On the assumption that the interpreter can in principle succeed, what the interpreter must assume is constitutive of his subject matter. It is therefore constitutive of speakers that they are largely right about their environments. This provides a transcendental guarantee that we are mostly right about the world, yet one which does not make the world mind-dependent. Rather, since what the interpreter goes on in identifying our attitudes is what typically causes them, reflection on the procedures of the radical interpreter leads to the conclusion that what we think is logically dependent on what typically

KIRK LUDWIG AND URSZULA M.ŻEGLEŃ 9

causes our beliefs. This is the sense in which our beliefs, and other propositional attitudes, whose contents must be coordinated with the contents of our beliefs, are backward-looking causal concepts, like jet-lag or hay fever: something has them only if it has had a certain kind of causal history. This transcendental guarantee is a guarantee of the truth of most of our beliefs about our environments, not of our having knowledge. But once it is in place, our usual procedures for justifying beliefs will suffice for determining which are justified. In effect, the guarantee that most of our beliefs about our environments are true allows us to test the truth of any given belief by the degree of its coherence with the rest of our beliefs. Second, it is clearly a condition on successful radical interpretation that the interpreter can come to know the minds of speakers on the basis, ultimately, of behavioural evidence. A transcendental guarantee of the success of radical interpretation therefore guarantees the knowability of the minds of other speakers. This is not to say that sentences about psychological attitudes are reducible to sentences about behaviour. Applications of psychological predicates supervene on, but are not reducible to, applications of behavioural predicates. We will consider in the next section Davidson’s argument for their nonreducibility. Third, in order to interpret another speaker, we must also assume that he knows what the meanings of his words are, and what he believes. This can be seen from reflecting on the crucial role that the principle of charity plays in interpretation. The principle of charity says that most of a speaker’s beliefs about his environment are true. In applying the principle of charity to arrive at interpretations of a speaker’s words, we must assume that the conditions under which he holds true sentences are the conditions which his beliefs are about, and that he holds true those sentences because he knows what he believes and knows what the sentences of his language mean. That is, we must represent his holdtrue attitudes as arrived at on the basis of knowledge of the meanings of his sentences and the contents of his beliefs. Because this is grounded conceptually in reflection on the radical interpreter’s procedure, it is supposed to explain also why we do not go on evidence in our own case when we attribute beliefs and desires to ourselves, and yet do go on behavioural evidence in attributing attitudes to others. It is clear that the argument for the knowability of the external world, other minds and our own minds rests on the assumption that the concepts we deploy in interpretation are necessarily deployed primarily from the third-person standpoint, that is, that we could not have the concepts we do in fact have, if we conceived of them as grounded in the first instance in their applications to ourselves.

10 INTRODUCTION TO READING DAVIDSON

Philosophy of mind We have touched already on a number of important themes in Davidson’s work in the philosophy of mind. In this section, we discuss Davidson’s thesis of anomalous monism, a form of non-reductive materialism. Anomalous monism is the view that every event (and every object) is a physical event (or object) and that there are no strict psychological laws, i.e., no strict laws expressed using psychological predicates. If there are no strict psychological laws, then there are no strict psycho-physical laws, laws connecting events under mental descriptions with events under physical descriptions, and consequently there can be no strict lawlike correlations between mental event types and physical event types. Davidson’s monism, then, maintains every mental event is token identical with a physical event, but denies that falling under any mental type nomically suffices for an event to fall under any physical type, and vice versa. Anomalism rules out the latter. Davidson’s argument for anomalous monism rests on three assumptions (Davidson 1970b/1980a, p. 208): 1 Some mental events causally interact with physical events (the Principle of Causal Interaction). 2 Events related as cause and effect must fall under a strict law (the Principle of the Nomological Character of Causality). 3 There are no strict laws on the basis of which mental events can be predicted and explained (the Anomalism of the Mental). A strict law need not be deterministic, but is rather the limit reached in ‘a comprehensive closed system guaranteed to yield a standardised, unique description of every…event couched in a vocabulary amenable to law’ (Davidson 1970b/1980a, pp. 223–4). In ‘Mental Events’, in which Davidson first presented his argument, he did not argue for the Principle of Causal Interaction or the Principle of the Nomological Character of Causality. Davidson’s arguments in ‘Actions, Reasons and Causes’, and other articles on the philosophy of action, can be seen as part of an argument for the common sense view that the Principle of Causal Interaction is true. In a more recent paper, ‘Laws and Cause’ (Davidson 1995e), he has argued for the Principle of the Nomological Character of Causality. We will be primarily concerned with the most controversial of the assumptions, the Anomalism of the Mental, for which Davidson offers an intriguing, but notoriously elusive argument. Anomalous monism does not follow from 1–3 by themselves, but only as supplemented by a number of auxiliary assumptions. In particular, we need to assume that: 4 If there are no strict psychological laws, then all strict laws are physical laws;

KIRK LUDWIG AND URSZULA M.ŻEGLEŃ 11

5 If any event has a description under which it is subsumed by a strict physical law, then it is a physical event; and 6 Every mental event causally interacts (perhaps through a causal chain) with a physical event. Assumption 1 tells us that some mental events causally interact with physical events. Assumption 2 tells us that those pairs of events must be subsumed by a strict law. Assumption 3 tells us that the law does not contain mental predicates. With 4 this tells us that the law subsuming those pairs of events is a strict physical law. Thus, the mental event has a physical description of the sort that appears in a fundamental physical law, which, by 5, suffices for it to be a physical event. This shows that some mental events are token identical with some mental events, though by 3 there are no type identities between mental and physical events. Since 6, with 1, 2 and 3, secures that every mental event is subsumed by a strict causal law, it will follow that every mental event is token identical with a physical event. Assumptions 4 and 5 are implicit in Davidson’s discussion. Davidson does not argue for 6, but he recognises that it is required to derive the token-token identity theory. The argument for premise 3, the Anomalism of the Mental, turns on the claim that the physical and the mental vocabularies represent families of concepts which are governed by different sets of constitutive principles which determine that their application to objects is responsible to different sorts of constraints. This thought is already familiar in the case of the mental from our discussion above of Davidson’s work in the philosophy of action. Attributions of psychological attitudes to agents are responsible to the global constraint of taking the agents to whom they are attributed to be largely rational, and, in the context of radical interpretation, largely right about their environments. Attributions of physical predicates to objects, on the other hand, are not responsible to the same sorts of constraints. According to Davidson, ‘Lawlike statements are general statements that support counterfactual and subjunctive claims, and are supported by their instances’ (Davidson 1995e, p. 217). A general statement connecting events can be lawlike, then, only if its instances can give one reason to think that so far unobserved instances of the first type of event will be followed by instances of the second. The assumption that Davidson needs for the argu ment to go through is that when the predicates connected in a general statement are drawn from different families of predicates, i.e., families of predicates whose constitutive principles make their application responsible to different sorts of constraints, instances of the generalisation give one no reason to think that so far unobserved instances of the first type of event will be followed by instances of the second.

12 INTRODUCTION TO READING DAVIDSON

Contents of this volume The book is divided into four parts. The first, Polemics, contains essays discussing Davidson’s views in relation to those of Rorty and Horwich. The second, Meaning, truth and interpretation, deals with fundamental questions about Davidson’s theory of language, particularly with truth, meaning and interpretation. The third, Semantics and its applications, is more technical in character, and considers applications of Davidson’s semantics to ontology, and to the semantic analysis of quotation. The fourth part, Knowledge and mind, takes up the epistemological implications of Davidson’s views, particularly in regard to coherentism, scepticism, and knowledge of one’s own and other minds. Polemics begins with a discussion by Davidson of the question, ‘Is truth a goal of inquiry’, which has been raised recently in regard to his views by Richard Rorty (Rorty 1995). Paul Horwich continues the discussion with a response to Davidson’s discussion of deflationism about truth in ‘The Folly of Trying to Define Truth’ (Davidson 1996b). The second part of the anthology focuses on Davidson’s views on truth, meaning and interpretation. The guiding essay of this part is Kirk Ludwig’s, Theories of meaning, truth and interpretation. Ludwig distinguishes two projects in Davidson’s semantic programme, a modest and ambitious project, and argues that the ambitious project is required for Davidson to reach some of his most celebrated conclusions, but that this requires an a priori argument for the possibility of radical interpretation in a form which is unsupported by any intuitively compelling considerations about the publicity of language. Gabriel Segal in, How a truth theory can do duty as a theory of meaning, argues that the constraints on an adequate truth theory proposed by Davidson are neither necessary nor even sufficient for it to be interpretive. The question of radical interpretation is taken up in Peter Pagin’s chapter, Radical interpretation and compositional structure. Pagin also considers some connected questions about learnability, understanding of new utterances, holism and indeterminacy. The third part on semantics and its applications begins with Stephen Neale’s chapter From semantics to ontology, via truth, reference and quantification. This chapter focuses mainly on the problem of reference. Neale gives an axiomatisation of this notion and argues that a theory of truth treated as a theory of meaning cannot be constructed without making use of some notions of reference. According to Neale, the notion of reference given in his axioms (which pair bits of language with bits of the world) is not in conflict with Davidson’s semantics. Some technical questions connected with utterances and their interpretations are discussed by Herman Cappelen and Ernie Lepore in their chapter Semantics for quotation. They analyse direct, indirect, mixed and pure quotation. They offer a unified account of these constructions by extending Davidson’s account of quotation, and compare the result with accounts by Carnap, Church, Quine and others. The topic is continued by Reinaldo Elugardo in his chapter, Samesaying. Elugardo discusses Davidson’s paratactic analysis in

KIRK LUDWIG AND URSZULA M.ŻEGLEŃ 13

regard to the objection that T-sentences of an empirically adequate truth theory for a natural language are incapable of bridging the gap between a speaker’s utterance and his audience’s understanding. He argues for two claims about the relation of samesaying considered from a semantic and pragmatic point of view respectively: (1) if samesaying is synonymy, then Davidson’s paratactic analysis assigns the wrong truth-conditions to indirect speech reports; (2) if it is a pragmatic relation, then the truth-conditions of a sentential utterance may be correctly assigned in an analysed language even if no native speaker could be correctly interpreted (and vice versa). In his conclusion, Elugardo claims that the gap cannot be closed until the line between semantics and pragmatics is redrawn. The fourth part of the book begins with Roger Gibson’s chapter, McDowell on Quine, Davidson and epistemology. He attempts to show that McDowell mischaracterises Quine as a proponent of the Given and Davidson as a proponent of unconstrained coherentism. In particular, he explains Davidson’s destructive argument supporting his ‘coherence thesis’, that only beliefs can be reasons for beliefs, and Davidson’s constructive argument that belief is by nature veridical, which together allow him to avoid the false dilemma that one must give either a false or no answer to the sceptic. Explaining why McDowell rejects the constructive argument, Gibson sketches a Davidsonian reply to show that McDowell mischaracterises Davidson’s coherentism. Anita Avramides’ chapter, Davidson and the new sceptical problem, casts light on the question of the asymmetry between the way in which a knowing subject has access to his mind and the way in which he reaches another mind, and contrasts and critically compares Davidson’s and Strawson’s accounts of this asymmetry. Notes 1 See: Barwise and Perry 1981; Lepore and Ludwig, Typescript for further discussion, and Ludwig in the present volume for some alternative considerations against quantifying over meanings. 2 For a more detailed discussion and a critical evaluation, see Ludwig in the present volume. 3 See: Davidson 1969b; 1974b; 1982a; 1985d; 1985g.

Part I Polemics

1 Is truth a goal of inquiry? Discussion with Rorty Donald Davidson

Aside from abstract objects like sentences or propositions, if such there be, the only things in this world that are true are some utterances and some beliefs. When we say an utterance or a belief is true, we predicate truth of that utterance or belief, so I see no harm in holding that truth is a property: some beliefs and utterances have it and some do not. The word ‘true’ also operates as a truthfunctional connective, as in ‘It is true that Antarctica is not wholly within the Antarctic Circle’: its truth table is like that of negation with the truth values reversed. Because, in this use, it maps true utterances onto true, and false onto false, its function in conversation is rhetorical. Cognitively it is redundant. Realism as I understand it is the view that predicational use of truth can be explained in terms of a relation of correspondence. This would be an interesting claim if anyone could come up with an intelligible and illuminating way of individuating the entities to which true utterances or beliefs correspond, along with an acceptable semantics for talk about such entities. But there is no such account. Until there is, I see no point in declaring oneself a realist, or, for that matter, an anti-realist. I see no difference between a correspondence view of truth and the idea that utterances (or sentences) ‘represent’, except, perhaps, that if one understood the idea, one could talk of what false as well as true sentences represent. But if there is nothing for true sentences to correspond to, neither is there anything for them to represent. On these matters, Rorty has me right. He also has me right on the main contention of his present essay, that truth is not a goal of inquiry. This seems to me correct if understood as Rorty sometimes puts it: it is a goal of inquiry to find substantiating evidence for our beliefs, and there is nothing more we can do in trying to firm up our convictions. We can’t try to do what we know we can’t do. As Rorty says, truth isn’t a norm in addition to the norm (norms?) of justification. But it is a little misleading to say truth isn’t a goal: it’s the same goal, since we are bound to think, not always correctly to be sure, that the more evidence we have the more apt we are to be right. People who invest on the basis of inside information on the stock market, or who consult a tout before the races certainly believe so. Much of the time Rorty appears to endorse the view that truth and justification are identical: pragmatists are ‘suspicious of the distinction between

16 IS TRUTH A GOAL OF ENQUIRY?

justification and truth’. He also thinks justification is a norm. Shouldn’t he conclude that truth is a norm? However, Rorty knows there is a difference between our beliefs being justified and our beliefs being true. As he says, justification is relative to many things: the availability of evidence, the expense of obtaining it, our audience, our standards of evidential support, and so on. Truth is not relative in these ways, and I applaud Rorty for not going along with many of his followers who would say that it is. When Rorty speaks of the ‘cautionary’ use of the concept of truth, I take him to mean that it is often useful to remind people that being justified isn’t necessarily being right. Is it obvious, then, that there is no sense in which the distinction matters to practice? Why remind someone of a distinction if it doesn’t matter? Just the same, Rorty may be justified in calling me a quietist with respect to truth, since I reject correspondence theories, don’t think the idea of representation can be cashed in, and agree that truth is not a norm in addition to justification. I also disavow all other attempts to treat truth as an epistemic concept. If this is sufficient, in Rorty’s eyes, to make me a quietist, so be it. But it’s not clear how far he wants me to go. For example, he finds James’s suggestion that ‘once you understand all about justification, you understand all there is to understand about truth’ (Rorty 1995, p. 282) persuasive. If truth differs from justification, as Rorty seems to allow, then there has to be something about truth you don’t understand when you understand ‘all about’ justification. I certainly think there is a lot more to say about truth. As Rorty recognises, I believe that what I call the content of the concept of truth can be brought out only by relating it to speech, belief and the evaluative attitudes. Rorty confesses that he wobbles between ‘trying to reduce truth to justification’ (ibid.) and embracing some form of minimalism like ‘Tarski’s breezy disquotationalism’ (ibid.). But he knows truth isn’t identical with justification, and he is wrong about Tarski. The Wahrheitsbegriff is about as far from breezy as you can get, and the reason is that it isn’t a form of disquotationalism. You can say if you want that each Tarski-style truth definition has an element of disquotationalism, but none of these definitions is a definition of truth, as Tarski proves. Fortunately, Rorty seems to be willing, perhaps reluctantly, to let me have my more full-blooded view of truth. I hope this is pretty much all that needs to be said. But I have an uneasy sense that it may not be. Rorty wants to explain away my claim that most of our simplest and most basic beliefs are true as ‘saying that most of anybody’s beliefs must coincide with most of our beliefs’ (ibid., p. 286) or that ‘the pattern truth makes is the pattern that justification to us makes’ (ibid.). I agree with these claims, but do not agree that they give my reason for holding that most of our beliefs are true. The beliefs I have in mind are our perceptual beliefs, the beliefs that are directly caused by what we see and hear and otherwise sense. These I hold to be in the main true because their content is, in effect, determined by what typically causes them. This is not the place to expound the argument. The point is that I believe in the ordinary notion of truth: there really are people, mountains,

DONALD DAVIDSON 17

camels and stars out there, just as we think there are, and those objects and events frequently have the characteristics we think we perceive them to have. Our concepts are ours, but that doesn’t mean they don’t truly, as well as usefully, describe an objective reality.

2 Davidson on deflationism Paul Horwich

Deflationism about truth, as I see it, comprises the following theses: 1 That the utility and raison d’être of the truth-predicate is to enable the explicit formulation of schematic generalisations. (For example, ‘pv-p’ becomes ‘Every proposition of the form “pv-p” is true’). 2 That the meaning of this predicate—i.e. our understanding the word ‘true’ the way we do—consists in the fact that our overall use of it stems from our inclination to accept instances of the ‘equivalence schema’: the proposition u (that p) is true if and only if p. 3 That the explanatorily basic facts about truth are those expressed by instances of that schema; therefore truth has no underlying nature—the truth of a proposition does not consist in its possessing some more fundamental property. Donald Davidson endorses none of these deflationary theses. As far as the first is concerned, he thinks that truth is much more than a mere device of generalisation; he regards it as a basic conceptual building block—so fundamental that, in order to possess any concepts at all, one must possess the concept of truth. On the second point he holds that no simple formula can capture the meaning of ‘true’; in particular, my proposal won’t do because the equivalence schema and its instances are incomprehensible: ‘I do not understand [them]’. For that reason he must also reject my third thesis. As far as I know he makes no explicit claim about the constitution of truth, but he says nothing to preclude the possibility that truth has an underlying nature—at least, roughly speaking. On the contrary, his view of truth as a causal, explanatorily potent property would suggest that it is as good a candidate for reduction as any other such property. In what follows I would like to elaborate slightly the deflationary position, to examine Professor Davidson’s main criticisms of it, and to offer a defence against these criticisms. The role of the truth-predicate as a device of generalisation is illustrated by a thesis whose correctness is disputed by realists and anti-realists, namely Any true proposition is verifiable

DAVIDSON ON DEFLATIONISM 19

This is a ‘generalisation’, in some sense, of the particular statement If there are black holes, then it is verifiable that there are But the generalisation cannot be constructed in the normal way, simply by replacing a singular term with a universal quantifier. In order to arrive at it, we first have to deploy an instance of the equivalence schema There are black holes iff it is true that there are black holes in light of which our particular statement is seen to be equivalent to If it is true that there are black holes, it is verifiable that there are or, in other words That there are black holes is verifiable if true which does generalise in the normal way, yielding Every proposition is verifiable if true My first deflationary thesis is that wherever we employ the concept of truth— whether it be in philosophy (as above), in logic (‘All instances of modus ponens preserve truth’), or in everyday life (‘What he said was true’)—it is nothing more than a device of generalisation. If this is correct then, since the truth-predicate’s capacity to fulfil its function derives, as we have just seen, from the equivalence schema, we can infer that our inclination to accept the instances of that schema is the causal/explanatory basis for our overall use of the word ‘true’. But it is plausible to suppose (since our use of a word is symptomatic of what we mean by it) that the meaning of a word is constituted by whatever basic facts explain its overall use. Thus we arrive at my second deflationary thesis: that the meaning of ‘true’ is constituted by the fact that its deployment derives from the equivalence schema. Third, it can be argued, on the basis of theses 1 and 2, that none of the facts formulated with the help of the truth-predicate could be explained by supposing that there is some other predicate ‘F’ with which ‘true’ is co-extensive. For the explanatorily basic facts about truth are instances of the equivalence schema: these instances will suffice, when conjoined with basic facts about other matters, to account for every familiar fact about truth. Moreover, given their a priori status, we can be fairly sure that there will be no reductive analysis resulting in explanations of those instances. Thus, no facts about truth will be explained by a theory of the form ‘(x) (x is true iff x is F)’. But, in general, the basis for supposing that a property, S, is constituted by an underlying property, U, is that the co-extensiveness of ‘S’ and ‘U’ would explain the characteristics of S. It follows therefore that truth is not constituted by some more fundamental property. This, in a nutshell, is the (or at least my) deflationary picture of truth. Let me now turn to what Davidson describes (in Davidson 1996b) as his two principle objections to it—which are focused on the thesis that ‘true’ is implicitly defined by the equivalence schema. One of his objections concerns the relationship between truth and other central psycho-semantic concepts such as meaning and belief. An implication of the deflationary view is that the concepts of belief and of meaning are more basic

20 PAUL HORWICH

than the concept of truth: they can be fully grasped by someone who has no concept of truth; and such a person can then acquire that concept by acquiring the inclination to accept instances of the equivalence schema. But Davidson holds, on the contrary, that truth is more fundamental a concept than either belief or meaning, and that no-one can acquire the concepts of belief and meaning unless they already have the concept of truth. In the case of meaning, Davidson’s anti-deflationary position is presumably based on the success of his own truth-conditional analysis of sentence-meanings, together with the apparent absence of any feasible alternative. For the following reasons, however, this basis for dissatisfaction with deflationism strikes me as weak: 1 As Davidson himself has frequently acknowledged, there are constructions for which we do not yet see how to provide a Tarski-style derivation of truth conditions, and hence how to implement the Davidsonian approach to the compositionality of meaning: n.b. statements of intention, counterfactual conditionals, probability claims, etc. Perhaps the explanation for the difficulty in some of these cases is that a Tarski-style truth definition is simply not possible for all reaches of natural language? 2 Despite concerted efforts over the last thirty years, no-one has yet been able to articulate a conception of ‘truth condition’ (i.e. of ‘u is true if and only if p’) that would be sufficiently strong to constitute the facts about meaning (i.e. ‘u means that p’). I have in mind the well-known, unsolved problem— perhaps insoluble—that although ‘grass is green’ is true if and only if snow is white, it does not mean that snow is white. 3 The use theory of meaning offers a non-truth-theoretic way of specifying the properties in which word-meanings consist. For example (and very roughly speaking) the word ‘true’ arguably means what it does in virtue of our inclination to accept instances of the equivalence schema. The word ‘red’ means what it does in virtue of our disposition to hold true ‘That is red’ when a red surface is under observation. The word ‘and’ means what it does in virtue of our tendency to infer ‘p and q’ from ‘p’ and ‘q’, and vice versa. These examples illustrate that we can find meaning-constituting properties that do not invoke the concept of truth. Rather—just as in Davidson’s own theory of interpretation—attributions of meaning are based upon what sentences are accepted and in what circumstances1. Notice that it will sometimes be a requirement of our meaning what we do by a certain sentence (e.g. ‘That is red’) that we accept it only when its truth condition is satisfied. But to say that this requirement is met by those who understand the word, is not to imply that the requirement is conceptualised by such people. Therefore (contrary to Davidson’s suggestion), it is not an implication of my use-conception of meaning that the concepts of truth condition, and hence of truth, are needed for the possession of more mundane concepts.

DAVIDSON ON DEFLATIONISM 21

4 It can be shown, I believe, that we need not go along with the Davidsonian orthodoxy regarding the compositionality of meaning: namely, that it can be explained only in terms of a Tarski-style account of truth together with a reduction of sentence-meanings to truth conditions. I would argue that understanding a sentence consists in nothing more than understanding its parts and appreciating how it has been constructed from them2. But if this is so, then compositionality is ensured, no matter what view is taken of how the meanings of words are constitùted. In particular, if the meaning of a word derives from its conforming to a certain basic regularity of use, then the meaning of a sentence will consist in its being constructed in such and such a way from primitives whose uses conform to such and such basic regularities. These four considerations cast doubt on Davidson’s truth-conditional analysis of sentence-meaning, and hence undermine his reason for challenging the deflationist thesis that meaning is conceptually prior to truth. Let me now turn to his second objection—the alleged unintelligibility of the equivalence schema and of its instances (e.g. the proposition that dogs bark is true if and only if dogs bark). One might suspect that Davidson’s attitude derives from scepticism about propositions; however he is quite explicit that this is not the objection. But in that case—if there really are such things—how can the expressions specifically designed to refer to them be unintelligible? And if, for example, ‘The proposition that dogs bark’ is not unintelligible, how can there be a problem about using it to say that its referent, in certain specified conditions, is true? Davidson’s answer is that it is obscure how the referents of such alleged singular terms could be determined by the referents of their constituent words. Now, insofar as this is so, it will be obscure how to develop a Davidsonian truthconditional meaning theory for such expressions. But are we really entitled to conclude that they are unintelligible? Why not instead suppose—as Davidson himself has done in certain other cases—that they are amongst the various recalcitrant constructions that, though perfectly meaningful, have so far resisted assimilation to the truth-conditional paradigm. After all, especially if we are waiving any objection to propositions as such, the expressions in question would appear to be on a par with ‘The hypothesis that dogs bark’, ‘The supposition that dogs bark’, etc.—which have no technical flavour and which are even more obviously meaningful singular terms. Moreover, if there really is an overwhelming reason to conclude that the referents of such expressions could never be shown to derive from the referents of their parts, then—instead of questioning the intelligibility of what would seem to be perfectly comprehensible ways of speaking—should we not be induced to look with an even more sceptical eye at the Davidsonian truth-conditional approach to compositionality?

22 PAUL HORWICH

An alternative to that approach is the use-theoretic point of view according to which an account of what we mean by that-clauses is given by an explanation of their deployment—roughly as follows. We accept a principle of the form u expresses the proposition that p iff Int(u)=my ‘p’ where the content of the right-hand side is that our procedures of interpretation, when applied to the utterance u (given the context in which it occurs) yields the sentence ‘p’ of the interpreter’s current language. Following this rule puts us in a position to deploy constructions such as ( u)(S holds-true u, and u expresses the proposition that p) ( u)(S utters u, and u expresses the proposition that p) ( u)(S wants-true u, and u expresses the proposition that p) which capture what it is for S respectively to believe, assert and desire that p. Thus the overall use of that-clauses may be explained. Needless to say, the credibility of the deflationary perspective must rest on more than a rebuttal of objections: a positive case for the view must be provided. As I have indicated, I think that the skeleton of such an argument involves, first, the observation that all non-trivial uses of the truth-predicate are displays of its generalising role; second, the recognition that this role requires nothing more or less than our acceptance of the equivalence schema; and third, the acknowledgement that such a predicate cannot be subjected to the sort of questions regarding ‘the underlying nature of the property for which it stands’ that are reasonable in the case of empirical, causal/explanatory notions. Notes 1 For an explanation and defence of the use theory of meaning see Horwich 1998. 2 See Horwich 1997.

Part II Meaning, truth and interpretation

3 Theories of meaning, truth and interpretation Kirk Ludwig

Introduction We owe to Donald Davidson the suggestion that a truth theory used as an interpretation theory for a speaker can do duty as a meaning theory for his language. This is a brilliant suggestion, but there are some oddities in the development of this idea in Davidson’s work which need to be brought to light, and the project, in the form it takes in Davidson’s hands, in the end is too ambitious to succeed. I begin by distinguishing three questions: 1 What is it for a word or expression to be meaningful? 2 What is it for a word or expression to mean what it does? 3 How do the meanings of complex expressions in a language depend upon those of their parts? A full answer to 1 would include answers to 2 and 3; but it is clear that one could pursue question 3 independently of 1 and 2, and 2 independently of 1. My discussion of the development of Davidson’s programme in semantics is organised around the distinction between these three questions. The second part of this chapter discusses Davidson’s development of the project of employing a truth theory in a central role in a recursive meaning theory for a natural language. Davidson’s project, as first stated in ‘Truth and Meaning’ (Davidson 1967c/1984a), undergoes an important but easily overlooked transformation from having the relatively modest goal of answering question 3 to that of answering simultaneously questions 1–3. Plausibly, a truth theory that meets certain constraints can be pressed into service to answer question 3. But the unremarked shift to the more ambitious goal imposes the requirement that the constraints should not appeal to any semantic facts. This has the effect of obscuring the possibility of a relatively straightforward answer to the question what constraints such a theory needs to serve a central role in a compositional meaning theory, if we put aside questions 1 and 2. When Davidson turns from his first proposal

KIRK LUDWIG 25

about what constraints are needed to his proposal in ‘Radical Interpretation’ (1973c) that the theory be confirmed by a radical interpreter, the effect of not having laid out the straightforward constraints is to leave unclear what it is that the radical interpreter has to confirm to succeed. Separating more cleanly the projects of answering question 3, and answering 1–2, helps us to evaluate the prospects of each, and, in particular, the prospects for the success of radical interpretation. The third part of this chapter considers the prospects for finding an answer to questions 1 and 2 by way of a description of the procedures of such a radical interpreter, who takes a recursive truth theory to give the basic structure of a meaning theory for another’s language. I distinguish a modest and ambitious programme in Davidson’s work, and argue that Davidson is best understood as engaged in the ambitious one. The ambitious programme assumes an a priori guarantee of success at radical interpretation. The modest programme eschews this. If the ambitious programme fails, then a number of important results Davidson grounds in reflection on the radical interpreter’s procedures are unattainable. I argue that the prospects for the ambitious programme are dim, and that this becomes clear once we see what is required for success in radical interpretation. In the fourth part of this chapter, I consider two a priori arguments for the ambitious programme suggested by some of Davidson’s recent work, and argue that neither succeeds. Truth and Meaning The suggestion that a truth theory can serve as a meaning theory is introduced in ‘Truth and Meaning’. Ostensibly, the paper begins with the following project: a satisfactory theory of meaning must give an account of how the meanings of sentences depend upon the meanings of words. Unless such an account could be supplied for a particular language, it is argued, there would be no explaining the fact that we can learn the language: no explaining the fact that, on mastering a finite vocabulary and a finitely stated set of rules, we are prepared to produce and to understand any of a potential infinitude of sentences…. I want to ask what it is for a theory to give an account of the kind adumbrated. (Davidson 1984a, p. 17) The project is to show how meanings of complexes depend on their significant parts and combination, and thus aims to answer question 3. Importantly, the aim is not to explain why primitive elements of a language are meaningful (question 1), or what it is for them to have the meanings they do (question 2). Davidson, in the course of his discussion of traditional attempts to carry out the project, says that ‘the task [is] to give the meaning of all expressions in a certain infinite set on the basis of the meaning of the parts; it [is] not in the bargain also to give the meanings of the atomic parts’ (ibid., p. 18). The aim is only to explain how

26 THEORIES OF MEANING, TRUTH, INTERPRETATION

mastery of a finite vocabulary and a finite set of rules suffices for the mastery of a potential infinitude of meaningful sentences. What are the constraints on such a theory? First, it must have as theorems for the target language L all true sentences of the form: (M) Φ 28in L means that p where ‘Φ’ is replaced by a structural description of a sentence of L and ‘p’ by a sentence giving its meaning (ambiguity can be removed by formulating the theory for a regimented version of the language). Second, the theory must be finitely specifiable so that a finite speaker can grasp it, and so that we can understand how a finite speaker could speak the language it is about. The argument leading to the adoption of a truth theory as the form of a compositional meaning theory aims to make us despair of getting the job done by appeal solely to meanings as entities assigned to expressions, or to an apparently intensional locution in our axioms such as ‘e in L means w’ where ‘e’ is replaced by a structural description of a primitive expression in the language and ‘w’ is replaced by an expression that gives its meaning. In brief,1 the difficulty for the second strategy is that it is unclear how to carry out the recursion necessary for generating meaning specifications for an infinity of sentences from a finite base without using the usual quantificational apparatus. For the first strategy, the most telling objection is that appeal to meanings as entities in a meaning theory could not in itself do the work required to generate true M-theorems. We need a sentence in use on the right-hand side of (M) that interprets the one mentioned on the left. Treating ‘means’ as taking a referring term τ in its second place, designating a sentence meaning, doesn’t guarantee τ is synonymous with the sentence whose meaning it denotes. In this case, grasp of the theorem would be insufficient for understanding the target sentence. Thus, appeal to meanings as entities is insufficient to specify appropriately the meaning of a sentence. If merely pairing an appropriately used sentence with a mentioned one is both necessary and sufficient, it begins to look as if quantification over meanings is at best a misleading technical device dispensable in principle. How do we proceed without quantifying over meanings? We want a theory that ‘provides, for every sentence s in the language under study, a matching sentence that, in some way yet to be made clear, ‘gives the meaning’ of s’ (ibid., p. 23). In a famous passage, Davidson continues: One obvious candidate for matching sentence is just s itself, if the object language is contained in the metalanguage; otherwise a translation of s in the metalanguage. As a final bold step, let us try treating the position occupied by ‘p’ [in ‘s means that p’] extensionally: to implement this, sweep away the obscure ‘means that’, provide the sentence that replaces ‘p’ with a proper sentential connective, and supply the description that replaces ‘s’ with its own predicate. The plausible result is

KIRK LUDWIG 27

(T) s is T if and only if p What we require of a theory of meaning for a language L is that without appeal to any (further) semantical notions it place enough restrictions [my italics] on the predicate ‘is T’ to entail all sentences got from schema T when ‘s’ is replaced by a structural description of a sentence of L and ‘p’ by that sentence [or, he might have added, by a translation of it]. (Davidson 1984a, p. 23) This last restriction is Tarski’s Convention T; thus, our predicate ‘is T’ will have in its extensional all and only true sentences in the language. As Davidson remarks, The path to this point has been tortuous, but the conclusion may be stated simply: a theory of meaning for a language L shows ‘how the meanings of sentences depend upon the meanings of words’ if it contains a (recursive) definition of truth-in-L. (Davidson 1984a, p. 23) Davidson came to realise that this overstates the case. Modifications to a Tarskistyle theory are required to adapt it to natural language, in particular to accommodate context-sensitive sentences. Employing disquotation to give the truth conditions of ‘I am hungry’, yields the absurd, ‘“I am hungry” is true iff I am hungry’, which blandly ignores the fact that the truth of utterances of ‘I am hungry’ depends on speaker and utterance time. Simplest is to relativise the truth predicate to speaker and time. The form of our theorems will be (T*). (T*) For all speakers s, times t, Φ is true in L as (potentially) spoken by s at t iff p Likewise, (M) must be modified so that ‘means that’ is relativised to a speaker and time, as in (M*). (M*) For all speakers s, times t, Φ means in L as (potentially) spoken by s at t that p (Henceforth I abbreviate these semantic predicates as ‘is true[s, t, L]’ and ‘means[s, that’.) We must also modify the convention we want the appropriate theorems to satisfy. Instead of ‘p’ being replaced by a sentence that translates Φ, we will want ‘p’ to be replaced by a sentence or sentence form which in the theorem interprets Φ. This condition will be met just in case replacing ‘is true in L as (potentially) spoken by s at t iff’ in a T*-sentence with ‘means in L as (potentially) spoken by s at t that’ yields a true sentence. Call this ‘Davidson’s Convention T’ to distinguish it from Tarski’s. (Call theorems which satisfy this convention ‘T-sentences’.)

t, L]

28 THEORIES OF MEANING, TRUTH, INTERPRETATION

More importantly, since the predicate ‘true-in-L’ is simply an unarticulated predicate which is to have in its extensional all and only true sentences of L, simply having a true recursive definition of it is not sufficient for the theory’s target theorems to meet Davidson’s Convention T—though Davidson apparently initially hoped that it would be. One can give an extensionally correct definition of such a predicate without the expressions one uses to do so, giving the meaning or interpretation of the sentences for which they fix the truth conditions. Given one materially correct definition, we can generate others by adding to the satisfaction conditions for various predicates any true sentence. From this we can prove non-interpretive materially adequate theorems. For example (suppressing relativisation to speaker and time for now), if For all functions f, f satisfies ‘x is red’ iff f(‘x’) is red is true, then so is, For all functions f, f satisfies ‘x is red’ iff f(‘x’) is red and Alexander crossed the Indus in 326 BC, where ‘f ’ ranges over functions assigning values to variables and ‘f(“x”)’ is the value assigned byn f to ‘x’. From this we can prove non-interpretive materially adequate theorems.2 What can we know about an extensionally correct truth definition for a language which enables it to serve as a meaning theory? We must know (1) that it has as theorems all T-sentences for the language, and (2) how to pick them out from the totality (any such theory with a rich enough logic will have as theorems sentences of the form ‘Φ is true-in-L, iff p’ in which what replaces ‘p’ does not provide an interpretation of Φ; e.g., ‘“snow is white” is true-in-L iff snow is white and T’, where ‘T’ is replaced by a logical truth). It would suffice to know that the theory had all the T-sentences as theorems to know that the axioms of the theory provide interpretations of the primitive expressions of the language, and that a standard procedure for proving T-sentences (a canonical proof procedure) yields theorems that, relative to the assumption that the base axioms are interpretive, satisfy Davidson’s Convention T. For example, for English, we would give for the predicate ‘is a rose’ (suppressing relativisation) the base axiom ‘For all functions f, f satisfies “x is a rose” iff f(“x”) is a rose’. Productive expressions receive the usual recursive axioms using the same device in the metalanguage as in the object language; for ‘and’, we give the recursive axiom, ‘For all functions f, sentences Φ, ψ, f satisfies Φ and ψ iff f satisfies Φ and f satisfies ψ’. A truth theory satisfying this desideratum is an interpretive truth theory. If we know, and know we have, an interpretive truth theory for L, and we know a canonical proof procedure, and that it is one, then we can straightforwardly provide (M*) theorems for each sentence of L, since each T-sentence yields a true (M*) sentence when ‘is true in L as potentially spoken by s at t’ is replaced

KIRK LUDWIG 29

by ‘means in L as potentially spoken by s at t that’. This enables us to answer question 3 for declarative sentences. We know that the axioms of the theory satisfy the desideratum that the expression for which each gives satisfaction conditions is interpreted by the expression used to give the satisfaction conditions. Meaning determines truth relative to context of utterance, and, thus, we can see, by way of the canonical proof of a T-sentence, how the meanings of the parts of the sentence in combination determine its meaning. We have not also answered questions 1 and 2. But we did not set out to. The theory is not irrelevant to these questions. If our languages have a compositional semantics, then an account of that must be a part of the answers to questions 1 and 2. The main problems for this approach are the technical problems associated with providing a truth theory for a natural language, such as extending the theory to constructions which resist incorporation into an extensional truth theory without introduction of intensional entities,3 and extending the theory to handle nondeclaratives.4 Davidson does not give this account. Instead, he offers conditions of adequacy on a truth theory, which presuppose nothing about what object language expressions mean. Davidson first suggests that any extensionally correct truth theory can be used as a meaning theory. The reason lies in two connected thoughts. The first is that a ‘theory of meaning…is an empirical theory’ (ibid., p. 24).5 The second is that the relativisation of truth to speaker and utterance time requires the theory to project correctly the truth conditions of sentences containing demonstratives such as ‘This is red’. Davidson apparently hoped that this would provide enough additional resolving power to ensure that any extensionally adequate theory satisfied Davidson’s Convention T. However, while this helps with sentences such as ‘For all speakers s, times t, “Grass is green” is true[s, t, L] iff snow is white’, this won’t ensure a theory that generates T-sentences.6 This represents a shift in the focus of Davidson’s project completed by ‘Radical Interpretation’, which, in his ‘Reply to Foster’, he presents as a criticism of his ‘own earlier attempts to say exactly what the relation is between a theory of truth and a theory of meaning’, and an attempt ‘to do better’ (ibid., p. 171). It is a shift from focusing on question 3, to questions 1 and 2. If Davidson wanted only to answer question 3, he could have stopped at the point above. To connect the question of when a truth theory can be used as a meaning theory with how to confirm it empirically from an interpreter’s standpoint is to connect it with the question of what it is for the primitive expressions in a language to have the meanings that they do, for the evidence that Davidson restricts himself to when he develops this project excludes any of the facts the theory is about. The shift occurs without remark in ‘Truth and Meaning’. It is unclear whether Davidson makes it deliberately. ‘Truth and Meaning’ creates the impression that Davidson begins with question 3, but switches to questions 1 and 2 at the point at which a clear answer to that question nearly emerges. This obscures how a truth theory can be employed as part of an answer to question 3 taken by itself. I will argue

30 THEORIES OF MEANING, TRUTH, INTERPRETATION

that this has further, unfortunate, consequences in Davidson’s appeal to radical interpretation in answering questions 1 and 2. To summarise: Davidson initially represents the project of ‘Truth and Meaning’ as showing how to provide a compositional natural language semantics. This can be accomplished with a truth theory if (a) we can solve the problem of giving recursive truth theories for natural language declaratives, (b) we can extend the approach to non-declaratives, and (c) they are interpretive. Meeting these conditions, for every canonically provable theorem we can infer a corresponding M*-sentence (or some corresponding form for non-declaratives). Davidson’s approach differs in what he puts in place of (c): the requirement that the truth theory be extensionally adequate. This represents a shift in focus from question 3 to questions 1 and 2. This point about the development of Davidson’s project has important implications for how Davidson ought to have, but did not, structure his more ambitious project. In the remainder of this paper, I consider that project as providing an answer simultaneously to questions 1–3 by appeal to radical interpretation. Many of Davidson’s more ambitious philosophical theses are associated with this project. I provide a reconstruction of the project of radical interpretation in the light of the discussion above about using a truth theory to provide a compositional meaning theory, and evaluate its prospects. I argue for two points. First, in light of the above, there is an especially difficult problem for the interpreter in arriving at a correct starting set of data. This problem has not been solved, and I suspect cannot be (see the section on ‘Radical Interpretation’, below). Second, in any case, for the project to have the results that Davidson conceives it having, he must make a strong assumption about the publicness of meaning facts, which is unsupported by intuitively compelling considerations about requirements for having a language. In connection with this, I criticize two arguments Davidson might plausibly give in support of the requirement he needs (see ‘Motivations for the ambitious programme’, below). Radical Interpretation The description of radical interpretation is to provide us with an answer to questions 1–3. It is to do so, not by providing a reductive account of meaning, but by illuminating the connection between the concept of meaning and others by showing how semantically neutral evidence can be marshalled to support an interpretation theory for a speaker. As Davidson presents it at the beginning of ‘Radical Interpretation’, the project is structured by two questions: (q1) What could we know that would enable us to interpret all of a speaker’s actual and potential utterances? (q2) How could we come to know this on the basis of evidence that does not presuppose that we already know it?

KIRK LUDWIG 31

I suggest that we already have an answer to (q1): what we could know that would enable us to interpret all of a speaker A’s actual and potential utterances is (K). (K)

(a)

(b)

that axiom A1 means that…, axiom A2 means that…,…; that…is a canonical proof procedure for T; that T is interpretive.

(c) (d)

that axioms…constitute a truth theory T for A’s language;

Knowing that T is an interpretive truth theory is knowing that its axioms are such that the satisfaction conditions give interpretations of the expressions for which they give satisfaction conditions, and a canonical proof procedure is one that, relative to the axioms providing interpretive satisfaction conditions, generates Tsentences.7 This Davidson ought to have said. Davidson’s actual answer substitutes for (d) above (d*): that T can be confirmed by a radical interpreter.8 Call the modified answer ‘(K*)’. This is a mistake that obscures what is required to confirm an interpretation theory. If we give (K*) in answer to (q1), then we cannot determine when an interpreter has succeeded in confirming a theory that can be used as a meaning theory. The theory he confirms is interpretive, not just if it is true, but if it is true and he confirms it! But his confirming it in itself adds nothing substantive to the theory. We need an independent characterisation of the kind of theory needed, given that an extensionally adequate theory is insufficient. In addition, giving (K*) as the answer to (q1) would require (q2) to be answered by saying how we could come to know inter alia that a radical interpreter had confirmed a truth theory for a speaker. But this is not Davidson’s answer to (q2). Davidson’s answer to (q2) is a description of the procedure of the radical interpreter. This is appropriate if we have a characterisation of what the interpreter aims to confirm that is independent of his success. Henceforth, I will assume that it is (K) which the interpreter seeks to know. Why should a description of the radical interpreter’s procedure be the right answer to (q2), even so? The question requires only that the evidence does not include facts about how to interpret the speaker’s utterances. If this is the only constraint, we can help ourselves to facts about attitudes, perhaps excluding those about word meaning. This is Grice’s project of intention-based semantics (Grice 1989). Why can’t Davidson help himself to this? The answer is that he aims for deeper illumination by choosing a more austere set of data. Motivation for the particular choice of data rests on the thought that language is essentially a medium for communication, and, hence, necessarily social and public. We must, therefore, be

32 THEORIES OF MEANING, TRUTH, INTERPRETATION

able to interpret one another on the basis of public cues. Thus, the appropriate standpoint from which to investigate meaning is that of an interpreter. To answer the question, how we could come to know (K), we need to say what our evidence is and how it supports the theory. Davidson helps himself to the attitude of holding (i.e. believing) a sentence true; in recent work he has based the interpretive project on the attitude of preferring true.9 Since the later approach incorporates the earlier, and my concerns focus on the stage which goes through holding true, I treat the basic evidence as consisting of facts about a speaker’s hold-true attitudes. Holding true is an attitude, so holding true a sentence is not the same as asserting it. It is to be disposed to assert the sentence sincerely in appropriate conditions. If one understands the sentence one holds true, then this suffices for one to be said to believe that p, where ‘p’ is replaced by the sentence one would assert or a translation of it (with appropriate adjustment for context sensitivity). Is Davidson entitled to his starting point? If he aimed only to show that certain data sufficed to confirm an interpretive truth theory, there could be no objection. He could choose what data he wanted; it might be expected to illuminate meaning by showing systematic interconnections between the concepts deployed in describing the data and the theoretical concepts. However, the uses to which he puts the results of radical interpretation, particularly in response to radical scepticism, require him to suppose that hold-true attitudes are identifiable from purely behavioural evidence (Davidson 1983a/1986). (The anti-sceptical argument requires that any speaker in any environment is interpretable by any interpreter. Having no magical access to hold-true attitudes, the interpreter identifies them from behaviour. Otherwise, the assumption the anti-sceptical argument requires is false.) This would require identification of syntactic types and grammatical structure in the language, as well as attitudes on behavioural grounds.10 Davidson has assumed rather than argued that we can do this. However, let us suppose that we can identify hold-true attitudes on the basis of behavioural evidence. An interpreter can also use information about the speaker’s interactions with his environment (non-intentionally described) without begging any questions. Given this, we could learn what sentences a speaker would hold true under various circumstances. We can now reformulate our second question more precisely. It is not just the question of how we could know (K) without presupposing what it tells us, but of how a radical interpreter could know (K). Our second question becomes: (q2’) How could we come to know (K) from the standpoint of the radical interpreter? We can now raise an important question about the project. What is that status of (q2’)’s presupposition that a radical interpreter can learn (K)? On one reading, ‘the modest programme’, we offer no a priori reason for its truth, but set out to see whether one can come to know (K) in radical interpretation, and what is necessary in order to come to know (K) in this way, if it is possible. On another

KIRK LUDWIG 33

reading, ‘the ambitious programme’, we offer an a priori reason to suppose that (q2’) can be answered. Davidson’s project makes sense only on the ambitious reading. The modest reading requires an independent criterion for success in interpretation. That radical interpretation can succeed is a contingent truth. This undermines the attempt to get out of reflections on radical interpretation any a priori arguments for the indeterminacy of interpretation and the inscrutability of reference, the impossibility of radically different conceptual schemes, the impossibility of our empirical beliefs being massively mistaken, and the necessity of first-person authority for speakers. The arguments for these theses presuppose that assumptions required in radical interpretation reflect necessary facts about its subject matter. To get these results, we need an a priori guarantee that speakers are radically interpretable, that is, an a priori guarantee that (q2’) has an answer. I return below to the question what defence Davidson could mount for this assumption. For now, suppose (q2’) has an answer. What must be true for radical interpretation to succeed? Part of the account of how we can interpret another speaker must rely on an a priori theory of speakers. Speakers, performing speech acts, are rational agents; thus, a large part of our theory will be an a priori theory of rational agents. Davidson’s work in the philosophy of action provides elements of the a priori theory of speakers. This work also grounds Davidson’s so-called holism about attitude content, not his reflections on radical interpretation. (Davidson’s holism about attitude content is not so radical as sometimes represented. He does not hold that believing p requires one to have any other specific beliefs, but, rather, that one must have (indefinitely) many beliefs with related contents, though no particular ones. Davidson’s holism has none of the wild consequences, such as that no two people can disagree, or that no beliefs are retained through belief change, which it is routine to attribute to holists, among whom Davidson is usually counted.)11 In interpretation, then, we help ourselves to three things: (a) the speaker’s hold-true attitudes; (b) the speaker’s interactions with his environment; (c) an a priori theory of speakers and rational agency. From this, how could we assign meanings to the speaker’s utterances and contents to his beliefs? The procedure is as follows. (a) Identify those hold-true attitudes for the speaker s which are caused by conditions in his environment; (b) Use this evidence to justify universally quantified bi-conditionals of the form, (L) (t)(ceteris paribus: s holds true Φ at t iff p)

34 THEORIES OF MEANING, TRUTH, INTERPRETATION

where we replace ‘p’ by a sentence or sentence form (to allow for a variable ranging over times) which gives in (L) conditions under which s holds true Φ (an Lsentence). (Time and speaker together should suffice to fix other contextual features relevant to truth conditions.) An L-sentence is a ceteris paribus law. We can think of this as a (possibly irreducibly probabilistic) law which holds in some set of unspecified conditions. (It cannot come out that L-sentences express strict laws, on pain of contradicting Davidson’s thesis of anomalous monism: Davidson 1970b). Now we add the principle of charity in its application to L-sentences. The principle of charity has two different elements. One consists of requiring the interpreter to assume that the speaker is mostly right about his/her perceptual environment. The other consists in requiring the interpreter to find the speaker largely a rational agent. The former is motivated by appeal to what an interpreter must assume to get past evidence in the form of L-sentences. The latter is drawn from the a priori theory of speakers. I am concerned with the former only. The principle of charity aims to enable us to derive from L-sentences sentences of the form (Tcp), (Tcp)

(t)(ceteris paribus: Φ is true[s, t] iff p)

which can then be treated as evidence for a correct truth theory which can be used to specify meanings for the speaker’s sentences (Tcp-sentences). To derive Tcp- sentences from L-sentences, we need both the conditions that, ceteris paribus, when a speaker is caused to hold true a sentence, by and large it is; and that, ceteris paribus, when a sentence about the speaker’s environment is true, by and large the speaker is caused to hold it true. Let ‘C’ represent the ceteris paribus conditions. Call sentences of s’s language for which some L-sentence is a true lawlike sentence occasion sentences.12 The principle of charity must validate at least: (PC1) If C, a speaker holds true an occasion sentence Φ at t iff Φ is true[s, t] Many interpreters have taken (PC1) to be the principle of charity. But (PC1) is too weak. (PC1) guarantees only that the evidence yields an extensionally adequate theory, which is insufficient for interpretiveness. We need a guarantee that Tcp-sentences, by and large, meet Davidson’s Convention T, i.e. something strong enough to guarantee (PC2). (PC2) In true L-sentences, by and large, the sentences (sentence forms) used on the right-hand side provide an interpretation of the sentences denoted on the left hand side If the principle of charity were (PC1), it could not secure (PC2). However, the principle of charity is properly not about sentences held true, but about beliefs.

KIRK LUDWIG 35

This could be clearer in Davidson’s writings,13 but the following passage from ‘Radical Interpretation’ supports this interpretation: This method is intended to solve the problem of the interdependence of belief and meaning by holding belief constant as far as possible while solving for meaning. This is accomplished by assigning truth conditions to alien sentences that make native speakers right when plausibly possible, according, of course, to our own view of what is right. (Davidson 1984a, p. 137) The connection between the assumption that a speaker is mostly right about his environment and (PC2) is made by assuming that if a speaker believes that p, then he holds true a sentence Φ which expresses that belief because he believes that Φ means that p; that is, (leaving aside relativisation) since he believes that p, and knows that Φ means that p, and that if Φ means that p, then Φ is true iff p, he comes to believe that Ф is true. We must also assume that by and large if a speaker holds true a sentence Φ, he holds it true because he knows what Φ means and believes what Φ expresses. Thus, we can reason as follows. Take a speaker s, and sentence Φ. Suppose, by charity, that for some substitution for ‘p’ about s’s environment (1) is true. Next invoke our assumption that for every belief a speaker has he holds true a sentence which expresses it for the reasons above (and, roughly, vice versa) to get (2). Our explanation of the truth of (2) gives us (3). Since if Φ means that p, then Φ is true iff p, we infer (4). (1) s believes that p iff p; (2) s believes that p iff s holds true Φ; (3) Ф means that p; (4) Φ is true iff p. It is natural to assume that in identifying the conditions that prompt hold-true attitudes, we identify the conditions under which the speaker believes what the sentence expresses, i.e. that the procedures of the radical interpreter yield (5). From this we can infer (6) and (7), which is what is needed to validate (PC2). (5) s holds true Φ iff p; (6) Ф is true iff p; (7) Φ means that p. To validate (PC2) we need to assume not just that speakers are most right about their environments, but also that the L-sentences which the interpreter confirms identify the interpretive truth conditions of the true beliefs which explain the speaker’s hold-true attitudes. Thus, after arriving at L-sentences, the next step is (c) Employ (PC2) to arrive at “Tcp-sentences.

36 THEORIES OF MEANING, TRUTH, INTERPRETATION

The radical interpreter amasses evidence in the form of Tcp-sentences, and on that basis formulates hypotheses about the axioms of a truth theory whose canonical theorems include the confirmed Tcp sentences, or as many as possible, all else being equal. He tests the theory against further evidence and modifies it until the theory works smoothly in providing meaning and content assignments that exhibit the speaker as intelligible and rational. How is the principle of charity justified? In the modest programme, it cannot be justified. To do so, we would have to be able to identify independently attitudes and meanings. But the constraints on our project forbid us from doing so. Thus, in the modest programme, we get at best a conditional result. If the principle of charity is true, then such and such. The ambitious programme, however, can offer the following justification. (i) Necessarily, question (q2’) admits of an answer. (ii) It is possible to interpret another speaker radically only if the principle of charity is true. (iii) Therefore, by (i) and (ii), the principle of charity is true. Interpretability requires also that the ceteris paribus conditions for the most part obtain, and so this immediately gives us the conclusion that most of the speaker’s hold-true attitudes about the world are true. Now for criticism. My charge is that the radical interpreter cannot know something he needs to know to employ the principle of charity. Let ‘LΦ’ represent an L-sentence about a sentence Φ of s’s language. The interpreter must know not just that most of a speaker’s world-directed beliefs are true, but also that the L-sentences he identifies are those which identify the interpretive truth conditions of the speaker’s beliefs. Thus, he must know something sufficient to narrow down the appropriate substitutions for ‘p’ in an LΦ-sentence to just those which express the belief on the basis of which s holds true Φ. However, if one LΦ-sentence is true, many will be. Part of the problem of deciding whether (PC2) is justified is deciding whether we can specify appropriate a priori constraints on selecting from among true LΦ-sentences only one, or a small number, which do not, in conjunction with (PC2), lead to genuinely different interpretations of Φ. There is a trade-off between how much indeterminacy one allows and what additional constraints must be added to arrive at interpretations of occasion sentences. If one says that every sentence or sentence form, which, when substituted for ‘p’ in LΦ, makes the resulting sentence true, expresses the same interpretation, then one is guaranteed uniqueness of interpretation, but also a great deal of indeterminacy, since a very large number of sentences or sentence forms could make, for any speaker, and any occasion sentences Φ in that speaker’s language, an LΦ-sentence true. (Indeterminacy can arise elsewhere also.) In this case no additional constraints are required, but what we gain in simplicity is lost in plausibility. Indeed, if there are more laws than one expressed in true LΦ-sentences, we cannot take this route, for if every sentence or sentence form which yields a true sentence when substituted for ‘p’ in LΦ expresses the same interpretation, then every resulting sentence has the same meaning, and, hence, expresses the same law. Even putting aside empirical considerations, it is clear that there are many different laws which will be expressed by LΦ sentences.

KIRK LUDWIG 37

Given one true LΦ-sentence, we can generate others by replacing the sentence substituted for ‘p’ with any nomically or logically equivalent one when the ceteris paribus conditions hold. It is not plausible that differences in sentence meanings cut no finer than nomic equivalence. In any case, the interpreter does not have a choice. He investigates the world and discovers what laws hold. There will be, for any sentence Φ of the object language, many different true LΦ-sentences. There are at least as many different possible interpretations as there are LΦ-laws. The interpreter has incompatible starting points for his/her interpretation theory. At most one can completely and correctly give the interpretations of the speaker’s sentences. He needs to choose those L-sentences which identify the interpretive truth conditions for the speaker’s beliefs that explain his hold-true attitudes. If he has no grounds for choosing between them, and at most one can be correct, he cannot know that he has started with the right set, and, hence, can never be in a position to know (K). In this case, the project cannot be carried out. There are two responses to this problem. First, one can appeal to contingent knowledge about the speaker’s psychology, for example to facts about what he is likely to notice in his/her environment, about his/her sensory experiences, about his/her desires, etc. But this presupposes knowledge of facts which we have set out to discover without already presupposing any such knowledge. To make this appeal is to give up. Second, one can try to give a priori reasons why some true L-sentences are not acceptable. One might, for example, argue that only one choice of L-sentences would exhibit the speaker as rational. The case for this has not been made, and it is difficult to see how rationality constraints, which are neutral with respect to content, and involve rather patterns among contents, could dictate a unique set of assignments. In the absence of any argument to show that there must be a unique set of assignments compatible with rationality constraints, this suggestion is implausible. I consider in the next section what a priori arguments I think can be discerned in Davidson’s work to try to establish that there must a way of narrowing down the choices. If one despairs of showing that only one set of L-sentences could provide an interpretation of another as a rational agent, then the only way I can see to respond to the difficulty is to hold that none of the interpretations is correct, but that each empirically acceptable one captures equally well the facts of the matter. One might claim that each possible starting point results in a theory equally good at explaining the speaker’s behaviour. The interpreter’s position with respect to the speaker could be likened to ours with respect to non-linguistic animals, whose attitudes we track with our sentences despite the behavioural evidence being insufficient to choose between different incompatible contents. But we could not invoke this as a general defence of our failure to arrive at unique interpretations. Consider two speakers of the same language, A and B. A would have to conclude that B’s language was not exactly translatable into his/ her language. But by hypothesis that is false. Thus, we are committed to there

38 THEORIES OF MEANING, TRUTH, INTERPRETATION

being a way to select objectively an LΦ-sentence that correctly interprets Φ, at least in the case of other speakers of our language. To summarise: the reconstruction I have offered of Davidson’s project answers (q1) by appeal to (K). The procedure for verifying an interpretive truth theory requires us to marshal evidence for axioms of the theory by appeal to Tcpsentences which are interpretive. The position is non-reductionist because of its ineliminatable reference to radical interpretation. We saw that it was necessary for the interpreter to assume that he can identify the conditions under which a speaker holds true occasion sentences which are also the interpretive truth conditions of the belief on the basis of which s/he holds them true. To succeed, the interpreter must choose correctly among different non-equivalent sets of nomic conditions for the truth of occasion sentences. Furthermore, for reflection on radical interpretation to have the results Davidson aims for, the interpreter must solve this problem without appeal to contingent facts about the speaker’s psychology. The difficulty cannot be avoided by appeal to indeterminacy, because the interpreter must treat the different conditions as non-equivalent, if he treats them as involving different laws. Motivations for the ambitious programme It is not easy to tease out of Davidson’s work an argument for the presupposition of (q2’). I consider two arguments, both suggested by things he says, but neither of which I am very confident he would accept. My first suggestion is that it is derived from the thought that being able to communicate with other speakers is essential to having a language.14 Here is how the line of argument goes15 : (1) Necessarily, speakers can interpret other speakers and be interpretable by other speakers [a priori truth]. (2) We do not have direct epistemic access to other speakers’ thoughts or utterance meanings [a priori truth?]. (3) By (2), the only evidence available for interpretation is third-person evidence —evidence equally available to every potential interpreter of a speaker. (4) From (1) and (3), any speaker can be interpreted on the basis of third-person evidence in any environment by any other speaker. (5) From (4), question (q2’) admits of an answer. (4) secures, in conjunction with the procedure of the radical interpreter outlined above, the impossibility of massive error about one’s environment, the impossibility of radically different conceptual schemes, and that whatever is not discoverable on the basis of the sort of evidence available to the radical interpreter is no part of meaning or the content of propositional attitudes. It ensures, given that we are linguistic beings, that we are now radically interpretable, and that everything is true of us which that requires.

KIRK LUDWIG 39

It is sometimes objected that if the conclusion was correct, then we would not know that we were speakers, and so not know most of our beliefs were true. This shows that the argument must also assume that we know the concepts employed in it apply to us, and that sceptics have assumed this all along. I am willing to grant this. My objection is that the argument, as I’ve stated it, is not sound; and I cannot convince myself that there is any adequate repair. The first premise is ambiguous. It doesn’t say whether we have to be able to interpret every other speaker, or to be interpretable by every other speaker, or only some in each case. It is most plausible if we interpret it as claiming that we must be able to interpret some other speakers, and that some other speakers must be able to interpret us. Interpreted this way, the problem is that (1) and (3) do not entail (4). In addition, (3) does not follow from (2). To start with the first of these objections, being interpretable by other speakers need require no more than that one is interpretable by some other speaker in some circumstances. This doesn’t entail that one is interpretable by every other speaker, even in principle (and so doesn’t entail that all languages are intertranslatable), or that in every environment one is in one is interpretable (and so doesn’t entail that one is mostly right about one’s present environment or about any environment one has ever been in, if that is a condition on being interpretable in it), and it doesn’t entail that third-person evidence is ever sufficient for interpretation. One could strengthen (1) so that it did entail that every speaker is interpretable by every other in any environment which he is in. But this is not supported by the unassailable observation that to speak language is to have a capacity to communicate with others. The second objection is that the step from (2) to (3) is invalid. The reason is that while we do not have direct epistemic access to others’ thoughts or meanings, we do to our own, or at least we do know what our own mental states are, by and large, without observing our behaviour. This can be relevant to how we interpret other speakers because we could have good reason to think them similar to us in observable respects which make it likely that they are similar in certain psychological respects. For example, we have good reason to think that general features of a being’s psychology are fixed by its biological (or physical) kind. Knowing what those features are in our case, we can project them onto our conspecifics. However, this would clearly undermine the ambitious programme’s conclusions. It might be said that the weak interpretation of the first premise refutes scepticism if being able to interpret others entails it is possible justifiably to interpret someone. This would require us to be able to know something about someone’s behaviour, and so to be justified in believing something about the world around us, contrary to radical scepticism. Everything hinges on whether the first premise so interpreted is true. There is clearly something right about it. But it need come to no more than that if we can justifiably come to believe things about the world, we have the capacity, relative to certain general assumptions about likeness of psychology, to arrive justifiably at interpretations of appropriate others.

40 THEORIES OF MEANING, TRUTH, INTERPRETATION

The second line of argument is in a number of recent papers of Davidson’s, but I am not confident that I have interpreted it correctly. The papers are: Davidson 1990f/1991, 1991a and 1992b. Some of the themes in some of the versions of the arguments have appeared in earlier papers, particularly Davidson 1975b and 1982a. I present a composite argument which I think represents the main line of argument in these papers.16 (1) To have a belief, one must have the concept of a belief. (2) To have the concept of a belief, one must have the concept of error, or, equivalently, of objective truth. (3) The claim that a being has the concept of objective truth stands in need of grounding; in particular, there must be scope for the application of the concept in the being’s experience. (4) We can understand how a being in communication with others could have the concept of error, as a tool used in interpretation to achieve a better rational fit of another’s behaviour to the evidence we have for his beliefs and meanings: the concept would have work to do for interpreters. (5) Only in interpretation is there scope for the application of the concept of error. (6) Therefore, from (3)–(5), to have the concept of objective truth, one must be (or perhaps have been) in communication with others. (7) Therefore, from (1), (2) and (6), to have beliefs, one must be a speaker, and be actually (or perhaps have been) in communication with others. If successful, this argument secures not only our interpretability, but our actually being (or having been) interpreted. An amazing result! But there is an obvious weak link: premise (5). (This is not the only weak link.) We can grant that if we were in communication with others, then the concept of objective error would be a natural (perhaps necessary) concept for us to have. But why should it be necessary to be in communication with others to have the concept? First, if anything is necessary, why should it not merely be that we have experiences as of interpreting others? This is a familiar difficulty for transcendental arguments that infer from conditions on having concepts to their correct application: it seems enough, once we have the correct story, that all of the psychological elements are in place, without contact with what the representational elements of the story are about. Second, why is even this much necessary? The concept of error has a point for us if there is a point to thinking ourselves mistaken. There could be a point to this outside the context of communication; for example, whenever one has incompatible beliefs and this makes a difference to one (as it would were one, for instance, concerned to believe the truth). There is no reason to suppose that having the concept of objective truth requires that we be (or have been) or even think we are (or have been) communicating with others. This second argument also fails to establish that we are actually interpretable or interpreted if we are speakers.

KIRK LUDWIG 41

Conclusion We get the clearest view of Davidson’s project in the theory of meaning by separating it into two parts. The first part aims to answer question 3 in our original list by appeal to an interpretive truth theory. The second part aims to answer questions 1–3 by appeal to the procedures for confirming an interpretive truth theory from the standpoint of the radical interpreter. Prospects for the success of the latter, more ambitious programme, are not bright. That we need as our evidential starting point not just true Tcp-sentences but interpretive Tcp-sentences highlightst he problem of selecting from many different true LΦ-sentences the one that interprets Φ. Unless the problem of selecting only suitable starting points can be solved without empirical assumptions about speakers’ psychologies, the project must fail. I doubt that this problem can be solved. Furthermore, Davidson’s ambitious programme requires that he support a priori the presupposition of (q2’), that radical interpretation can succeed. The first of two arguments for this claim is invalid unless supplemented by assumptions which are not supported by uncontentious claims about the publicity of language; the second assumes without argument that having the concept of objective truth requires communicating or having communicated with others. To the extent that we can see a ground for the claim that communication with others suffices for having the concept of objective truth, there are conceivable conditions which do not require this. The suggestion that an interpretive truth theory can serve as a compositional meaning theory remains an important and permanent contribution to semantics. Study of how empirically to confirm such a theory for another speaker remains an important project, but we must recognise— itself an important point about meaning—that success in interpretation requires empirical assumptions about the speakers we interpret, namely, that they are much like us. Notes 1 See Lepore and Ludwig Typescript for more discussion. 2 The deficiency is not removed by reading ‘iff’ as having the force of ‘nomically equivalent’ or ‘necessarily equivalent’; two sentences may be nomically or necessarily equivalent but non-synonymous. 3 See Ludwig and Ray 1998 for a truth-theoretic semantics for opaque contexts, which avoids quantifying over intensional entities. 4 See Ludwig 1997 for an extension of the truth-theoretic approach to nondeclaratives, which keeps a truth theory at the core of a meaning theory without reducing nondeclaratives to declaratives. 5 The hint here that in the empirical conditions for confirming the theory we find reasons to believe that it does the intended job becomes a major theme in Davidson’s later work. In Davidson 1984a, p. 27, we already find the first sketch of the project of radical interpretation, with many of its central themes. See also ‘In Defense of Convention T’:

42 THEORIES OF MEANING, TRUTH, INTERPRETATION

If we treat T-sentences as verifiable, then a theory of truth shows how we can go from truth to something like meaning—enough like meaning so that if someone had a theory for a language verified in the way I propose, he would be able to use that language in communication. (Davidson 1984a, p. 74) 6 For present purposes I assume ‘snow is white’ is meant timelessly. See Lepore and Ludwig 1998 for a truth-theoretic semantics for tensed expressions. 7 This avoids objections raised by Foster 1976 and Loar 1976 to the effect that the interpreter can’t put his knowledge together so as to be able to interpret another. 8 This seems to be the only way to interpret Davidson’s emphasis that we know we have the right truth theory for interpretation when we know that it meets certain ‘formal and empirical restrictions on the theory as a whole’ (Davidson 1984a, p. 134), where the empirical restrictions are that it is confirmed from the standpoint of the radical interpreter. The retrospective footnote 11 of ‘Truth and Meaning’ (1984a) is also significant; Davidson there emphasises the lawlike character of Tsentences in defending the use of a truth theory as an interpretation theory, a theme emphasised also in ‘Reply to Foster’; but this would make sense only if he were thinking of a condition on its use as a meaning theory being its empirical confirmation. 9 See Davidson 1985f and the appendix to Davidson 1990e, where the argument is repeated. 10 While Davidson describes a procedure for identifying the logical constants in 1985f, he presupposes that he has a partial syntactic analysis of the speaker’s language. 11 See Fodor and Lepore 1994 for an example. 12 I borrow Quine’s terminology (Quine 1960, Chapter 2) for obvious reasons, though the project described here is different from Quine’s. 13 See Lepore and Ludwig typescript for further discussion. 14 See Davidson 1979c/1984a, p. 235: ‘The semantic features of language are public features. What no one can, in the nature of the case, figure out from the totality of the relevant evidence cannot be part of meaning.’ 15 I criticise an argument of this general form in Ludwig 1992. 16 In Davidson 1991c it looks as if Davidson is appealing to the argument that is attributed to the Wittgenstein of the Philosophical Investigations (1958), according to which one cannot follow a rule by oneself, since it is not possible to make sense of one’s going on in the right way if the only check on whether one has is whether one thinks so. Since we can follow rules, the argument goes, the standard by which we can correct ourselves must be external, and the only candidate is how others in our speech community go on. This is not a persuasive argument: as many commentators have noted, it is mysterious why the community’s judgements are not subject to the same objection.

KIRK LUDWIG 43

REPLY TO KIRK LUDWIG Donald Davidson Along with numerous objections which are mentioned but not argued, Ludwig seems mainly concerned to show that radical interpretation, as I have described it, cannot hope to accomplish what I claim for it. Since he seems to have seriously misunderstood my position in a number of ways, let me first try to set the record straight. The fairly radical change of mind he thinks he discovers in ‘Truth and Meaning’ is not there. The point of the article, which I think has been clear to most readers, is that a Tarski truth definition, applied to a speaker, and with the last step which converts an axiomatisation to a definition omitted, so that the truth predicate is left undefined, can provide an interpreter with what s/he needs in order to understand that speaker. It is hard for me to comprehend how I could be thought to have suggested that such a theory would show only how the meaning of sentences depended on the meaning of their parts, but would give no idea how the parts were to be interpreted. Semantics, at least as I (and Tarski) conceive it, gives the semantics of sentences, and so must, of course, specify the semantic roles of the parts. The passage Ludwig quotes from p. 18, that it is not ‘part of the bargain…to give the meanings of all the parts’, means only that the analysis of individual words, a job often confused with interpretation, is not part of the project I was outlining. This point is made clearly later in the article (p. 31). I have adopted from the start Quine’s idea that we catch on to the interpretation of basic predicates in ostensive situations; if the speaker has an unknown tongue, we notice the situations in which he is prompted to accede to or dissent from a sentence of the form ‘This is red’, ‘That is a dog’, etc. This process, which almost every linguist and philosopher has agreed must be basic, is the key to learning, whether as an interpreter or as the acquirer of a first language, the meanings of the relevant predicates. I call this a form of ‘charity’ in the sense that it assumes meanings are more or less the same when relevant verbal behaviours are the same. The other aspect of charity is the assumption of a degree of rationality (consistency) on the part of a speaker. This has always seemed to me an a priori assumption because what are to be interpreted are sentences, and the propositional contents of sentences, which have logical relations to one another, partly determine what the sentences mean. I find no argument against this in Ludwig’s criticisms, and so I am at a loss to understand his repeated claim that I am making unsupported assumptions in the application of the principle of charity. Finally, Ludwig seems to think I assume, without argument, that radical interpretation will work only for creatures who are much alike. I do not think of this as an assumption, but as a partial condition on what is correctly called language or thought. Thought and language are features and functions of rationality; to deny this would be to quarrel uselessly over words. If this is the

44 THEORIES OF MEANING, TRUTH, INTERPRETATION

‘assumption’ of which I am guilty, I plead guilty. But interpretation requires more similarity than this: we could only understand another creature which was tuned to some of the main features of the world we are tuned to. Far from denying, or playing this down, I have repeatedly emphasised it.

4 How a truth theory can do duty as a theory of meaning Gabriel Segal

Donald Davidson has made the remarkable, insightful and initially incredible claim that a Tarskian truth definition—a mere material truth theory—can ‘do duty’ for a theory of meaning. I will explain what I think Davidson means by the claim and how I think he defends it. I will offer a sort of objection to Davidson’s defence. Then I will go on to offer my own defence of the claim, which beds it within a philosophical perspective that differs somewhat from Davidson’s. Let me first define some terms, since I’ll be using some vocabulary in a slightly untraditional way. A T-theory for a language, L, is a finite, compositional theory that contains the resources to prove, for each sentence of L, a T-theorem of the form (T), where S would be replaced by a metalinguistic description of the L-sentence, and p is a sentence of the metalanguage that has the same truth value as the sentence named by S: (T) S is true iff p I call a T-theorem ‘interpretive’ if the ‘p’ on the right translates the sentence specified on the left. And I call a T-theory ‘interpretive’ if it entails only interpretive T-theorems. So an interpretive T-theory will entail at least one interpretive T-theorem for each sentence of L and no uninterpretive ones. A theory of meaning for L is one that specifies all the information required to interpret every sentence of L on the basis of its syntax. So, if you knew a theory of meaning for, say, Japanese, and you were able to identify sentences of Japanese on the basis of their syntax, then you should, in principle at least, be able to use the theory to deduce the meaning of each Japanese sentence.1 Davidson’s idea is that an interpretive T-theory for a language can in some sense, as he tends to put it, ‘do duty’ for a theory of meaning for that language. Whatever that means, there’s an obvious problem. For T-theories, even interpretive ones, don’t specify what the words and sentences of their object language mean. An interpretive T-theory does not itself say or entail that it is interpretive. One might put it this way: if there is one interpretive T-theory for a language, L, then there will also be many uninterpretive ones. If all you knew about the semantics of L was a particular T-theory for it, then you would have no

46 GABRIEL SEGAL

way of telling whether the T-theory you knew was interpretive or not. So, even if the T-theory was in fact interpretive, you wouldn’t know enough to know what the sentences of L mean. What would be involved in this further knowledge? Apparently rather a lot. Compare the two T-sentences in 1: 1(a) ‘Les fleurs sont perissables’ is true iff flowers wilt. 1(b) ‘Les fleurs sont perissables’ is true iff pigs have curly tails. Sentence (a) is interpretive and sentence (b) isn’t. What would you need to know in order to know this? You would need to know that ‘Les fleurs sont perissables’ means that flowers wilt, and that it does not mean that pigs have curly tails. But that is just the kind of information a theory of meaning was supposed to provide in the first place. So now it looks as though a T-theory could only do duty for a theory of meaning if it were combined with a theory of meaning. That is no doubt true, but not helpful. Davidson’s solution to this problem has the following abstract form: he aims to articulate a set, C, of constraints that will ensure that a T-theory is interpretive. If these constraints are correctly formulated, then any T-theory meeting C will be interpretive. This will allow him to say that what suffices for interpretation is knowing some T-theory that meets C, and knowing that it does meet C. If the theory meets C, it will in fact be interpretive. And if one knows that it meets C, one should therefore know that it is interpretive (assuming that it is a priori, that C is indeed sufficient to its task). Davidson imposes a general meta-constraint on the constraints: C should be formulated with as little recourse to linguistic semantic notions as possible. There are obviously many correct but unilluminating ways in which one might formulate C. The meta-constraint serves to rule these out. The underlying need for the meta-constraint flows from Davidson’s more general philosophical concerns. He wishes to provide some illumination of the relationships between the concepts of linguistic semantics (what a sentence means, what a sentence is understood to say, and so on) and other concepts that are ‘better understood or clear or more basic epistemologically or ontologically’ (Davidson 1984a, p. 137). Concepts of the latter variety might include such things as speakers’ dispositions to produce and react to utterances of sentences. Davidson’s aim is not to provide definitions or reductions or analyses of linguistic semantic notions for the more basic ones. It is the more realistic aim of detailing the interrelationships. His idea seems to be that if we can formulate constraints on interpretivity that satisfy the meta-constraint, then we will have, as it were, a conceptual map locating the place of the linguistic-semantic in the rest of the world. It is not hard to find also in Davidson’s work a different kind of metaconstraint, something that functions as an adequacy condition on C.

A TRUTH THEORY AS A THEORY OF MEANING? 47

C’s meeting this adequacy condition will ensure that he’s got it right, that any Ttheory meeting C really will be interpretive. The adequacy condition is the principle of charity. Very roughly, the idea is that a method of interpretation for L should interpret utterances of L-speakers so that as many of them come out true as possible. ‘True’ of course must mean true by the interpreter’s lights. So the idea is to maximise agreement between interpreter and interpretee. More realistically, one should interpret in order to minimise inexplicable disagreement and ensure ‘the right sort of agreement’ (Davidson 1984a, p. xvii). As he says ‘The “right sort”, however, is no easier to specify than to say what constitutes a good reason for holding a particular belief’ (ibid.). To sum up, Davidson aims to discover a set C, of constraints, formulated so far as possible without recourse to linguistic semantic notions, such that any Ttheory meeting C will be interpretive. So, if you were to use a T-theory meeting C as a theory of meaning, using the ‘p’s on the right-hand side of T-theorems as interpretations of the utterances of the sentences specified on the left, the interpretations thus yielded would be charitable. What, then, are C? If we read pp. xiv, 26 and 174 of Davidson (1984a) we find a certain emphasis on the point that T-theories are empirical theories about speakers, and so should not only be true but also lawlike. We also find mention of counterfactuals. A distinction between the interpretive 1(a) and the uninterpretive 1(b) is that the former is lawlike and the latter is not. Laws support counterfactuals. Contraposing counterfactual versions of the relevant sentences we get: 2(a) ‘Les fleurs sont perissables’ would not be true if flowers were not to wilt. 2(b)‘Les fleurs sont perissables’ would not be true if pigs were not to have curly tails. Sentence 2(a) appears to be true, and 2(b) appears to be false: even if the world were to differ from the way it actually is in respect of the truth of ‘Les fleurs sont perissables’, pigs would still have curly tails. Lawlikeness alone leaves some room for some uninterpretive results. If 1(a) is a law, so is 3: (3) ‘Les fleurs sont perissables’ is true if (flowers wilt and 2+2=4). But these cases can be ruled out by another standard feature of empirical theories: they should be as simple as possible (in some sense of ‘simple’ that philosophers of science try to explain). In any event, it is often not too hard to tell in practice when a T-theory contains unnecessary complexity either by imposing unnecessary structure on object-language sentences or by adding superfluous wrinkles to the axioms. Some residual difficulties might seem to remain. It is a law that electrical and thermal conductivity co-vary in metals. Consequently, 4 appears to be a law too:

48 GABRIEL SEGAL

(4) ‘Copper conducts heat’ is true iff copper conducts electricity. And 4 by itself doesn’t seem to be unnecessarily complicated. However, there are at least two routes Davidson might take to avoid this sort of counterexample. First, he might appeal to the requirement of compositionality: the T-theory must yield lawlike T-theorems for every sentence of the language. Sentence 4 would presumably not be a theorem of an appropriately constrained T-theory that also entailed theorems for many other sentences containing ‘heat’ and ‘electricity’. Second, he might point out that although 4 is a law, it is ‘gruelike’, arrived at by combining the laws of two quite separate sciences. He could then modify the requirement of lawlikeness and appeal to a specific kind of lawlikeness. I myself wouldn’t like to bet either way on whether these tactics will succeed across the board. But I don’t see why Davidson shouldn’t be forgiven optimism over the score. With this much in hand, Davidson can offer a philosophical account of linguistic meaning. The core of the account could be expressed, roughly, along the lines of (P): (P) S means that p in L iff a canonical theorem of a maximally economical, lawlike T-theory for L entails that S is true iff p. (P) must be taken informally, since (1) the functioning of the variable ‘p’ is at best unclear, and (2) it is awkward to quantify into an ‘entails that’ context. But it serves to give the flavour of the idea. Further, (P) would by no means exhaust the account. The project of limning the relationships between linguistic/semantic concepts and others requires that we know the role of non-linguistic/semantic evidence in confirming that we have a T-theory of the requisite character. Here, Davidson proposes a certain kind of decision theory that can be used to confirm when a speaker holds a sentence true, or, as he has more recently suggested (Davidson 1990e) prefers one sentence true to another. Holding true, or preferring true, are unanalysed linguistic/semantic notions. But they are purely extensional. If a person holds a sentence true, or prefers one sentence true to another, such a thing holds however the sentences (or the person) is described. Moreover, the content of these notions is easily understandable given their role in the theory. They are detectable from (though not reducible to) observable behaviour. A speaker, Kurt, utters S when and only when it is raining; this provides evidence for the claim that Kurt holds true S when and only when it is raining. And, in their turn, these notions provide evidence for the T-sentences that the T-theorist aims to prove (Davidson 1984a, p. 135). Davidson also addresses under determination: there will be more than one lawlike, maximally simple T-theory that is confirmable on the basis of the relevant evidence. In essence, his view appears to be that all such theories are

A TRUTH THEORY AS A THEORY OF MEANING? 49

true, each offering its own version of the reference and satisfaction relations that are drawn upon to generate the theorems.2 Davidson’s overall picture is surely ingenious and appealing. It is also plausible for someone with Davidson’s more general philosophical outlook. However, for someone with a different philosophical outlook, it won’t quite do. Davidson’s picture relies on what one might think of as a hint of veriflcationism and a hint of behaviourism. These are only hints. But to someone with a more naively realistic outlook, this serves to weaken the edifice. As I am of the latter persuasion, I have a sort of objection, or two. The problem is that passing charitable tests would seem to be neither necessary nor sufficient for interpretiveness. Let us look at each of these points in turn. I will express them crudely first, relying on crass examples. I will then abstract away from the example-based style of objections, and gesture towards the more general principle underlying them. Imagine someone who knows a language but who manifests no observable sign of this knowledge. He neither speaks, nor appears to understand. Indeed, there are cases of temporary aphasia that seem to be just like that. During the aphasic period the subject loses all, or large parts of, their capacity to speak and understand. However, their condition improves and their abilities return. It is natural to suppose that during the aphasic period, the subject still had, or knew their language. (If not, it is hard to explain the return of their abilities: they don’t need to relearn the language from scratch.) Take someone whose aphasia has the same underlying cause as that of the temporary cases, but suppose they never recover. In such a case, a T-theory would pass no charitable tests, since the subject doesn’t provide evidence by which the theorist could test it. Nevertheless, it might be interpretive for the language. Now I don’t think that Davidson is strictly committed to the necessity of charity in this sense. (Indeed, in conversation he has confirmed this.) Nevertheless, it is not clear that his outlook permits a satisfactory way of dealing with the case. What is it that distinguishes the aphasic from a worm or a prelinguistic neonate, or from someone who had a language but had all their linguistic circuitry unfortunately destroyed by a bullet? I would suggest that for Davidson the answer would lie in certain counterfactuals: the subject would manifest understanding (would assent to S when and only when it is raining) if they were functioning better, or were in (endogenously) better condition. But counterfactual moves of this sort usually run into trouble, since there is rarely any workable non-question-begging way of specifying the relevant conditions.3 And if one is prepared to accept that there is no such specification, then one must be prepared to suppose that connections between the concept of having or knowing a language (that which determines whether a given T-theory is interpretive for a subject or not) and the concept of charity are not as clear as Davidson might wish. I leave the aphasic and move on now to the second kind of crass example, and the insufficiency of charity. Passing charitable tests is not sufficient for

50 GABRIEL SEGAL

interpretiveness, since a giant but stupid computer might behave just like a competent speaker, but understand nothing. (The argument that follows is drawn from Block, 1981.) The kind of computer Block envisages is a giant look-up table. All it contains within it is a huge list of canned responses to stimuli. And all it can do is identify the stimulus, locate the canned response and produce it. To get a handle on how the machine might work, let’s begin with one that is designed to pass the Turing Test. All its inputs and outputs are typed sentences of English. What it does is produce apparently sensible responses to the sentences that it receives as inputs. It does this, for, let’s say, an hour. Here’s how it works. It has all possible hour-long, English, sensible conversations (or conversations that are sensible at least on the part of one interlocutor) represented in data structures within it. On being given the first input, it locates a conversation beginning with that sentence, and produces the response that is associated with it. On receiving the next input, it locates the conversation which begins with the pattern of sentences thus far encountered, and produces the next response. And so on. The machine just described, of course, doesn’t much resemble a human in its panoply of apparently rational behaviour. But we can imagine enlarging the machine so that it does so resemble the human, without changing the basic structure of the program. First, we enlarge its conversational abilities to cover a lifetime instead of an hour. Then we deal with the problem of having it behave appropriately in non-linguistic ways. To do this, we note that we have finite sensory receptors and finite possibilities of bodily movement, and that there are only finitely many possible finite sequences of stimuli we can receive in a lifetime and finitely many possible finite sequences of responses we can produce in a lifetime. We now construct a giant look-up table that works as before: for each sequence of stimuli an intelligible response is pre-programmed. Now, it is very likely that building such a machine is far beyond any capacity that humans will ever attain. The machine is perhaps even nomologically impossible, given the combinatorial explosions that it would involve. But it does seem that the machine is conceptually and mathematically possible. And if there were such a machine, I would say that it doesn’t mean or understand anything by its words. Yet a well-chosen T-theory for the language of its inputs and outputs would pass charitable tests. And so, by Davidson’s lights, the T-theory ought to be interpretive for its language. Note that this is not like John Searle’s Chinese Room argument (Searle 1980). The point about Block’s machine is not that it is a computer, but that it is a very stupid computer. I abstain on the question of whether a machine with the right programming would, merely by virtue of that programming, be a genuinely intentional system. I said that the example-based arguments were crass and I certainly wouldn’t want to place too much weight on them. The second one, in particular, is of limited intellectual value. Intuitions generated by thought experiments that require a stretch of the imagination easily inspire doubt. But the examples do

A TRUTH THEORY AS A THEORY OF MEANING? 51

serve illustrative purposes: one can sometimes understand a view better by looking at its consequences in extreme cases. The underlying point behind the examples is really the simple one that all that is gold does not glitter and that not all that glitters is gold. Interesting surface phenomena are usually evidence of something more interesting underlying them. Human linguistic capacity is remarkable in its scope, power and subtlety. So one would expect something interesting to underlie the surface phenomena of language. An empirical theory of the phenomena, like empirical theories generally, will inform us about underlying causes. And for that reason, the theory will not be conclusively (in the conceptual or logical sense) confirmed by any particular range of observable evidence. I think this is true of empirical theories generally, and the case of a theory of language should be no different. If all that is right, then it should in principle be possible to conceive of a different kind of causal structure that would give rise to the same surface phenomena, one that would not be the real thing not the same natural kind of thing—but a mere facsimile. The giant look-up table was meant to be a case in point. Likewise, one should be able to envisage the underlying causal structure cut off from its usual manifestations. The aphasic was meant to illustrate that possibility. But it’s not the cases in point that are the point, so much as the principle they might help illustrate: that outward criteria need an inward process, that language is a natural kind of phenomenon. I don’t expect someone with Davidson’s philosophical outlook to be moved by these considerations. But for those who might be moved, let me recommend a different tack. Davidson’s central insight is that an interpretive T-theory is as good as a theory of meaning for someone who knows that it is interpretive. This goes both for the speaker and for the semantic theorist. There are questions of a certain type that fascinate semantic theorists. We want to know how the language works, what the semantic structures of sentences are, how the meanings of the parts contribute to the meanings of the wholes, and so on. We like to ask, as Davidson once put it ‘What are these familiar words doing here?’ (Davidson 1984a, p. 94). We need a theory within which to raise and answer these questions, and Davidson saw that T-theories are just what we want: they have exactly the properties we want our semantic theory to have. T-theories are relativity limited in their scope: what they actually say only concerns reference, satisfaction and truth. Hence the problem we encountered at the beginning: they don’t say what words or sentences mean. But with limited scope comes formal tractability. The formal apparatus required to construct T-theories is limited, and, on the whole, well understood. Moreover, we know enough about how to go about constructing T-theories for natural languages to be reasonably optimistic that they might actually ‘turn the trick’. So if we could solve the problem of scope, the problem that T-theories don’t say what words and sentences mean, that would be great. Having registered unhappiness with Davidson’s solution to the problem, I must put something in its place. What I propose is that we work towards a theory of the causal structure underlying human linguistic capacity. Let’s hypothesise, first, that our linguistic

52 GABRIEL SEGAL

competence is owed to our actual possession of unconscious knowledge of a body of linguistic rules; the ‘language faculty’, roughly as conceived by Noam Chomsky (e.g. Chomsky 1980; 1986). The language faculty contains phonological, syntactic and semantic rules, and it is (in part) because we know these rules that we speak and understand as we do. Let us also make the further hypothesis that the semantic rules take the form of an interpretive T-theory. (Here I depart from Chomsky, who does not think that semantic theories for natural language should be formulated in the standard philosophical vocabulary of reference, satisfaction and truth.) My proposal initially runs straight into the problem that we encountered earlier. Even if we did unconsciously know an interpretative T-theory for our language, this would not account for semantic competence, since the theory doesn’t specify what the linguistic items mean. To address this problem, I need to invoke Chomsky’s ‘competence-performance’ distinction. Chomsky emphasises that it is one thing to have linguistic competence—unconsciously to know the rules of a language—and another thing to be able to put this knowledge to use in actual performances, such as finding words to express oneself, understanding a perceived utterance, or answering a linguist’s question about how one understands a particular sentence. We must therefore hypothesise further that linguistic knowledge is drawn upon by one or more cognitive systems that deploy it in the performance of the various tasks. The temporary aphasic presumably retains their linguistic competence, but suffers from impaired performance systems. The performance systems thus have access to semantic knowledge, and this takes the form of an interpretive T-theory. What I propose is that they deploy the T-theory on the assumption that it is interpretive. That is to say, they are so designed that they use the ‘p’ on the right-hand side of a T-theorem to interpret the S specified on the left. Thus, when one learns a language, what one acquires is knowledge of an interpretive T-theory for that language. This is built into the design of the language-acquisition mechanisms. Given this, it makes sense if the cognitive systems that deploy the knowledge are designed so that they automatically treat the T-theory to which they have access as interpretive. I am now in a position to make some more or less substantial claims about linguistic meaning, (M) and (M*): (M)

S means that p for individual i iff a canonical theorem of i’s internalised T-theory entails that S is true iff p. (M*) What an expression e means for i is given by the canonical clause for e in i’s internalised T-theory.

(M) and (M*) are not intended as philosophical claims, but as empirical hypotheses, roughly along the lines of water=H20, or heat in a gas=mean molecular kinetic energy. I will not here defend (M), (M*) or the general approach on which they are based.4 I will only say that it seems to me that the

A TRUTH THEORY AS A THEORY OF MEANING? 53

same kind of cognitivist theory to language—an approach that explains linguistic competence in terms of unconscious knowledge—is almost mandatory. It is the only kind of theory that addresses the data in any kind of depth and detail. As for the specific claim that semantic knowledge is knowledge of an interpretive Ttheory, I would only repeat what I have already said: interpretive T-theories have just the right properties, and it is not at all clear that any other kind of theory does. I conclude with a remark about the philosophical significance of my empirically oriented approach. As it stands, it offers little by way of genuinely philosophical illumination. However, if such be sought, I recommend a generally Davidsonian attitude to it. The strategy of seeking interesting, conceptually necessary and sufficient conditions for things (truth, justice, beauty, knowledge, and so on) has not proved tremendously successful over the past 2,000 years. We should seek rather to trace out interesting interrelationships among our concepts. My empirical theses about linguistic meaning could be integral to such an enterprise, were it combined with accounts of (1) the relationship of semantic knowledge to other (linguistic and non-linguistic) bodies of unconscious knowledge; (2) the theoretical role of unconscious knowledge in cognitivist explanation; and (3) the evidence that would confirm/disconfirm theories of unconscious knowledge in general, and unconscious semantic knowledge in particular. I would be glad to provide these, but constraints on space forbid it.5 Notes 1 Here and below I will ignore the complexities of indexicals, demonstratives and contextually dependent features of the meaning of utterances. 2 For details, see Davidson 1977b/1984a. 3 For more on this, see Segal 1994. 4 The theory is elaborated and defined in detail in Larson and Segal 1995. 5 An early version of this paper was presented at the Lublin (Poland) conference on ‘Tarski and Davidson’s Semantic Program’. I am grateful to participants at the conference for helpful comments and to the British Academy of Sciences for a grant enabling me to attend it.

REPLY TO GABRIEL SEGAL Donald Davidson There seem to be two main difficulties Segal finds with my approach to the problem of meaning. The first is that there are formidable difficulties in showing that radical interpretation based on a Tarski-type theory of truth can succeed in its aim of providing enough information to an interpreter to enable him to understand a speaker. The second is that by ignoring the actual mechanism that enables people to perform this feat, I have left myself open to the claim that an unthinking automaton might satisfy my conditions.

54 GABRIEL SEGAL

There certainly are good reasons to wonder whether my approach could succeed, but I don’t think Segal has pointed to the ones that worry me. I did at one time suggest that an interpreter needed to know more than a T-theory; s/he needed also to know that it was an adequate theory. I now think s/he doesn’t need to know either of these things. The interpreter doesn’t have to have a Ttheory; s/he needs to know just the T-sentences entailed by such a theory. In fact, I think most people do know some parts of such a theory (e.g. that a conjunction is true if and only if each conjunct is true). But theory is essential for two reasons. First, only such a theory can say what it is that a competent interpreter knows (all the unlimited number of T-sentences)—something the philosopher would like to be able to say. Second, the theory also gives substance to the claim that an interpreter could come to know all that, using evidence of a kind plausibly available. The sort of theory I have in mind is not, of course, any old theory that yields a true biconditional in the form of a T-sentence for every sentence of the speaker’s language. Being an empirical theory, it favours simplicity, like any other empirical theory. It also preserves entailment relations between sentences, which immediately eliminates a vast number of ‘non-interpretive’ theories. It also preserves, within reason, relations of evidential support between sentences (as emphasised in ‘The Structure and Content of Truth’), which should exclude Segal’s 4: ‘“Copper conducts heat” if and only if copper conducts electricity.’ Someone who knows the T-sentences of such a theory for a speaker can understand that speaker. Such an interpreter does not need the additional knowledge that this is an interpretive theory. I find little interest, or merit, in raising questions about a theory of interpretation that applies equally to any theory whatever. I’m afraid I am, as Segal conjectures, little impressed with science fiction stories about machines that act in every way like a normal person. Such stories can’t be told with enough detail to allow us to evaluate the counterfactuals involved. For example, a machine set to carry out any possible two-hour conversational exchange isn’t anything like any person I know. Until someone builds the machine that really does act like us, we don’t know what to say. For example, is it physically possible that a machine could act in every way just like us, and yet be built in a very different way? I very much doubt it, despite the speculations of Segal and Ned Block. But if it were done, I would probably say that it’s a person, and should be given the vote. I have nothing against the hypotheses of Chomsky, Fodor and others concerning the sort of mechanisms that are built into people. There seems sound evidence for the idea that first languages are not learned in anything remotely like the methods of my radical interpreter. But then, I never thought they were.

5 Radical interpretation and compositional structure Peter Pagin

In this chapter I shall be concerned with the relation between a particular account of linguistic meaning and the property of compositionality in natural language.1 The account, proposed by Donald Davidson, is that based on considerations about radical interpretation. I shall argue that there is a fundamental conflict between, on the one hand, the view that the meaning of expressions of natural languages is determined purely according to canons of radical interpretation, and, on the other hand, the view that natural languages exhibit compositional structure. I shall also argue that if there is such a conflict, this speaks against the proposed account.2 Radical interpretation versus compositionality My main aim in this paper is to explore the consequences of basing the entire account of meaning on considerations about radical interpretation. That Davidson seems to have moved in this direction is one reason for considering this option.3 Another reason is that compositionality is an important property, perhaps the most important property, of human languages, and if we aim at a fundamental account of linguistic meaning, we cannot just take that property for granted. That natural language does have compositional structure is a tenet that needs justification, and if we aim at a fundamental account, that justification should flow from its assumptions. This is a reason for looking to radical interpretation for such a justification. For this purpose I shall consider what I shall call the RI account of meaning. By the RI account of meaning I shall understand an account of meaning that involves the following two tenets: 1 A sentence, as uttered by a speaker X at a time t, means what it can be successfully interpreted as meaning. 2 Success of an interpretation is judged by the degree that the interpretation conforms to norms of charitable interpretation. What the RI account of meaning is supposed to be depends, clearly, on what we include in ‘norms of charitable interpretations’. Clearly the ideas that Davidson

56 INTERPRETATION AND COMPOSITIONAL STRUCTURE

himself, and some others, have advanced as making up ‘the principle of charity’ must be included: that speakers are mostly (massively) right in their beliefs, that they are rational, coherent agents with reasonably stable beliefs, desires and preferences. But I would also want to include in these norms general features of thought and attitudes, which do not themselves concern linguistic structure. For instance, the idea of holistic individuation of thought content may well be counted among those norms, if it is a basic and general feature of thoughts, irrespective of particular properties of the languages in which they are expressed. I only want to exclude principles which directly concern structural properties of languages or linguistic expressions. I do not claim that Davidson has ever advanced the RI account of meaning. One reason to be cautious about it is precisely that Davidson in fact may have thought all along, as I do, not only that compositionality is an essential aspect of natural language meaning, but that this feature is not derivable from the RI account of meaning alone, and has to be added by separate considerations, such as our pre-theoretic knowledge of English as a language with significant semantic structure. That compositionality is not derivable from the RI account of meaning alone is something I shall argue for. More precisely, I shall argue for the following three tenets: 1 The view that natural language speakers speak a compositional language, or the view that natural language interpreters have a compositional competence, cannot be justified from the RI account of meaning alone. 2 If natural languages have the semantic features which the most successful interpretational theories ascribe to them, then they are not compositional. 3 If 1 and 2 are correct, then the RI account of meaning is false. The tenets can be alternatively rendered as follows:4 Suppose we shall come up with a theory of the language L of a speaker X at a time t. The theory shall state the meaning of the sentences of L, and hence provide a long list of meaning specifications in some form.5 The criterion of truth of such a theory is provided by the RI account of meaning. That is, the theorems are true if, and only if, utterances of X can be successfully interpreted as meaning precisely what the theorems specify. Then the question is whether there is any justification for the view that such a theory shall exhibit compositional structure, for example as Davidson’s T-theories do. Tenets 1 and 2 can then be reformulated as concerning such theories of particular languages or speakers at times. RI and the justification of compositionality I shall here consider a number of possible reasons for believing in the compositionality of natural language. I shall start with some classical arguments, and see how these fare under the RI account of meaning, and then I shall

PETER PAGIN 57

consider some further reasons derivable from this account itself. I shall try to show that in none of these cases do we get a justification of compositionality. Two things should be kept in mind. First, I am not going to argue that intuitively successful radical interpretation, i.e. radical interpretation which is successful by ordinary, intuitive standards, does not require compositionality (indeed, the intuitive standards may themselves involve pre-theoretic conceptions of semantic structure). That question is considered in passing by Martin Davies (1986, pp. 149–52). Second, I am not going to complain that the RI account cannot offer any proof of the compositionality principle, or that it cannot provide a knockdown argument. I don’t believe that the compositionality principle (as applied to natural languages) can be proved, or that there are any knockdown arguments for it. I think there are good arguments for the principle, but these are not of the knock-down kind. The argument I favour is an inference to the best explanation, or, somewhat stronger, an inference to the only reasonable explanation. But in both forms, the argument is disputable. What I shall complain is that the RI account does not offer any reasonable argument at all. Or, to be more precise, I shall try to show that all the candidates below, which are those I (and others, so far) can think of, are unsatisfactory, but not really that no better candidates can be had. Learnability One of the classical arguments for compositionality is based on the idea that natural languages must be learnable, since speakers do in fact learn them. This was Davidson’s own argument (1965/1984a, pp. 8–9). This argument presupposes that natural languages do have an infinite (or extremely large finite) number of meaningful sentences (with different meanings), and also a few rather uncontroversial things about human cognitive capacity. Let’s say that such a language is very rich. If we are allowed to assume that very rich languages exist as entities to be learned, then we do have a good argument for compositionality. That a very rich language has a compositional structure is (part of) a good explanation of how a speaker can learn it. It is probably the best explanation, and possibly the only reasonable explanation. The question is whether we are allowed to make the assumption. In the present context, that question comes down to whether the assumption can be justified on the basis of the RI account of meaning. Since this account proceeds by way of interpretation of individual speakers, that justification must involve at least three steps. In the first step we would have to find a justification for ascribing to a single speaker the property of being a speaker of a very rich language. In the second we would have to ascertain that two or more speakers speak the same very rich language. And in the third we would need a reason for believing that at least one of them has learned that language from the others.

58 INTERPRETATION AND COMPOSITIONAL STRUCTURE

Now clearly, if we can justify the assumption that speakers speak a very rich language without compositional structure, then we know in advance that the learnability argument is flawed, because, if speakers learn such a language from each other we know that it cannot be correctly explained by means of compositionality, simply because we know that the language is not compositional. On the other hand, if we can justify directly the assumption that speakers do speak very rich compositional languages, then we already have an argument for compositionality, and need not add any extra considerations about learnability. The only reason for turning to learnability as an argument would be that we can justify the assumption that speakers do speak very rich languages without being able, directly, to justify the belief that they are compositional. But I cannot see at all how we could end up in that situation, for I don’t think there could be any good reason for ascribing to a speaker the property of speaking a very rich language except on the basis of believing that a compositional theory gives the best account of the finitely many utterances he has actually made. If this is correct, then the learnability argument fails, in the sense that compositionality would already have to be justified if we had good reasons for the premises of the learnability argument. In fact, I think that this holds generally, not just within the context of the radical interpretation approach to meaning. In this context, the conclusion is even sharper. For the judgement that a speaker does speak a very rich language must be based on radical interpretation, issuing in a theory about the speaker’s language. That theory is either compositional or non-compositional, definitely rendering the learnability argument either superfluous or flawed. And if it is added that radical interpretation must respect compositional structure because of the learnability argument, then the whole reasoning is viciously circular, since the learnability argument, if it isn’t flawed, depends on the assumption of compositional outputs of radical interpretation. Understanding new sentences Closely related to the argument from learnability is the argument from understanding new sentences. It is said that a language must be compositional in order for it to be possible to understand new sentences—sentences you have never used or encountered before. This argument, however, can be understood in two different ways. One way to understand it is to think of a language as an independent entity, containing many sentences. Only a fragment of all its sentences have actually been used by its speakers. Frequently, however, speakers are confronted with sentences never used before, and they manage to understand them. Compositionality is then invoked for explaining this. Again, however, we must ask for a justification of the assumption that there are sentences which have never been used and which nonetheless have a welldetermined meaning. I think the situation here is exactly parallel to the situation

PETER PAGIN 59

we had with the learnability argument. The only reasonable justification of the view that sentences never used before do have a well-determined meaning (different from meanings of sentences that have been used) is compositionality. And so there will be no need to add extra considerations about the understanding of new sentences. The other way to understand the argument is more effective. We shift to utterances, or sentences as used by speakers. Then we can ask how it is possible for an audience to interpret a new sentence, which it may never have used or encountered before, as the speaker intends it to be interpreted. The natural answer is that it can be understood because of its compositional structure. The speaker employs the compositionality of her language in order to find an expression of her thought, and, by grasping the meaning of the parts and the mode of composition, the interpreter manages to understand the utterance in the right way.6 This does provide a good argument for compositionality on certain empirical assumptions. If we can take for granted that speakers often do understand each other, and often right away, even when they use sentences not used before, and express thoughts not expressed before, then we have good reason for believing that we need compositionality for explaining this. But from the viewpoint of the RI account of meaning this cannot simply be taken for granted. Rather, whether two speakers understand each other is something that would have to be ascertained by applying radical interpretation.7 Either we interpret them both, and check whether they understand each other, or they apply the methods of charitable interpretation to each other. Either way, we will again find ourselves in this situation: if the two speakers are successfully interpreted as speaking a compositional language, then we don’t need special considerations about communication for justifying compositionality, and if the speakers are successfully interpreted as speaking non-compositionally, then an argument from communication will be flawed, from the standpoint of the RI account. So the results about the compositionality, or non-compositionality, of the speakers’ languages must be established before the question whether they understand each other can be answered. We must therefore do without assumptions about mutual understanding when it comes to justifying compositionality within the RI framework. What we need to ask is whether it is a necessary feature of interpretation that linguistic communication, normally and typically, proceeds by employing compositional structure. And it seems that, from the standpoint of RI, it is not. If we are to get an argument for compositionality from the understanding of new utterances, we must get it on the assumption that the interpreter has the knowledge needed for understanding the utterance before the utterance is made. His linguistic competence, including knowledge of the compositional structure of the speaker’s language, should allow him to know in advance what that particular sentence, as used by that speaker, means, or would mean if uttered, because that is precisely what compositionality is good for. If we put it in terms

60 INTERPRETATION AND COMPOSITIONAL STRUCTURE

of having a theory, as Davidson often does, we can say that the interpreter should be able to predict what the new sentence will mean in the mouth of the speaker.8 But there is no general requirement in the RI account of meaning that the interpreter should be able to predict meaning. All that is required is that the interpreter should be able to arrive at the right interpretation, even if he can do so only after the utterance has been made, exploiting some extra knowledge he has gained in the circumstances of the utterance or later (even much later). Davidson himself stresses this when criticising the idea that linguistic meaning as such is determined by conventions. There is no general requirement of successful prediction (and that is why, according to Davidson, conventions aren’t necessary for linguistic communication).9 Correctness of an interpretation is judged solely by the norms of charitable interpretation. And charity can only be applied to attitudes already revealed, i.e. with respect to utterances already made. Unless compositionality is assumed, interpretations of unuttered sentences are irrelevant to the evaluation of the interpretation scheme as a whole.10 If you add compositionality, certain interpretations of sentences uttered commit you to certain interpretations of sentences not uttered, and so induce predictions. But if you only have charity to begin with, predictions are not induced and not required.11 Hence, within the framework of the RI account, we cannot appeal to compositionality to explain the understanding of new sentences. So within this framework, this version of the classical argument for compositionality fails. Instead, to justify compositionality within RI we have to look for specific features of the account itself. Regularity It is often stressed that for a speaker to be interpretable, he must use language in a regular way. If we could not find any correlation between his assent to sentences and features of his surroundings, then we would never be able to interpret even his observation sentences. And if he often changed his linguistic habits, suddenly starting to express different thoughts with the same old sentences, then again we could never get evidence enough for a correct interpretation. I am not sure to what extent this is true, but be that as it may. The point I am making does not depend on the assumption that the speaker does, or even may, deviate in any way from her previous language or speech habits. The speaker may be perfectly regular. My point is just that you don’t get compositionality from the demand that a speaker speaks regularly in the sense of never changing the meaning, or use, of any single sentence. That demand does not even suggest compositionality. To motivate compositionality, there would have to be a more complicated kind of regularity, in the use of sentence parts, since what is needed is a pattern for relating sentences which the speaker has not so far used to those sentences which she has used. You don’t get this kind of regularity by requiring

PETER PAGIN 61

only that the speaker be consistent in her use of each single sentence. This places no requirement at all on sentences not yet used. The question is whether anything belonging with the RI account can motivate the requirement of semantic regularity in the use of sentence parts. Again, my challenge to the RI account does not depend on any assumption on my part that the speaker can vary the use of sentence parts over time. We may well assume that the speaker doesn’t, and cannot, but the question is how to learn about it from the principles of the RI account. Holism That the interpreter need not be able to predict the meanings of new utterances does not, however, imply that the interpreter need not have a system of interpretation. There can be other considerations, based on the norms of charitable interpretation, that force the interpreter to be systematic, and perhaps compositional too. Perhaps the holistic nature of thought provides one such consideration. So let us assume that thought is holistic in roughly the sense of Davidson’s presentations of this idea: There are good reasons for not insisting on any particular list of beliefs that are needed if a creature is to wonder whether a gun is loaded. Nevertheless, it is necessary that there be endless interlocked beliefs. The system of such beliefs identifies a thought by locating it in a logical and epistemic space. (Davidson 1975b/1984a, p. 157) In order to entertain the thought that the gun is loaded, for example in the form of a desire that it be loaded, the speaker must have a number of true beliefs about guns and about being loaded, and these other beliefs will interlock with yet further beliefs. And the same goes for the interpreter. So the interpreter cannot ascertain that the speaker is expressing a belief that the gun is loaded unless he can also ascertain that the speaker has a great number of other beliefs. And this in turn requires the interpreter to provide interpretations of a number of other utterances of the speaker. The interpreter must therefore interpret not only a great number of utterances of the speaker, but the beliefs and other attitudes she thereby ascribes to the speaker must form a coherent pattern. The interpretations must support one another. This much does, I think, follow from the holistic nature of thought, together with principles of coherence and rationality as norms of interpretation. But the next step, to the conclusion that there must be some compositional structure in the expressions of the attitudes, is not taken. To be sure, if there are semantic relations between attitudes expressed by sentences, then there are also corresponding semantic relations between the sentences expressing them. But

62 INTERPRETATION AND COMPOSITIONAL STRUCTURE

there is still no requirement, and it is not even suggested by these considerations, that there be syntactic relations, between those sentences, that mirror the semantic relations (for example that, for at least some sentence parts, sentences containing that part are about the same thing). It is quite sufficient that the thoughts themselves form a system of the required kind, and that the utterances and sentences are interpreted as expressing them.12 Syntax isn’t touched by holism. Note that I do not claim that it is possible for a speaker to have a coherent system of beliefs without speaking compositionally. In fact, I believe that to be impossible. I only claim that as far as the standards of charitable interpretation go, there is no reason to doubt that possibility. A possible further point about holism is this. Since holism requires the existence of a whole system of beliefs as a condition of having thoughts, and since every thought must be expressible by a sentence, it might be that the sheer required number of meaningful sentences could only be realised by a compositional language. That would be very plausible, if infinitely many beliefs, and therefore infinitely many sentences, were required. But that does not seem to be required by holism, on any conception of holism that I know of.13 Even so, it can be claimed that it is inconsistent with holism that a person has a fixed finite number of beliefs.14 Rather, a person frequently forms new beliefs. Therefore, there cannot be a fixed finite number of sentences in a person’s language, for the new beliefs must be expressible also, and this might be held to imply that the language expressing the beliefs must be compositional. But this is not right either. As long as a person is free to fit the words to his thoughts as he pleases, he can freely make up the new sentences, one by one, which he needs to express his new thoughts. There is nothing for compositionality to explain there, and therefore no justification of it from holism. Of course, it can be objected that we are not free to mean whatever we want by our sentences, and I would say that this claim is correct, but the reasons for this go far beyond holism. So they give no support whatsoever of the idea that holism can justify compositionality. Underdetermination You might, however, want to protest here: it should not be enough to assign meanings to expressions to the effect of ascribing a coherent system of attitudes to the speaker. For clearly, unless you impose some further constraint on how to match sentences with thoughts, you can come up with many different attributions, since many different sets of attitudes form a coherent system. In order to eliminate this massive underdetermination, you must assume a pattern of correspondence between the thought and the expression of the thought. Indeed, underdetermination is one of the basic themes in Davidson’s account of meaning. He has reminded us many times that an assertive utterance depends on two factors: what the speaker believes and what his words mean. We can

PETER PAGIN 63

interpret the words one way, provided we are prepared to attribute to the speaker the corresponding belief, but for each utterance there are many pairs of belief and meaning that can explain why the speaker holds the sentence true. Moreover, Davidson has given as a reason for not attributing propositional attitudes to animals that the attributions will be heavily underdetermined by the evidence for them (Davidson 1974b/1984a, p. 164). And clearly, by requiring compositional structure, we can reduce the number of possible interpretations, since we then require that syntactic relations between sentences correspond to semantic relations between them, and many otherwise possible interpretations will fail to satisfy this requirement. But there are three problems to solve before you have the right to conclude that radical interpretation requires imposing compositional structure. First, you must explain why the plurality of possible interpretations is unacceptable. That is, you must explain why the method of charitable interpretation fails to distinguish between correct and incorrect interpretations when compositional structure is not required. The alternative view is that ‘Indeterminacy of meaning or translation does not represent a failure to capture significant distinctions; it marks the fact that certain apparent distinctions are not significant’ (Davidson 1973c/1984a, p. 154). That is, if you trust your method of interpretation, you will say that it does capture all that is objective about meaning, and that intuitions to the contrary are misleading. Davidson thought so about his preferred method of radical interpretation (which included compositionality). The general question is why some underdetermination is acceptable and some not. Until there is a basis for drawing the line that can be derived from the RI account of meaning itself, it has not been shown that this account will force the further restrictions on interpretation that we asked for.15 The second problem is this. Even if it can be established that the RI account of meaning must involve further restrictions in order to reduce underdetermination, there is so far no reason why that restriction should be compositional structure. Any odd restriction, like trying to match numbers of letters in words with number of words in sentences, can serve to reduce underdetermination. Too much underdetermination is itself just a problem. It does not direct you to the solution. The third problem is an extra twist to the second. There may be some reason, derived from the norms of charitable interpretation, for preferring some extra restrictions over others. You may want the restrictions you choose to preserve those interpretations which fulfil the norms to the highest degree, and which rule out others. You will then get compositionality as your preferred restriction, if the compositional interpretations do get the highest marks. There is, however, reason to doubt this. I return to this question in the next section.

64 INTERPRETATION AND COMPOSITIONAL STRUCTURE

The interpreter’s theory I want to conclude this part of the chapter by considering Davidson’s claim that we still do need a recursive representation of the interpreter’s competence at any moment of speech transaction (Davidson 1986b/Lepore 1986, p. 441; Davidson 1994b, p. 5). Davidson claims that this is necessary, since there are infinitely many utterances which the interpreter can interpret, and must be able to interpret, at any given time. However, as far as I can see, Davidson has himself undercut the reasons for taking this view. Even if, from purely syntactic considerations, we take seriously the idea that the competent interpreter must, at any given time, be able to interpret infinitely many possible utterances by the speaker, there is still no requirement that the interpreter must be able to give correct interpretations of any such possible utterance until after it has been made, with the possibility of gathering the required evidence after the event. And even if the interpreter does have a recursive prior theory, and the speaker does make an utterance correctly interpretable by that theory, there is no requirement, in the RI account of meaning, or, to my knowledge, in Davidson’s considerations on interpretation, that the interpreter rely only on that prior theory, and indeed, no requirement that he rely on it at all. He might as well start from scratch. Of course, he cannot start from scratch in respect of evidence. If he treats each new utterance as the first piece of empirical evidence, then he will not get anywhere. But he can always start from scratch in respect of theory. What matters is how well a theory measures up to the norms of interpretation, not how old it is. I agree with Davidson that actual interpreters do have an infinite competence, one which does require a recursive representation, or at least a recursive ingredient in its representation, and I think it is clear that they do use it, but I also think that the reasons for this view are quite external to the RI account of meaning. Which are the best theories? One natural way, and obviously the best way, if it can be had, to justify compositionality from the RI account of meaning, is to point out that, when we compare the different possible interpretations of a speaker, in respect of their satisfying the norms of charitable interpretation, it will be the compositional interpretations that come out with the best results. If I believed this to be the case, then I would not think that there is any problem. But I don’t believe it. The reason is simple. The norms of charity are supposed to be met by way of maximising, or ‘optimising’, certain properties of the speaker’s attitudes: for example, maximising the frequency of truths among his beliefs. And it is quite clear that it will be much easier to reach such a maximum if your interpretation isn’t hampered by the restrictions of compositionality. You can, if you want to,

PETER PAGIN 65

decide not to interpret any utterance as expressing a belief you disagree with, or a desire you find absurd. Or, if you prefer to think that the speaker is wrong about certain things, you can interpret his utterances so that he does reveal false beliefs about those things. Whatever the norms, whatever the properties you wish to maximise, you will find it easier to maximise those properties without compositionality; and whatever your background anthropological theory about the speaker, you will find it easier to interpret her as confirming that theory if you need not impose compositionality on your interpretation.16 So in all likelihood, ad hoc non-compositional interpretations will get higher marks than compositional alternatives. At this point you may object that the compositional theories stand out precisely by assigning meanings to the infinitely many sentences in the speaker’s language, and that the non-compositional alternatives don’t. I think that that objection is twice mistaken. First of all, no strong reason has so far emerged for thinking that the speaker does have a language with infinitely many sentences, each with a well-determined meaning. Until you can provide such a reason, you should not count it an advantage of a theory that it does assign meanings to infinitely many sentences. But second, the objection is technically wrong. You can, quite trivially, assign meanings to the infinitely many sentences which, syntactically, belong to the speaker’s language. To take the most trivial option, assign some arbitrary meaning, the same meaning, for example the meaning that the moon is made of cheese, to each of the infinitely many sentences to which the speaker still has not revealed any attitude. That will most clearly be wrong, but as long as the speaker has not revealed any attitude to any of them, there is no empirical evidence against it, and so no problem with charity. There will be evidence against it, as soon as the speaker, for example, uses one of those sentences to make an assertion. But when this happens, you simply revise your theory to accommodate the new data, by assigning a different meaning to the uttered sentence. There could not be anything wrong with that, from the point of view of the RI account of meaning, since that account does not require predictive success of interpretational theories. All that is asked for is successful accommodation, i.e. correct interpretation after the event. If predictive success had been required, then clearly ad hoc non-compositional theories would be quickly eliminated. But it was not, and your non-compositional theories still come out as the best.17 And if you say that you need to impose restrictions, in order to reduce underdetermination, the sound conservative line to take is to impose restrictions which allow you to keep your best theories, i.e. some of the non-compositional ones, and to eliminate others. So the non-compositional ones still come out as the best. If this conclusion stands, there is an even stronger result. It is not just that you cannot justify compositionality from the RI account of meaning. If it will always

66 INTERPRETATION AND COMPOSITIONAL STRUCTURE

hold that your best interpretational theories are non-compositional, then the proper conclusion is that natural languages are non-compositional. But this conclusion seems just wrong. The only reasonable explanation of why natural language speakers understand each other so often, so easily, in so diverse and complicated matters, and often with so little background knowledge about the other speaker, is that they understand each other through the grasp of the composition of what is spoken.18 For an unbounded number of sentences, ordinary speakers are in a position to predict what a sentence would mean in the mouth of a fellow-speaker, since in many cases the clues provided in the speech situation explain very little, and in many cases there are no such clues worth mentioning. Yet understanding is immediate, and thus cannot reasonably depend on anything other than knowledge that the interpreter had before being exposed to the utterance. Compositionality does essential explanatory work for actual interpretation, and compositional structure is an essential feature of the meaning of sentences that have it. But I have argued that compositionality cannot be derived purely from consideration about radical interpretation. If this is right, then the radical interpretation approach to linguistic meaning is methodologically misguided.19 Notes 1 By ‘compositionality’ I understand a feature of a language, by virtue of which the compositionality principle holds: the meaning of a complex expression is determined by the meaning of its parts and its mode of composition. As stated there is much that must be further specified in order to make it reasonably precise, not least as regards the term ‘meaning’. For now, by ‘meaning’ I shall intend semantic properties which are relevant for the present discussion, not necessarily intensional ones. I shall further assume that natural languages do exhibit syntactic structure (but not necessarily that there is always a uniquely correct syntactic analysis). 2 Note that I shall not claim that you cannot combine compositionality and radical interpretation. Clearly you can, and Davidson has impressively done so. I shall only claim that if radical interpretation is your only basic idea about linguistic meaning, then there is a conflict, for compositionality cannot be justified from that idea. 3 I don’t claim that Davidson has actually taken this position. 4 This way of putting the matter was suggested to me by Barry Smith. 5 That form can be ‘s means that p’ or ‘s is true iff p’ or something else. Which it shall be can perhaps not be decided in advance. Davidson’s original reason for preferring ‘is true iff’ over ‘means that’ was that the former is extensional, and hence allows that standard extensional logic may be applied in deriving the theorems. But the derivations are needed precisely in order for the theory to exhibit the compositional structure of the object language sentences. If there is no reason to believe in compositional structure, this reason for preferring ‘is true iff’ over ‘means that’ is not available. 6 This situation is depicted in the opening passage of Frege’s ‘Gedankengefüge’ [‘Compound Thoughts’), and Frege also provides compositionality as the

PETER PAGIN 67

explanation, in the form of a structural correspondence between the sentence and the thought. 7 Davidson explicitly says that the assumption that two speakers speak the same language is in need of justification. Speakers of the same language can go on the assumption that for them the same expressions are to be interpreted in the same way, but this does not indicate what justifies the assumption. All understanding of the speech of another involves radical interpretation. (Davidson 1973c/1984a,p. 125)

8 It is not, of course, a matter of predicting what a speaker will say. It is a matter of knowing what a sentence would mean if it were uttered by the speaker (barring, of course, slips of tongue and other aberrations). 9 Davidson 1984b/1984a, p. 278 and Davidson 1986b/Lepore 1986, pp. 440–4 It seems that in these papers Davidson has been thinking about prediction in terms of possible deviation from previous use. Understanding is possible in spite of deviation, which is why you don’t need conventions. But the possibility of deviation is irrelevant when compositionality is at stake. 10 You may abstract away from the idea of a real interpreter needing real observations, and think of the principle of charity as applied directly to the speaker’s attitudes, whether revealed or not. Still, the point against the RI account remains, since there would be no requirement of interpreting any sentence to which the speaker does not already have any attitude (like holding true). The point would be lost only on the assumption that a speaker in general has an attitude to every sentence of his own language. 11 In Davidson 1990e, p. 313, Davidson says, ‘Thus, a theory of truth is a theory for describing, explaining, understanding, and predicting a basic aspect of verbal behavior.’ There is no doubt that Davidson wants predictive capacity, as there is no doubt that he wants compositionality. But the question is how to get the requirement of predictive capacity out of the criteria for correct interpretation, when judged by charity. As far as I can see, you plainly don’t. 12 As far as I understand, Davidson slurs over this distinction when he writes ‘If we suppose, as the principle of charity says we unavoidably must, that the pattern of sentences to which a speaker assents reflects the semantics of the logical constants, it is possible to detect and interpret those constants’ (Davidson 1990e, p. 319). 13 The relation between holism and compositionality is studied in Pagin 1997. 14 I see no reason myself why this should be inconsistent with holism, even though, of course, it is quite implausible in itself. 15 In several passages, for example in Davidson 1977b/1984a, pp. 224–5, Davidson compares indeterminacy of meaning with measurement. His idea is that the difference between two acceptable T-theories for the same speaker is like the difference between the Fahrenheit and Celsius temperature systems. The numerical values are different, but the assignments are still equivalent. As the pattern of assignment of numbers to physical states is the same in these two systems, so the pattern of assignment of truth conditions is the same across acceptable T-theories.

68 INTERPRETATION AND COMPOSITIONAL STRUCTURE

16 You can, if you like, also impose a vacuous compositionality, i.e. one that satisfies a formal definition but could not serve any explanatory purpose. For examples of this, see Janssen 1997 and Westerståhl 1998. 17 It is natural to protest here that the compositional theory will be better than the proposed ad hoc rivals simply because it will be more true, i.e. contain a greater number of true meaning assignments. But recall that the criterion of truth of an interpretational theory, as far as this discussion goes, is given by the clauses of the RI account. A theory is true only to the extent that it conforms, as well as any theory, to the norms of charitable interpretation in its account of utterances made. 18 I think that Stephen Schiffer’s attacks, in Remnants of Meaning (1987) and many other places, on the idea that compositionality is required in an account of linguistic communication, are unsuccessful, but that is a topic for another occasion. Also, I cannot here go into the question whether speakers must be ascribed some tacit knowledge of compositional principles as part of the compositionality explanation of understanding. This question has been discussed, for example, by Gareth Evans, Crispin Wright and Martin Davies. 19 I am indebted to Hermann Cappelen, Donald Davidson, Kathrin Glüer, Sören Häggqvist, Mattias Högström, Mikael Janvid, Ernie Lepore, Dag Prawitz, Gabriel Segal, Barry Smith, Fredrik Stjernberg, Folke Tersman, Dag Westerståhl and Åsa Wikfors.

REPLY TO PETER PAGIN Donald Davidson There is much in this chapter with which I agree, and more that I admire. Many readers have sensed a conflict between my early concentration on compositionality and my subsequent insistence on the fact that speakers frequently invent or stumble on novel idioms, scramble their grammar, and also differ from one another in the ‘language’ they speak. Yet as interpreters we take most of this in our stride, usually hardly noticing the random malapropism, the deliberate new use for an old word, the forgotten tense or number a prior expression calls for further down the line, and so on. What adds to the feeling of conflict is that I did change the emphasis. In my first writing on the philosophy of language, ‘Theories of Meaning and Learnable Languages’ (1965a), I was mainly interested in the fact that the heated discussions about logical form, particularly of indirect discourse, seemed to proceed without anyone knowing what the rules were. Reading Tarski, I was struck by how neatly he solved the problem (for quantificational structure) of showing how the semantic values of sentences depended on the semantic properties of the parts of those sentences, and their structure. This led me to recognise that familiar proposals, by Church, Quine, Sellars and others, concerning adverbs, quotations, indirect discourse and other idioms, could not satisfy this natural demand; in fact, their proposals led in each case to an infinite primitive vocabulary. Working along these lines, I paid little attention to the question of how a theory which satisfied the requirement of compositionality might be shown empirically to be the language spoken by

PETER PAGIN 69

someone. But my next article, ‘Truth and Meaning’ (1967c), took up this question. From the start, I realised that a theory for interpreting a speaker had to be tailored to a particular speaker, and even relativised to a time, just as a theory of preference under uncertainty (another problem I was working on at the time) must be. I also realised, of course, as any sane person must, that though speakers of what we call the ‘same’ language differ in many details, we manage to understand them if we have a broadly correct compositional theory, and use our common sense in applying it. ‘A Nice Derangement of Epitaphs’ (1986b) is defective in a number of ways. It does not sufficiently stress the practical advantage that comes from a large degree of commonality among groups of speakers; the distinction between the prior and passing theories is clumsy, and makes a fairly simple point seem difficult; and there is, if not an actual confusion, a sort of fusion of two separable claims. The two claims are, first, that mutual understanding does not require that people speak the same language. It follows directly from this (obvious) point that shared ‘rules’ or ‘conventions’, however convenient they might be, cannot illuminate what is essential to linguistic communication. And of course there is no conflict at all between this claim and an insistence on compositionality. The second, and trickier, claim is that speakers can deviate not only from other speakers, but from their own earlier regularities, and can still (often) be understood. This is an ability that on occasion engages all our intelligence, sympathy, sense of humour and knowledge of people and the world. This second point, which surely records something we all know from experience, does force us to modify the idea of a rigid compositional structure learned in advance as a condition of successful interpretation. But of course the result can’t be that compositionality plays no role at all. Knowledge of a general compositional structure remains essential to the understanding not only of language but of belief and all the other propositional attitudes. What my attack on the standard picture of language, and the appeal to conventions and rules challenges is not compositionality, but the idea that a shared, fixed set of rules and conventions is basic to linguistic communication. Neither the fact that conversing speakers may not fully share idioms nor the fact that they often deviate from their own practice has any tendency to show that the semantic properties of their sentences do not depend on the semantic properties of their parts and their structure. Pagin puts much emphasis on the idea that compositionality doesn’t follow logically from my conception of radical interpretation. I can’t grasp the point of this claim. He characterises radical interpretation as having two elements: that an utterance ‘means what it can be successfully interpreted as meaning’ and that ‘Success of an interpretation is judged by the degree that the interpretation conforms to norms of charitable interpretation’. He doesn’t view the matter at all as I do. I assume (on the basis of learnability, creativity and common sense) that any successful interpretation of an utterance must fit that utterance into a compositional scheme. And the ‘norms of charitable interpretation’ (by which I

70 INTERPRETATION AND COMPOSITIONAL STRUCTURE

guess he means in accord with the ‘principle of charity’) include, and have always included for me, at least these two elements. First, a sufficient number of what Quine calls observation sentences are, as a start, interpreted as being true in the situations that prompt them. (I say ‘as a start’ because as theory grows, putting the pieces together will reveal cases where a speaker errs.) This is the simple-minded idea that if a speaker says (apparently believing what he says) ‘Gavagai’ when and only when a rabbit is in sight, then that’s a reason, subject to revision, for interpreting this utterances as meaning ‘There’s a rabbit’. That’s one application of ‘charity’. Another part enters in the interpretation of the logical constants, as Quine explained. Charity plays a role also in the interanimation of sentences, not only in constraining us to interpret speakers as basically rational (consistent) logically, but also in aiding the interpretation of sentences containing theoretical terms by noting the degree to which assent to them is increased or decreased by assent to other sentences evidentially related. Clearly, on my understanding of radical interpretation, compositionality is built in at every stage. Does Pagin perhaps think, as some of my other commentators seem to think, that for me charity just enjoins us to make as good sense of each utterance, taken in isolation, as we can? Or that charity simply means ‘maximise truth’? I urge such readers to reread the introduction to Inquiries into Truth and Interpretation (1984a). The reason why compositionality and ‘radical interpretation’ cannot ever conflict, given my understanding of radical interpretation, is that I view it as a given that any theory of meaning is compositional, and then, and only then, ask how we can tell that a speaker is speaking in accord with a specific compositional theory. It is only at this point that radical interpretation has a role to play. This picture may not always have been as plain in my writings as it should have been, but it has always been there, and it has certainly been the position I have been at pains to emphasise in all my writings in recent years.

Part III Semantics and its applications

6 From semantics to ontology, via truth, reference and quantification Stephen Neale1

If we have the semantics of a language right, the objects we assign to the expressions of the language must exist. (Donald Davidson) Preliminary remarks on reference and ontology It would be natural to say that the names ‘Maurice’ and ‘London’, as they occur in an utterance (or inscription) of the sentence ‘Maurice’s mother lived near London’ refer to things (a person and a place, respectively), and that someone using these words would be so referring in normal circumstances. It would also be natural to say that the occurrence of ‘Maurice’s mother’ in an utterance (or inscription) of the sentence refers to something (although wrong philosophically if Russell is right), and that someone using these words would be so referring. In sharp contrast, it would be decidedly unnatural to say that any of ‘lived’, ‘near’, ‘near London’, ‘lived near London’, or ‘Maurice’s mother lived near London’ referred to things, or that someone using these words would be so referring. But of course this will not prevent philosophers and linguists from arguing that verbs, adjectives, prepositions, prepositional phrases, verb phrases and even sentences do refer to things. A theoretical notion of reference may have to replace our ordinary notion as far as a theory of meaning is concerned. And it is no news that the recent history of semantics is littered with theories according to which functions, sets, properties, types, propositions, states of affairs, situations and facts serve as the references of all manner of expression. From the point of view of constructing a systematic and comprehensive semantics for a natural language, it is not an intuitive or obvious requirement that a final theory appeal to the idea that anything other than names and other singular terms have references. It makes no sense to stipulate in advance of detailed investigation that any adequate theory will, or will not, treat, say, sentences as referential devices. Our ordinary uses of words such as ‘refer’, ‘reference’, and so on suggest that it would be at least a little unnatural to view sentences (or even sentences relative to contexts) as referring to things (except in

STEPHEN NEALE 73

the sense in which we say, loosely, when examining a text that a particular sentence or verse refers to, say, a recent historical event). But of course the semanticist who appeals to the idea that sentences have references does not have to be entirely shackled by ordinary usage. Certainly we do well to begin with ordinary usage and depart from it grudgingly; but it is well attested that although (many of) our theoretical terms (in both the sciences and the humanities) derive from ordinary usage, as theorising becomes more complex, such terms may begin to diverge in meaning from their ordinary language counterparts, a normal pattern of linguistic development in any interesting discipline. Donald Davidson (1984a; 1990e) has argued (a) that we can go about constructing theories of meaning that do not assign references to anything other than our standard referential devices (singular terms, including variables under assignment); (b) that such theories will take objects and events, construed as particulars, to be the sorts of things to which singular terms refer; (c) that as far as the theory of meaning is concerned, we can free ourselves of the idea that predicates refer to properties and the idea that individual sentences refer to, stand for, correspond to or are made true by particular facts (or by particular states of affairs or situations that obtain). In addition, Davidson has presented a ‘slingshot’ argument designed to show (d) that even if we wanted facts to serve as the references of true sentences, we could not have them; and he has argued (e) that if there are no facts, then we cannot make sense of ‘correspondence’ theories of truth, for such theories are built upon the idea that a sentence is true if and only if there is some fact to which it corresponds. Meaning theory, truth theory and truth definition A very useful way of characterising Davidson’s approach to meaning is as follows (Davidson does not state it quite like this, but I think what I am about to say is equivalent in all crucial respects to what he says). Point (a) is that a theory of truth for a language L is any theory that entails, for each sentence S of L, a true theorem of the form (1) S* is true (in L)↔p a sentence in a metalanguage M used to talk about L in which ‘S*’ is a structural descriptive name (in M) of S and ‘↔’ is the material biconditional. (For the moment, let us ignore the existence of indexical and other context-sensitive expressions.) Instances of (1) are called ‘T-Sentences’. Point (b) is that a theory of meaning for a language L is, for Davidson, any theory, knowledge of which would suffice for interpreting the utterances of L-speakers (made in L, of course). So (c) a theory of truth θ for a language L will qualify as a theory of meaning for L if knowledge of what θ states would suffice for understanding the utterances of L-speakers. We can call this the interpretation requirement on truth theories. Not every theory of truth for L will satisfy it. Imagine a theory of truth θ

74 FROM SEMANTICS TO ONTOLOGY

for French that delivered just the following T-sentence for the sentence ‘La neige est blanche’: (2) ‘La neige est blanche’ is true (in French)↔Milan is north of Rome Since the left- and right-hand sides of this biconditional are both true, (2) itself is also true. Knowing (1) would not suffice for understanding utterances of ‘La neige est blanche’. Taken by itself, this fact is of no import. But in the context of θ it has some bite. For θ, although delivering a true theorem of the form given in (1) for every sentence of French (including the sentence ‘la neige est blanche’), would not qualify as a theory of meaning for French because knowing θ would not provide the means of interpreting utterances of ‘La neige est blanche’. It is this fact that renders θ inadequate as a theory of meaning although it is impeccable as a theory of truth. One way of constructing a theory of truth that might seem to qualify as a theory of meaning would be to stipulate that in every T-sentence, the sentence of M that replaces ‘p’ must have the same meaning as the sentence that replaces ‘S’. A theory of truth satisfying this condition would be what Tarski (1956a) called a (materially adequate) truth definition for L (a definition of ‘true-in-L’). Tarski could help himself to the notion of sameness of meaning (or translation): his objective was to define (i.e. characterise) truth (in L), not meaning. But an appeal to sameness of meaning (or to translation) is not available to Davidson, nor does he claim it is: Davidson’s objectives are (a) to say what a theory of meaning is, and (b) to show how one might go about constructing one, and to do all of this without appealing to ‘meanings’, whether of whole sentences or of their parts. The crucial question, then, is whether it is possible to construct a theory of meaning for L without appealing to, say, sameness of meaning. Davidson claims it can be done: a truth theory θ for L will qualify as a theory of meaning for L if certain empirical and formal constraints are imposed on the way θ is constructed. The precise details of the constraints are irrelevant for present concerns. What is relevant is that the joint force of the constraints is supposed to ensure that knowledge of θ (together with the knowledge that θ is a truth theory for L) would suffice for understanding utterances of sentences of L. To speak roughly, the sentence replacing ‘p’ in (T) will always (it is hoped) have ‘the same meaning’ (intuitively speaking) as the sentence replacing S; more carefully, there is nothing more to being able to interpret the utterances of Lspeakers than knowing a truth theory constructed in accordance with the constraints in question. There is no circularity in this proposal: it is no part of either the truth theory itself or the constraints imposed upon the construction procedure that the sentence replacing ‘p’ gives the meaning of, or constitutes, a translation of S. (Thus an interesting difference between the projects of Davidson and Tarski: whereas Tarski can take the undefined semantic notion of translation for granted and use it to place constraints on what counts as an adequate definition of truth

STEPHEN NEALE 75

for L, Davidson uses a primitive notion of truth in order to characterise what he takes to be the most plausible way of constructing a theory of meaning.) Reference, quantification and ontology I want now to explain a double asymmetry between sentences and singular terms in Davidson’s programme. Notice that no appeal to the notion of reference (or predication, for that matter) has surfaced in the discussion of what a Davidsonian theory of meaning will look like. A theory’s deliverances with respect to whole sentences are all that matter: it is irrelevant how the internal workings of the theory treat the parts of sentences, as long as things come out right for the wholes: Since T-sentences say nothing whatever about reference, satisfaction, or expressions that are not sentences, the test of the correctness of the theory is independent of intuitions concerning these concepts. Once we have the theory, though, we can explain the truth of sentences on the basis of their structure and the semantic properties of the parts. (Davidson 1990e, p. 300) In principle, then, it might be possible to construct an adequate theory that doesn’t utilise a notion of, say, reference at all. Only attempts to build adequate truth theories will tell. According to Davidson, a theory of meaning for a language such as English will, in fact, utilise a lean, formal notion of reference, by virtue of containing axioms capable of handling quantification. If we were dealing with an infinite extensional language L containing only (a finite number of) names, predicates and sentence connectives, but no quantifiers, it would not be necessary to invoke a concept like reference: a finite theory consisting of one axiom for each atomic sentence and one recursive axiom for each sentence connective would suffice because L would contain only finitely many atomic sentences. So it is not the existence of singular terms per se that foists reference upon us. According to Davidson, it is quantification that forces us to take something like reference seriously, because it is quantification that forces us to abandon the construction of a straightforward theory whose axioms take the form (3) __ is true↔… in favour of a theory that takes a detour through satisfaction (‘a generalised form of reference’), a theory whose predicate and quantifier axioms take the form (4) ( s)(s satisfies __↔…s…) where ‘s’ ranges over (infinite) sequences of objects.

76 FROM SEMANTICS TO ONTOLOGY

According to Davidson, the need for such axioms is ontologically significant. The ‘logical form’ of a sentence S belonging to a language L is, for Davidson, the structure imposed upon S in the course of providing an adequate truth theory for L as a whole; and a truth theory is adequate only if it satisfies the interpretation requirement. Detailed empirical work on the syntax and semantics of natural language suggests that we will not get very far in our attempts to construct an adequate truth theory for English unless we view the logical forms of English sentences as encoding something very like the quantifier-variable structures familiar from formal languages such as the first-order predicate calculus. One formally useful way of pulling together names and variables (and other singular terms, if there are any) within a truth theory—a notational variant of many other ways, and perfectly consonant with Davidson’s approach—is to use the notion of reference relative to a sequence, which we can abbreviate as ‘Ref’. On this account, the axiom for the name ‘Maurice’ is given not by a simple reference axiom such as (5) but by (6): (5) The referent of ‘Maurice’=Maurice (6) ( s)(Ref(‘Maurice’, s)=Maurice) where ‘s’ ranges over sequences. Axiom (6) is a ‘reference’ axiom in the sense that it connects a piece of language with some entity or other, some piece of the world. (For ease of exposition, let us put aside names that allegedly fail to refer, if there are such expressions.) Axioms for individual variables are given by the axiom schema (7): (7) ( s)( k)(Ref( xk ,s)=sk) where ‘s’ ranges over sequences and sk is the k-th element of s. (The difference between a name and a variable, then, is that the Ref of a name is always the same entity, whereas the Ref of a variable depends upon the sequence in question. The constant/variable distinction should not be confused with the rigid/non-rigid distinction: variables (with respect to sequences) are just as rigid as names (in Saul Kripke’s sense: 1980). Two interesting things about reference axioms should be noted: (a) they pair bits of language with bits of the world (this gives them their ontological bite); and (b) the fact that reference is no more than such a pairing—any pairing that works—means that there is no need to provide an analysis of the reference relationship. Questions about whether Davidson’s story about reference conflicts with those of, say, Searle 1983; Donnellan 1971; Kripke 1980; or Evans 1982, don’t make a great deal of sense. Axioms for verbs, adjectival expressions and ordinary common nouns treat such expressions as predicates (for ease of exposition, let us put aside so-called ‘intensional’ predicates such as ‘fears’, ‘fake’, and so on).2 For example:

STEPHEN NEALE 77

(8) ( s α)(s satisfies α snores ↔Ref(α, s) snores) (9) ( s α β)(s satisfies α likes ß ↔Ref(α, s) likes Ref(β, s)) The utility of the axioms flowing from (7) lies in their interaction with axioms for quantifiers. If quantification in English turns out to be unrestricted and completely analysable in terms of the first-order quantifiers ‘ ’ and ‘ ’—a more realistic proposal will be considered shortly—a truth theory for English can get by with an axiom like (10), and its universal counterpart: (10 ( s)( k)( Φ)(s satisfies ( xk)Φ ↔there is at least one sequence differing ) from s at most in the k-th place that satisfies Φ The theory is completed by adding that a sentence is true if and only if it is satisfied by every sequence. Assuming an adequate background logic (e.g. extensional, first-order logic with identity), we could then prove theorems (T-sentences) such as the following: (11 ‘Maurice is Greek’ is true↔Maurice is Greek ) (12 ‘Maurice likes everything’ is true↔Maurice likes everything3 ) If it turns out that a theory of truth that is to succeed as a theory of meaning cannot be constructed without making use of some notion of reference (such as Ref above), i.e. without axioms that connect pieces of language with nonlinguistic entities, let us say that such a notion is ‘theoretically ineliminable’. According to Davidson, accepting that reference is theoretically ineliminable does not mean accepting that any particular set of reference axioms is ineliminable. Indeed, it is Davidson’s position, as it is Quine’s, that it is possible to transpose any adequate set of truth-theoretic axioms X into another adequate set Y that contains, as a subset, a set of reference axioms quite different from those contained in X (the predicate axioms would also differ, of course).4 A particular axiomatisation is tested only by its proper theorems, i.e. by its deliverances at the level of whole sentences. So the notion of reference employed by Davidson is philosophically lean in two senses: (a) no particular set of reference axioms is privileged; and (b) to claim that a well-understood notion of reference is appealed to by Davidson in constructing a theory of meaning for a language is not to claim very much; for on Davidson’s account to say that ‘Maurice’ refers to Maurice is just to say that there is a successful axiomatisation containing the axiom (6) above—or some notational variant that hooks up ‘Maurice’ and Maurice— where ‘Ref’ is straightforwardly an extensional functor or a description-forming operator. Two points need to be made in this regard. First, as Davidson (1993b.) stresses, the fact that an adequate truth theory might make an ‘unnatural’

78 FROM SEMANTICS TO ONTOLOGY

assignment of objects to individual words does not affect the overall ontology to which the language is committed. (Of course, no successful assignment is really ‘unnatural’ on Davidson’s account.) Second, no appeal to modes of presentations, causal chains, informational packages, or intentionality is needed in order to characterise the theoretical notion of reference that Davidson employs. As Richard Rorty observes, for Davidson ‘any ‘theory of truth’ which analyses a relation between bits of language and bits of non-language is already on the wrong track’ (Rorty 1986, p. 333). There is a crucial difference between the reference axioms and the predicate axioms in a successful truth theory. The former assign particular entities (individuals) as the semantical values (references) of expressions; the latter do no such thing. This is important. On Davidson’s account, since reference axioms are theoretically ineliminable, we must accept the entities that our singular terms stand for, even though we do not have to regard any particular reference function as privileged. Predicate axioms invoke no new entities: the only ontologically significant notion they use is satisfaction, and satisfiers are just sequences of objects. By hypothesis, a successful axiomatisation construes singular terms as standing for objects, and construes sentences as satisfied by sequences of objects. Hence Davidson’s ‘realism’ about ‘the familiar objects whose antics make our sentences and opinions true or false’ (Davidson 1984a, p. 198). On Davidson’s account, events—taken to be unrepeatable particulars—will get into the picture along with objects because a successful axiomatisation will have to deal with sentences which involve quantification over events, for example (13) and (14): (13 There was a fire and there was a short-circuit ) ( x)fire x & ( y)short-circuit y (14 There was a fire because there was a short-circuit ) ( x)(fire x & ( y)(short-circuit y & y caused x) Sentences containing action verbs and adverbs (‘John left quickly’) and those containing bare infinitives (‘John saw Mary leave’) also appear to require quantification over events.5 In addition to claiming that we will need to pair bits of language with objects and events as far as constructing a theory of meaning is concerned, Davidson suggests that if the need to posit a particular ontological category does not arise in the construction of an interpretive truth theory, then the need cannot arise at all. (Of course, linguistic categories and set theoretic entities such as sequences are posited by the metalanguage.) The thought behind this suggestion appears to be (roughly) the following (although I have not actually found it stated quite this way anywhere in Davidson’s work): A theory of truth for L, of the sort that can serve as a theory of meaning for L, delivers a true theorem of the requisite form for every sentence of L. So there is nothing one can say in L that outstrips the ontology revealed by an interpretive truth theory for L; so there is no sense to be made of ontological categories not forced upon us by

STEPHEN NEALE 79

the construction of a viable semantics. (It might be objected that in doing semantics we can appeal only to entities that we think exist, so a semantic theory offers us no more by way of ontological insight than ordinary reflection upon our thought and talk. But this misses the fact that an adequate semantics is systematic in ways in which ordinary reflection is not.) One interesting question that is left open concerns competing truth theories with different ontologies. If θ1 posits As, Bs and Cs, while θ2 posits only As and Bs, then if both are adequate truth theories for L, we have reason to posit only As and Bs. In this case our ontology is given by the intersection of the things posited by the competing theories. But what if θ2 posited only Bs and Cs? Perhaps it is unlikely that we will find ourselves in such a situation (we are unlikely to find even one theory that covers all the data), but the question is still of philosophical interest, if we take seriously the idea that ontology flows from semantics in the way that Davidson suggests. Presumably Davidson will prefer a theory that posits objects and events over one that posits (for example) events and properties on the grounds that (a) identity conditions for objects and events are considerably clearer than they are for properties, and (b) our best accounts of nature and our most cogent statements of traditional philosophical problems concerning (for example) causation, time, change, human action, and the mind-body problem appear to presuppose the existence of objects and events. Potential challenges and modifications I have tried only to present clearly the bare bones of Davidson’s ambitious programme. I want to conclude with a partial catalogue of ontological issues that could, in principle, have a bearing on the final shape of an interpretive truth theory. Davidson proposes an elegant semantics and a tidy ontology of objects and events. There are no facts, no situations, no states of affairs, no propositions and no properties because we can construct interpretive truth theories without positing them. Of course Davidson does not claim to have a proof that we need only objects and events. His claim is simply that there is no good evidence to date that we need posit more than objects and events. In a way, Davidson is throwing down the gauntlet: ‘Show me sentences that appear to require more than objects and events and I will do my best to show that objects and events suffice.’ Context-sensitivity Truth-theoretic Axioms for words such as ‘I’, ‘me’, ‘we’, ‘us’, ‘you’, ‘he’, ‘him’, ‘she’, ‘her’, ‘it’, ‘they’, ‘them’, ‘this’, ‘these’, ‘that’, ‘those’, ‘today’, ‘yesterday’, ‘tomorrow’, ‘here’, ‘there’, ‘now’, ‘then’, ‘hitherto’, ‘henceforth’, ‘present’, ‘current’, ‘contemporary’, ‘previous’ and ‘next’ must take into account contextual features of one sort or another. In view of this, the Davidsonian will naturally treat truth as a property of utterances rather than sentences, and then treat utterances as individual events. Simplifying somewhat, on such an account, axioms will be rewritten with metalanguage

80 FROM SEMANTICS TO ONTOLOGY

semantic predicates taking on a parameter for the utterance event e, giving us, for example, (15) for ‘Maurice’ and (16) for ‘I’: (15 ( s)( e)(Ref(‘Maurice’, s, e)=Maurice) ) (16 ( s)( e)(Ref(‘I’,s,e)=eu) ) where eu is the utterer, the person producing the utterance (e.g. the speaker or writer).6 Properties A sentence like (17) might be thought to present Davidson with an interesting challenge as it appears to involve quantification over colours, construed as properties: (17 This is the same colour as that ) The Davidsonian who does not want an ontology of properties must either find a way of understanding colours as things other than properties, or else show that once the logical form of (17) is revealed, there is no quantification over colours. Facts If we were to find that we needed to allow variables to range over, say, facts or situations in order to provide an adequate truth theory, then on Davidson’s account, facts or situations would also be part of our ontology. But Davidson suggests (a) that the need will not arise, and (b) that entities like facts and situations (under their most common construals) are ruled out independently, as can be demonstrated by a ‘slingshot’ argument. In the case of (a) the sorts of sentence that might tempt one to posit facts include the following: (18 The fact that there was a short-circuit caused it to be the case that there was ) a fire Davidson suggests that (12) above gives the logical form of (18) so an appeal to facts is unnecessary. A more difficult case for Davidson might be: (19 The fact that Mary left Bill’s party did not worry him, but the fact that she left ) so suddenly did Davidson is surely correct that events, just like objects, can be described in a myriad of ways. If Mary’s leaving was Mary’s sudden leaving, then how is it that (19) can be true? A promising approach to this problem utilises Grice’s (1989) notion of conventional implicature in an attempt to provide the basis of an explanation of our judgements concerning (19).

STEPHEN NEALE 81

With regard to (b) Davidson (1984a; 1990e; 1993b) argues that we cannot have facts of any significance because of a simple formal argument that has come to be called the ‘slingshot’ and which can be cast in a formal mode by (i) assuming that ‘the fact that…caused it to be the case that…’ is a non-truthfunctional sentence connective and then (ii) deriving the conclusion that it is truth-functional. I have discussed this type of argument in detail elsewhere (Neale 1995; Neale and Dever 1997). As far as facts are concerned, my main conclusion may be summarised thus: slingshot arguments do not actually rule out facts but they do impose very definite constraints on what theories of facts must look like, constraints that many proposed theories are incapable of satisfying. (Russell’s theory of facts passes—assuming his Theory of Descriptions; the theories of Wittgenstein and Austin fail.) Modality and possible worlds A truth theory for English must be able to handle sentences that contain modal adverbs such as ‘necessarily’ and ‘possibly’. Will this require an ontology of ‘possible worlds’? If the axioms of a truth theory are modalised, a richer background logic (a suitable modal logic) can be used to derive theorems of the requisite form while treating ‘necessarily’ and ‘possibly’ as non-extensional sentence connectives (Peacocke 1978). Of course, Davidson himself has little time for such connectives, but in view of the fact that Quine’s worries about the modalities were overstated, and the fact that slingshot proofs do not actually finish off modal sentence connectives, there is good reason to think that possible worlds are not required. Higher-order quantification It is well known that natural language quantification is more complex than it is in standard first-order languages. However, work in generative linguistics and mathematical logic has revealed elegant methods of extending Tarski’s insights, so that truth theories can be provided for quantified fragments of natural languages while making precise the relationship between the superficial grammatical form of a sentence and its logical form in Davidson’s sense (Wiggins 1980). Many of the details need not concern us here. Suffice to say that noun phrases such as ‘some man’, ‘every Athenian’, ‘no farmers’, ‘most tall soldiers’, etc. can be viewed as restricted quantifiers composed of quantificational determiners (‘every’, ‘some’, ‘no’, ‘most’, etc.) combined with simple or complex nouns (‘man’, ‘tall man’, etc.). Axioms such as the following make things run very smoothly: (20 ( s k Φ ψ)(s satisfies [some xk: Φ] ψ ↔ some sequence satisfying Φ ) and differing from s at most in the k-th place also satisfies ψ) (21 ( s k Ф ψ)(s satisfies [most xk: Φ] ψ ↔ most sequences satisfying Φ ) and differing from s at most in the k-th place also satisfies ψ) It is not necessary to pack any additional set theory or overtly higher-order machinery into the right-hand sides of (20) and (21); they have the same form as axioms for the traditional unrestricted quantifiers modulo the relevant restriction concerning the satisfaction of Φ.

82 FROM SEMANTICS TO ONTOLOGY

Notes 1 I thank Donald Davidson, Ernie Lepore, Peter Pagin and Bruce Vermazen for comments on ancestors of this paper. 2 There are notorious difficulties involved in providing a uniform predicational analysis of these categories. For present purposes, the differences between these categories can be ignored as they do not raise problems that bear directly on my main theme. 3 To be precise, we could prove such theorems using extensional first-order logic with identity if we treated ‘Ref (“Clinton”,s)’ either (a) as a Russellian definite description (‘the referent of “Clinton” with respect to s’) and hence as a first-order definable device of quantification; or (b) as a complex singular term formed from a singular term and a functional expression. If method (a) is selected, proofs will make use of Whitehead and Russell 1927 *14.15, a derived rule of inference for extensional contexts, rather than straightforward applications of the Principle of Substitutivity. If method (b) is selected, a version of first-order logic with functors must be selected. There are, I believe, reasons for preferring method (a). 4 It is debatable whether interpretive truth theories for languages containing Russellian descriptions together with words such as ‘necessarily’ and ‘possibly’ allow of the sorts of straightforward reference permutations Davidson has in mind. This does not matter for present concerns. 5 On these matters, see Davidson 1980a and Higginbotham 1983. 6 For sophisticated and more plausible implementations, see (e.g.) Burge 1974; Weinstein 1975; Taylor 1980.

REPLY TO STEPHEN NEALE Donald Davidson Neale gives a masterful and sympathetic account of my views on the sort of theory I favour as the basis for interpreting a speaker, and in particular of the relation between such a theory and ontology. He also raises a number of problems or challenges that arise in the attempt to apply such a theory to a natural language with normal expressive power. One such problem he mentions only suggests that perhaps the constraints on a satisfactory theory can solve it. Since Paul Horwich also mentions this problem (which he apparently thinks insoluble), let me say a few words about it. The problem is how to eliminate such T-sentences as (1) ‘Snow is white’ is true↔grass is green or Neale’s (2). I have from time to time made suggestions on how to cope with this difficulty, not all of them consistent with each other. My present view is that there is no point in trying to find a strictly formal solution. What is wrong with (1) is that it will not fit into an empirically satisfactory theory because it will lose track of relations among sentences. Simple inferences such as that from ‘This is

STEPHEN NEALE 83

snow’ and (1) to ‘This is white’ will not be validated by our theory, since it must treat ‘snow’ as ambiguous. (There is nothing wrong with (1) or Neale’s (2) if they are embedded in theories which systematically alter the extensions of all predicates and names in such a way as to preserve all inferences. Those who raise this difficulty don’t have such cases in mind; they think of the rest of the theory as providing the ‘standard’ translations.) The thing to bear in mind is that a theory of truth, as I want to employ it, is an empirical theory, and is constructed in the light of certain intuitively obvious kinds of evidence. These sources of evidence are mainly three: a matching of perceptual sentences like ‘This is snow’, ‘That is white’ to their socially identified causes; the logical relations inferred to hold among sentences; and relations of evidential support among sentences. (I have spelled out how I think this can work in ‘The Structure and Content of Truth’: 1990e.) I now address the challenges Neale sees to my relatively simple semantics. Context-sensitivity I recognised from the start that the truth conditions of many sentences are sensitive in systematic ways to context, and so a theory of truth for a natural language would have to treat utterances as the elements with which the theory treats. I made a simple suggestion of how this might go in ‘Truth and Meaning’, and my then-student Scott Weinstein made a far more sophisticated proposal. No doubt there is more work to be done. Properties Neale seems to me to confuse two issues about properties. The first is whether there is in general any objection to including properties in our ontology. The second is whether there is any advantage in introducing predicates as the sole semantic item used to explain the function of predicates. I have no objection to accepting properties, if they turn out to be needed to explain such sentences as ‘This is the same colour as that’. I have no general objection to abstracta, as long as they are shown to be useful; as I always say, they take up no space. I have deep misgivings, however, about semantic theories such as Russell’s, which treat predicates as referring to, or corresponding to, predicates. These misgivings have two related sources. First, such theories are apt, as Plato discovered, to lead to regresses. The second is that it is essential to my idea of how a theory of truth is verified that the right side of T-sentences not employ conceptual resources not employed by the sentence for which the truth conditions are being given. This constraint is obviously violated in the treatment of indexicals, but not in the basic way it would be violated if the work of every predicate were explained by appeal to a property (relation, etc.). Facts Neale’s splendid discussion of the slingshot argument against facts, while it shows the limitations of that argument in both Church’s and Gödel’s versions (both of which I repeated in various places), has encouraged me to stick to the conclusion that facts cannot be incorporated into a satisfactory theory of truth. Neale speaks of Russell’s view of facts as ‘acceptable’, but he does not show that it can be incorporated into a Tarski-style truth theory. Modality and possible worlds A satisfactory theory of a natural language must say something about modal adverbs, and I have not published anything on the

84 FROM SEMANTICS TO ONTOLOGY

subject. Neale is right that I am unhappy with non-extensional sentential connectives; I am even more unhappy with analyses that depend on our intuitions about possible worlds. I do think there are ways to explain in a nonintensional way how modal adverbs function, but here is not the place to explore this. Higher-order quantification I can only applaud the suggestion that the semantic functions of expressions such as ‘some man’ and ‘most people’ can be accommodated within a harmlessly expanded version of a theory of truth of the sort I favour.

7 Semantics for quotation1 Herman Cappelen and Ernie Lepore

Our quotation practices are multifarious. We can directly quote another, as in (1); indirectly quote him, as in (2); or we can purely quote an expression in order to talk about it, as in (3). (1) Clinton said, ‘I’ll cut taxes’. (2) Clinton said that he’d cut taxes. (3) ‘Lobster’ is a word. These are all philosophically familiar. There is another sort of quotation, however, which, though rarely discussed in the philosophical literature, is probably the most utilised form of quotation in natural language. We will call this mixed quotation, as in (4). (4) Clinton said that he’d ‘cut taxes’. The semantics of these four sorts of quotation cannot be entirely disunited. Any utterances of (2) and (4) respectively agree in one sense about what was said. This all by itself indicates: C1 Mixed and indirect quotation should receive overlapping semantic treatments. Similarly, any two utterances of (1) and (4) partially agree in another sense about what was said. This supports: C2 Direct and mixed quotation should receive overlapping semantic treatments. However, utterances of (1) and (2) might not be in agreement about what was said in any sense, at least not unless we assume that (1) attributes (understanding of) an English utterance to Clinton. For all we know, (1) may be ascribing to him an utterance of some other language that merely happens to share certain lexical

86 HERMAN CAPPELEN AND ERNIE LEPORE

items with English; or Clinton may not know any English but merely uttered ‘I’ll cut taxes’ in an effort to clear his throat. From all this we should conclude that: C3 Direct and indirect quotations must receive distinct semantic treatments. Finally, notice that (1) and (4) imply pure quotations (5) and (6) respectively: (5) A token of ‘I’ll cut taxes’ was uttered. (6) A token of ‘cut taxes’ was uttered. These data lend support to: C4 Quotations in pure, direct, and mixed quotation should receive overlapping semantic treatments. The chief aim of this chapter is to develop an account that satisfies constraints C1 to C4. We will begin by raising doubts about whether leading semantic accounts for indirect, and for pure (and direct) quotations do or can easily be refined so as to accommodate mixed cases. Semantics for indirect quotation (A) and (B) are standard assumptions among semanticists of attitudinal reports: (A) Propositional attitude reports assert that a relation obtains between an agent and a proposition (or a proposition-like content); (B) Any propositional attitude report is true only if the proposition (or proposition-like content) expressed by the report’s complement clause matches the proposition (or proposition-like content) of the agent’s attitude (or utterance). Assumptions (A)–(B) require that (4) is true just in case Clinton uttered something matching the propositional content of its complement clause. Mixed cases, we believe, make it difficult for any theory incorporating (A)–(B) to be inadequate. So, for example, according to Soames, (4) is true iff Clinton uttered a sentence S in an associated context C such that for some S′ that can be readily inferred from S, the content of S′ in C=the content of ‘I’ll cut taxes’ in the context of the report (Soames 1989, p. 411). Extending Soames’ account to mixed quotation, an utterance of (4) correctly reports Clinton’s utterance u, say, of ‘I’ll cut taxes’ just in case ‘he’d “cut taxes’” expresses the same proposition as (an utterance of) a sentence that can be readily inferred from ‘I’ll cut taxes’. Even if (as we think unlikely) that ‘I’ll “cut taxes”’ expressed a proposition, how could it express anything readily inferable from ‘I’ll cut taxes’? (T1) and (T2) are obvious truths about (4) and u.

SEMANTICS FOR QUOTATION 87

Figure 1.1

(T1) Clinton was talking about taxes; he was not talking about words. (T2) Quotation (4)’s complement clause includes quotation marks and therefore, in some sense, says something about linguistic entities. So, if (4) were an example of indirect quotation, we could indirectly quote another without using a complement clause matching the content of the reported utterance, contrary to Soames’s account. Here’s another challenge mixed quotation presents for accounts of the semantics of indirect speech respecting (A)–(B). Suppose propositional attitude reports relate an agent to an Interpreted Logical Form (ILF).2 An ILF is a logical form in the sense of Chomsky (1981) augmented by semantic values at each node in the phrase marker. So, an ILF effectively incorporates information about the proposition expressed by an utterance, the logical structure of that proposition, and the lexical means by which that proposition is expressed. Adapting an example from Larson and Segal (1995, pp. 438–40), ‘Peter said that Lori met Cary Grant’ is true just in case what Peter said was as represented in Figure 1.1, where semantic values appear in the square brackets to the right of each node. Mixed, pure and direction quotation have not yet been discussed among ILF theorists. But were an ILF theorist to treat quotes as functioning in the same way and as having the same semantic value in whatever linguistic context they occur, a constraint, following Davidson, that Larson and Segal (1995, pp. 36–7) and Larson and Ludlow (1993, p. 32) explicitly endorse, then, for mixed cases like (4), would be that expressions quoted must be annotated as semantic values in the

88 HERMAN CAPPELEN AND ERNIE LEPORE

ILF of (4)’s complement clause (because the semantic value of a quotation is the quoted expression) and that it must contain the quotation itself as a lexical element. This requires that there must be a node in which ‘cut’ occurs as a lexical item and the semantic value of ‘cut’, whatever that may be, occurs as well. So (4) would be claiming that Clinton stands in the saying-relation to such an ILF. However, Clinton does not bear this relation to any ILF as described above. ILF theorists disagree among themselves about whether standing in the sayingrelation to an ILF is to any extent pragmatically determined.3 We ourselves support appeals to pragmatic considerations in an effort to characterise our practice of indirect quotation (and other propositional attitude attributional practices). But these appeals cannot solve the problem posed by mixed quotation. It is not a pragmatic issue whether a mixed report requires the reported speaker to stand in the saying-relation to an entity containing lexical elements as semantic values. Typically, a mixed quoted speaker said nothing about words and it is a straightforward semantic fact that s/he did not. Any theory that does not respect this fact, or tries to dodge it by deporting it to pragmatics, is inadequate. So, to sum up, whatever its other merits, ILF theory establishes no advance over Soames’ account with respect to mixed cases. Whatever content, if any, is expressed by the complement clause of a mixed case, it must be at least partially about words. But any person correctly reported by a mixed quote such as (4) has not uttered anything about words. So no theory requiring for the truth of an indirect quotation that the reported speaker stand in the saying-relation to the proposition expressed by the complement clause of the report can accommodate mixed cases. One reaction to mixed cases would be, for example, applied to Soames’ account, to give whatever contribution quotation marks make, so to speak, wide scope over ‘says that’. This fits in well with the not implausible intuition that quotes in mixed quotation behave much like an afterword, adding ‘and, by the way, he used these words in saying it’. This intuition might lead 3one to modify theories satisfying (A)–(B) to treat cases such as (4) along the lines of (7): (7) Clinton said that he’d cut taxes and he said it uttering, in part, the words ‘cut taxes’. First, what may be a minor point, (7) fails to account for the fact that mixed quotations contain a single component serving two functions concurrently: the quoted part is employed both to report what the speaker said and to say, at least partially, what he actually uttered. In (4), ‘cut taxes’ (together with other words) functions to report what Clinton said, namely, that he’d cut taxes; and it also functions to report that Clinton tokened ‘cut taxes’. So ‘cut taxes’ unambiguously serves two functions. In (7), no component serves these two functions; and so, it cannot explain how a single token can serve double duty.

SEMANTICS FOR QUOTATION 89

Second, suppose Bob tells Ted, ‘Carol is an oenophile’. He can be correctly mixed quoted by (8): (8) Bob said that Carol is an ‘oenophile’. Someone who utters (8) might be mixed quoting Bob’s utterance of ‘Carol is an oenophile’ because he is uncertain about what ‘oenophile’ means. He might assume that Bob’s vocabulary is larger than his own and mixed quotes him in order to indicate that ‘oenophile’ is an unknown word for him. Or he might be convinced that Bob is linguistically incompetent and wants to make this transparent to others without committing Bob’s mistake himself. Either way, mixed quotation is used to report what another said when part of what was uttered is unintelligible or just plain odd to the reporter. An account of mixed quotation along the lines of (7) does not adequately extend to (8), say, along the lines of (9): (9) Bob said that Carol is an oenophile and he said it using, in part, ‘oenophile’. In uttering the first conjunct of (9), normal English speakers fail to report anything at all. Since Ted can correctly report Bob with (8), the account adumbrated by (7) fails. None of this proves that leading accounts of the semantics of indirect quotation cannot be extended to account for mixed quotation. But once C1-C4 are adopted as adequacy constraints on a unified semantic theory of quotation, these accounts face a so far unmet challenge. Semantics of pure quotation The two influential accounts of pure quotation are the proper name account and the description account; both pre-empt C1-C4. According to the proper-name account (e.g. Quine 1961b, p. 40 and Tarski 1956a, p. 159), quotations are unstructured proper names of the quoted expressions and so there is no systematic correlation between what occurs inside quotes and the referent of the entire quotation. Applied to (4), the proper-name account issues in ungrammaticality. Since on this account quotations behave just like proper names, (4) semantically parallels (10): (*1 Clinton said that he’d Hillary. 0) According to the description account (e.g. Quine 1960, p. 202; Tarski 1956b, p. 160; Geach 1957, p. 79; 1970), there is a set of basic units in the language, either the words (Geach) or the letters (Quine). At this basic level, we retain the propername account either by treating the basic alphabetic units as names à la Quine—

90 HERMAN CAPPELEN AND ERNIE LEPORE

for example, ‘a’ is the name of one letter, ‘b’ the name of another, etc.—or by treating the morphemes as names à la Geach—‘red’, ‘un-’, etc. Quotations with more than one basic unit, are treated as descriptions of concatenations of whatever the basic units are. So, for example, (4) (à la Quine) is construed as (11): (*1 Clinton said that he’d ‘c’-‘u’-‘ t’-‘ ‘-‘t’-‘a’-‘x’-‘e’-‘s’. 1) where ‘-’ is the sign for concatenation. Since according to the description account, when quotes surround basic units, the result is a name of the expression, (11) is analogous to saying that Clinton said that he’d Hillary-Chelsea-A1- etc. That’s not anything like what Clinton said. So, on the basis of this brief discussion of these two accounts, we see that any account of quotation, according to which the semantic function of word-tokens inside quotation marks is to refer to word-types (or some other type of linguistic entity), fails to assign the correct truth-conditions to (4). In summary, our criticisms of accounts of indirect, direct and pure quotation show that in order to account for mixed cases an account must do two things: it must account for how the complement clause of, for example, (4) can be employed to effect simultaneously a report that Clinton uttered ‘cut taxes’ and one that Clinton said he’d cut taxes. We turn to proposals for how to execute these. Davidson on pure quotation Davidson’s paper ‘Quotation’ (1979b) is mostly about pure quotation. Davidson construes (3) as (12): (12 Lobster. The expression of which this is a token is a word. ) where an utterance of the second sentence is accompanied by a demonstration of an utterance of the first. On his view, quotes are definite descriptions containing demonstratives. The demonstrative picks out the token within the quotation marks, and the definite description denotes an expression, instantiated by the demonstrated token. Extending Davidson’s idea to direct quotation, (1) would be semantically construed as (13): (13 Clinton said (produced) a token of the expression instantiated by that. I’ll ) cut taxes,

SEMANTICS FOR QUOTATION 91

where an utterance of the first sentence is accompanied by a demonstration of an utterance of the second. Unifying Davidson’s demonstrative account of pure and direct quotation incorporates four attractive features: We can explain why learning to quote is learning a practice with endless application Understanding quotation is understanding ‘the expression instantiated by that’. There is no mystery about how we acquire this capacity, nor about how to account for it in a finitely axiomatic semantic theory. We can do this while preserving semantic innocence A semantic account T for a language L is semantically innocent just in case what an expression of L means according to T does not vary systematically according to context (see, Davidson 1968–1969/1984a, p. 106; 1975b/1984a, p. 166). Semantic innocence is preserved at two levels. First, the account does not assume words take on new semantic values when quoted. Second, it makes the device of quotation unambiguous; quotes in pure quotation are treated semantically in exactly the same way as quotes in direct quotation, thus respecting C4. Semantic innocence so construed, however, is compatible with there being contexts in which what an expression means is not in active use. So, even though ‘Clinton’ names Clinton, it is semantically inert in (14): (14 ‘Clinton’ has seven letters. ) We can explain why quotational contexts are opaque Sentences containing demonstratives need not preserve their truth-value when different objects are demonstrated. If you substitute a word-token of one type for another of a different type as the demonstrated object, different objects are demonstrated and thus the truth-value of the original (utterance of that) sentence may change (see conclusion below). We can, thus, explain why quantifying into quotes in natural language produces absurd results The expression ‘ y(“boxey” is a word)’ cannot be inferred from ‘ “boxer” is a word’ nor can ‘ x(“x” is a word)’. The account explains why these inferences fail: since it makes no sense to quantify into a demonstrated object, it makes no sense to quantify into quotes on this account.4

92 HERMAN CAPPELEN AND ERNIE LEPORE

Davidson’s account can be modified so as to treat quotes as quantifying over tokens that stand in a certain relation, call it the same-tokening relation, to the demonstrated token. This suggests construing (3) as (15): (15 x((STx, that)→Wx). Lobster, ) where an utterance of the first sentence demonstrates the exhibited token of ‘lobster’, ‘ST’ means same-tokens, and ‘W’ means is a word (token). Whether two entities stand in a same-tokening relation to each other is not settled by the semantics. This demonstrative account of pure quotation can be extended naturally to direct quotation. (1) should be construed as (16), or, alternatively, as (17): (16 u(Says(c, u) & y(ST(y, these)→ST(u, y))). I’ll cut taxes. ) (17 u(Says(c, u) & ST(u, these)). I’ll cut taxes.5 ) ‘Says’ means says. So, Clinton said a token that same-tokens the demonstrated object. Davidson on indirect quotation In ‘On Saying That’ (1968–9), Davidson paraphrases (2) as (18): (18 He’d cut taxes. Clinton said that, ) where ‘that’ is accompanied by a demonstration of the first utterance and that second utterance is true just in case Clinton said something that same-says the demonstrated utterance. Expression (18) is more perspicuously represented as (19): (19 u(Says(u, c) & SS(u, that)). He’d cut taxes, ) where an utterance of the first sentence demonstrates an utterance of the latter sentence, ‘Says’ still means says, and ‘SS’ means samesays. We do not intend here to engage in an evaluation of Davidson’s account of indirect quotation.6 Our aim, instead, is to show how the accounts of direct and indirect quotation can be exploited and developed so as (at least) to satisfy C1– C4. Notice straight away that C3 is satisfied, i.e. indirect quotation is treated

SEMANTICS FOR QUOTATION 93

differently from direct quotation. The former invokes same-tokening; whereas the latter invokes samesaying. Mixed quotation Earlier we argued that the available semantic accounts of pure (and direct) and indirect quotation do not integrate mixed quotation and therefore fail to satisfy C1–C4. The semantic theories of indirect speech canvassed earlier treat the complement clause as a semantic unit referring to (or in some other way determining) a proposition (or something proposition-like). The semantic theories of quotation canvassed earlier treat pure quotes as singular terms referring to abstract objects, expression types. For anyone receptive to either of these ideas, mixed cases remain enigmatic. From these points of view, mixed quoting seems to involve two entirely different activities taking place at the same time in the same place. How could the complement clause of (4) both determine a proposition not about words, and, concomitantly, refer to words Clinton used? The merged demonstrative accounts offer an ingenious reply. Since the complement clause is in effect semantically excised from (4) and merely demonstrated, we can ascribe different properties to it. With one utterance we can say both that the demonstrated token samesays one of Clinton’s utterances, and say that it (or parts of it) same-tokens that utterance. Our suggestion, then, is to construe (4) as (20): (20 u(Says(c, u) & SS(u, that) & ST(u, these)). He’d cut taxes, ) where an utterance of the first demonstrative demonstrates an entire utterance of ‘He’d cut taxes’ and an utterance of the second demonstrative demonstrates (only) the (sub)utterance of ‘cut taxes’.7 According to the unified account, mixed cases like (4) can be utilised both to attribute the same-tokening relationship between one of Clinton’s utterances and the demonstrated (sub)utterance, and to attribute a samesaying relationship between Clinton’s utterance and the demonstrated utterance.8 In criticising other accounts, we appealed to reports such as (8): (8) Bob said that Carol is an ‘oenophile’. Our account treats (8) as (21): (21 u(Says(b, u) & SS(u, that) & ST(u, these)). Carol is an oenophile, ) where ‘that’ is accompanied by a demonstration of the token of ‘Carol is an oenophile’ and ‘these’ by a demonstration of the token of ‘oenophile’. Is our

94 HERMAN CAPPELEN AND ERNIE LEPORE

account better positioned to account for the truth of (8) than the account we criticised earlier? After all, an utterance t of (8), construed as (21), is true only if Bob’s utterance samesays t’s sub-utterance of ‘Carol is an oenophile’. But can any such utterance by an English-speaker express anything if ‘oenophile’ were not an English expression? Anyone who has this concern has not understood how the extension of SS is determined. What samesays what is determined by the practice of indirect quotation; we place no a priori constraints on what can samesay what. Mixed reports containing meaningless or unknown words can be true.9 It is simply part of the data that an utterance of an ill-formed sentence of English can samesay one of Bob’s utterances. Maybe this is difficult to understand, but difficulty is no excuse for denial or dismissal.10 So, what we actually do inside the complements of indirect reports must be reflected in the extension of samesay relation. Indeed, it constitutes this relation. Conclusion Adequate semantic accounts of pure, direct, mixed, and indirect quotation cannot disrespect C1–C4. A myopically developed account of either indirect or pure quotation is likely to do this. Our joint account, as far as we know, is the only one available satisfying all four constraints. Also worth noticing is that this unified account satisfies C1–C4 with a simplicity and elegance not likely to be shared by any competing account. According to C4, quotes in pure, direct, and mixed quotation should be treated uniformly. It is hard to evaluate alternative accounts before they are supplemented with an account of direct and pure quotation. The joint modified Davidsonian account of pure and direct quotation is, for reasons discussed above, enormously attractive. If you share this view, it is certainly natural, though not necessary, to combine it with Davidson’s account of indirect quotation. What makes it almost irresistible to combine a demonstrative account of pure and direct quotation with a demonstrative account of indirect quotation is that this results in a unified account of opacity. Pure and indirect quotation are paradigms of opaque contexts. From a methodological point of view, it is both plausible and desirable that there be a common explanation of their opacity. This probably accounts for so many efforts in the history of this subject to assimilate indirect quotation to direct quotation (Carnap 1937, p. 248; Carnap 1947; Scheffler 1954; Quine 1956; 1960, sections 30–2; Sellars 1955; even Church 1954 given his metalinguistic solution to Mates’ problem). Though the unified demonstrative account provides a uniform account of quotation, it does not do so by assimilating either form of quotation to the other, and, therefore, does not fall prey, as did its predecessors, to the standard Church arguments (Church 1950, p. 97).

SEMANTICS FOR QUOTATION 95

Anyone who finds the account of quotation served in this paper unpalatable (perhaps because of the numerous objections to Davidson’s account of indirect quotation) needs an alternative. But the situation here is unlike indirect quotation. There are no large number of more or less acceptable competing accounts to choose from. In other words, if C1–C4 are acceptable constraints on a general account of reported speech, then any account of indirect quotation is incomplete until supplemented with an account of pure and direct quotation. Still, we’d like to end with a challenge for those philosophers unalterably convinced that Davidson’s account of indirect quotation is wrong. First, take your favourite theory of indirect quotation and show that it can (be extended to) account for C1–C4. Second, either develop a theory of pure quotation that combines with your favourite theory of indirect quotation to yield a unified account of opacity or explain why opacity doesn’t admit of a unified account. When you’ve completed these tasks, compare your results with the unified account with respect to simplicity and elegance. Notes 1 This paper is an abridged version of Cappelen and Lepore 1997a. 2 Proponents include Higginbotham 1986; Segal 1989; Larson and Ludlow 1993; and Larson and Segal 1995. 3 See Higginbotham 1986; Segal 1989; Larson and Ludlow 1993, pp. 335–42; Larson and Segal 1995, Chapter 11 for the pros and cons of this debate. 4 We are not claiming it’s illegitimate to introduce a quotation-like device into English that allows quantification. The point is rather that ‘ordinary’ quotation, which we are discussing, doesn’t allow such quantification. Here we agree with Quine 1961b and Davidson 1979b. 5 Given certain not implausible assumptions about ST, (16) and (17) are equivalent since: ST(u, these)↔ y(ST(y, these)→ST(u, y)) 6 See Higginbotham l986; Lepore and Loewer 1989b; 1990; Segal 1989. 7 Actually, there might be reasons for complicating this logical form by at least adding another quantifier ranging over a distinct (sub)utterance; and also, perhaps, by adding another predicate in logical form indicating that whatever utterance this second quantifier ranges over must be a part of whatever utterance the first quantifier ranges over. We will ignore these technical niceties here. 8 Notice that ‘cut taxes’ is mentioned in (20) if ‘mention’ is defined as the following: an expression e is mentioned in an utterance u just in case the token of e occurring in u is produced in order to be demonstrated so as to talk about tokens that sametoken it. 9 This, of course, doesn’t mean that these utterances are utterances of grammatical sentences; that is something the account behind (9) requires. On the demonstrative account we are demonstrating an utterance and not ourselves asserting it, but merely attributing a relational property to it. 10 See Cappelen and Lepore 1997b for our views about what determines the extension of ‘samesaying’.

96 HERMAN CAPPELEN AND ERNIE LEPORE

REPLY TO HERMAN CAPPELEN AND ERNIE LEPORE Donald Davidson This is an elegant, clear paper, and I find nothing in it with which to disagree. Anyone interested in a serious semantics for natural language should accept the four principles, C1–C4. But as Cappelen and Lepore argue, it’s hard to find an alternative to the theory they promote which satisfies the principles. There is an unimportant matter on which I would like to defend myself. Cappelen and Lepore object that if Clinton said, ‘I’ll cut taxes’, his utterance did not instantiate the pattern demonstrated by the written words ‘in any obvious sense’. I assumed it was obvious that there is a perfectly good sense in which we can say that the spoken and written tokens of the same expression have the same pattern or shape. But anyone who objects can fall back on the word ‘expression’. I can’t see how it matters. Nor is the issue nominalism; expressions, sentences, words, shapes are all abstract entities, and no one has a good idea how to do syntax or semantics without appeal to them. Probably we can get along with the relevant sets and let the intensional entities go, if it matters. But it is helpful to contrast the ideas of same-tokening and samesaying, not for the sake of nominalism, but for the sake of making clear why my accounts of direct and indirect quotation are not the same.

8 Samesaying Reinaldo Elugardo

In this chapter, I will focus on a proposed solution to a certain objection that some of Davidson’s critics have raised against his semantic programme.1 The objection is that the T-sentences of an empirically adequate truth theory for a natural language are incapable of bridging the gap between perceiving a speaker’s utterance and coming to understand what the speaker said with that utterance.2 Some of Davidson’s defenders have argued that his paratactic analysis can help bridge the gap because it uses the notion of samesaying. (Lepore and Loewer 1989a, pp. 65–78).3 I shall argue that the analysis cannot solve the gap problem for two reasons. First, if samesaying is a synonymy relation between utterances, as Davidson sometimes suggests, then his analysis will assign the wrong truth-conditions to indirect speech reports. Second, if samesaying is instead a pragmatically fixed relation, which it seems to be, then either his paratactic account is incorrect or the gap widens between what a truth theory T for a natural language L assigns as the content of a sentential utterance and what a speaker of L said by that utterance. The reason for the split is that the content reported by an indirect speech report is partially fixed by the pragmatic features of the reported utterance, whereas the T-theorems of a Tarskian truth theory for a natural language are not fixed at all by the pragmatics of language use.4 And so, the content correctly assigned to a sentential utterance by a correct truth theory need not be the content of what a speaker said by that utterance—the latter may depend on certain pragmatic features of the speaker’s utterance. Consequently, we still have the problem of explaining how a truth theory can be used to interpret what speakers of the language say by their utterances. This paper is divided into three parts. In the first part, I present Davidson’s paratactic analysis and two possible readings of ‘samesaying’. In the second part, I argue for my first claim stated above and answer some possible objections to my argument. In the third and final part, I argue for my second claim and reply to some other objections. Davidson’s paratactic analysis Consider the indirect quotation:

98 REINALDO ELUGARDO

(1) Alice said that life is difficult to understand. According to Davidson, an utterance of (1) has the logical form of two paratactically related sentential utterances, (2) Alice said that. Life is difficult to understand. Here an utterance of ‘that’ in the first sentence demonstrates an utterance of the second sentence (Davidson 1984a, pp. 105–6). Thus, according to his analysis, the complementiser of an indirect quotation functions as a demonstrative at the level of logical form.5 The second sentence in (2) occurs with its standard semantic properties in place, which are semantically inert because the sentence (or some utterance of it) is demonstratively displayed in an utterance of (2). Since the complementiser functions as a demonstrative, an utterance of the first sentence in (2) must be semantically evaluated relative to a context in which an utterance of the second sentence is demonstrated. In this case, an utterance of ‘Alice said that’ is true or false relative to the context in which ‘Life is difficult to understand’ is demonstratively uttered. According to Davidson, the first utterance is true relative to that context if and only if the following condition holds: (3) There is an utterance u such that Alice uttered u and u samesays that. [Life is difficult to understand.] where an utterance of the demonstrative demonstrates an utterance of the displayed sentence, ‘Life is difficult to understand’ (ibid.). The truth-conditions of an utterance of the second sentence in (2) are the standard ones: an utterance of ‘Life is difficult to understand’ is true just in case life is difficult to understand. Thus, the two sentential utterances that make up an utterance of (2) are semantically independent given that their truth-conditions are logically independent. Davidson’s account can explain some important features about indirect quotation. First, in contrast to inscriptional and quotational accounts of indirect discourse, it can explain why the complement sentence of an indirect quotation has semantic structure (Davidson 1984a, pp. 02–3).6 Second, it can explain why extensionally equivalent expressions cannot be mutually substituted, salva veritate, in indirect quotation.7 Third, in contrast to Fregean and other intensionalist accounts, it honours the principle of semantic innocence, according to which linguistic expressions do not shift in their semantic properties from direct discourse to indirect discourse or vice versa.8

SAMESAYING 99

Samesaying According to Davidson, a reported utterance must samesay an utterance of the complement of an indirect quotation if the report is to be true. What is it, then, for one sentential utterance to samesay another? He gives what could be regarded as two different answers to that question. According to the first, samesaying is the synonymy relation between two utterances, for example the reported utterance and the reporter’s own (demonstrated) utterance of the sentence complement of ‘said that’ (Davidson 1984a, pp. 104, 106–7). Some of Davidson’s commentators think that he is committed to this interpretation of samesaying. The reason is that, since indirect quotation is a species of indirect discourse, to quote indirectly what someone said, one must produce an utterance which, at the very least, is similar in meaning and in reference to the reported utterance, if not in psychological mode (Burge 1986, p. 192). Burge may conceivably have taken Davidson’s use of ‘synonymous’ as providing some indirect support for a literal semantic reading of samesaying.9 In any case, that understanding of the samesay relation is plausible given Davidson’s paratactic analysis. Davidson also gives a second, quite different, answer to our question about the notion of samesaying. In discussing the report, ‘Galileo said that the Earth moves’, he says that: The second utterance, the introduced act, may also be true or false, done in the mode of assertion or of play. But if it is announced, it must serve at least the purpose of conveying the content of what someone said. (Davidson 1984a, pp. 106–7, italics added). So, if an utterance of ‘the Earth moves’ is ‘announced’ by an utterance of ‘Galileo said that’, then the former utterance must serve to convey the content of what Galileo had reportedly said. This idea is sometimes put in terms of the notion of translation (Davidson 1984a, pp. 53, 176–7).10 There are important conceptual differences between synonymy and translation. Synonymous utterances must have the same referential truthconditions and, of course, must be the same in meaning. Thus, if samesaying is the relation of synonymy, then it follows that two truth-evaluable utterances samesay each other only if they have the same meaning, and thus have the same truth-conditions. By contrast, one utterance can be said to be an adequate translation of another without their being the same in meaning. For in translating another person’s speech, the reporter’s job is to ‘convey the content’ of what the speaker said. In some cases, that can be done by uttering a sentence that differs in its conventional lexical meaning from the sentence-token produced by the reported speaker. Here is a simple example: I can convey what you said by your utterance of ‘I am sleepy’ by pointing to you and saying ‘You are sleepy’. I conveyed what you said about yourself by using a sentence that differs in

100 REINALDO ELUGARDO

meaning from the one you used. Also, a reporter can sometimes give a true indirect quotation even though the sentence complement she tokens differs in its truth-conditions from those of the sentence that the reported speaker actually uttered. Thus, if samesaying is a translation or an interpretation relation, then two utterances can samesay each other even though they differ in their literal meanings and in their truth-conditions. With this distinction in mind, we can move on to the question of the connection between a truth-theoretic account of semantic meaning and a paratactic account of indirect discourse. The gap between truth and understanding As I mentioned at the beginning of the chapter, some philosophers think that the paratactic account can be used to answer a formidable objection to Davidson’s semantics. As a way of setting up the problem, consider the following example. Suppose that you hear Mary utter (4) as she runs frantically toward a fire exit: (4) There is a fire in the building! Upon hearing her utterance, and knowing that she uttered (4) and the context in which she uttered it, you infer and come to believe, (5) Mary said that there is a fire in the building. given that you understand English. If you understood Mary’s utterance, then your belief about (5) is quite justified.11 What kind of reason, then, might you have such that, if you had it, you would be justified in believing (5), assuming that you know that Mary uttered (4) and are aware of the context in which she uttered it? Davidson’s answer is that, provided that a certain constraint is met, it suffices for one to know the Tsentences entailed by a formally correct, empirically and materially adequate, Tarskian truth theory for English, in particular, the T-sentence of Mary’s utterance of (4) (Davidson 1984a, pp. 125–39). The constraint in question is that ‘the totality of T-sentences should optimally fit evidence about sentences held true by native speakers’ of English (p. 139). In addition, one would also have to know that the theory in question entails the T-sentence for (4), and that the theory met all the relevant empirical and formal criteria. If one knows all of that, in addition to certain facts about the context in which Mary uttered (4), then one can justifiably use the T-sentence of (4) to interpret Mary’s utterance as saying that the building is on fire. In which case, one would be justified in believing (5) upon hearing Mary utter (4). Lepore and Loewer reject Davidson’s answer (Lepore and Loewer 1989a, pp. 72–3). However they think that Davidson’s paratactic account can be used to supply the needed warrant (ibid., pp. 74–5, 79).

SAMESAYING 101

Lepore and Loewer’s proposal is that a correct truth theory, T, for a natural language L must meet the following condition if it is to be a correct interpretative theory of L: for every truth-evaluable sentence S of L, T assigns P as the truthcondition of an utterance of S if and only if a speaker of L will have said that P by that utterance of S. On the paratactic analysis of indirect quotation, a speaker said that P by an utterance of S just in case her utterance of S samesaid that: [P], where an utterance of ‘that’ demonstrates an utterance of the displayed sentence that replaces ‘P’ in the schema. Thus, on Lepore and Loewer’s proposal, the samesaying relation links the truth-conditions of an utterance, as determined by the correct truth theory for the language in which the utterance is made, with what a speaker of the language said by the utterance. Given their proposal and Davidson’s paratactic analysis, it follows that T correctly assigns P as the content of an utterance u of S in L if and only if any utterance that samesays u is a case of saying that P. Suppose that same-saying is the synonymy relation. In that case, T correctly assigns P as the content of u if and only if any utterance synonymous with u is a case of saying that P. I will argue in the next section that the right-hand side of this last biconditional is false, and thus, is not a necessary condition on the correctness of T as a truth theory for L. In other words, if samesaying is the synonymy relation, then Davidson’s paratactic account assigns the wrong truth-conditions to indirect speech reports. If that is correct, then the account will not bridge the gap between a truth theory for a language and indirect discourse. Let us assume that two utterances samesay one another if and only if they are synonymous.12 I will now try to show that Davidson’s paratactic analysis does not provide a necessary and sufficient condition for the truth of an indirect quotation on that reading of ‘samesay’. Example (1) Suppose that Alice never used the sentence, ‘Life is difficult to understand’, to state that life is difficult to understand. She never made such a claim. She instead once uttered: (6) Either life is difficult to understand or life stinks. Because she assertively uttered a disjunction, she did not use the words, ‘Life is difficult to understand’ to state anything about life in that context. Nor was that ever her intention. Thus, my report of what she said in that context, namely (1), is false.13 On Davidson’s analysis, (1) comes out true. After all, Alice did utter ‘Life is difficult to understand’ when she uttered the disjunction (6). Let u* be her subutterance of the first disjunct. When she uttered (6), she said something that is either true or false. Her disjunctive utterance has the particular meaning and truth-value it has, in part, because of the particular meanings and truth-values that

102 REINALDO ELUGARDO

each of the two component utterances has, one of which is u*. Hence, u* has a truth-condition in the context of Alice’s utterance of (6). In that same context, u* samesaid the sentential utterance that is demonstrated by an utterance of ‘that’ in Davidson’s analysis of (1), (3) There is an utterance u such that Alice uttered u and u samesaid that. [Life is difficult to understand.] Alice’s sub-utterance of ‘Life is difficult to understand’ has exactly the same meaning and truth-conditions as any literal utterance of that same sentence, which is displayed in (3). By hypothesis, having the same meaning suffices for samesaying. And so, u* meets the conditions expressed in (3), given the context in which Alice uttered (6) and the context in which I uttered (1). Thus, Davidson’s analysis of (1), namely (3), is true in this example. However, my indirect speech report, (1), is not true. Report (1) is true only if Alice made a claim to the effect that life is difficult to understand. But, by hypothesis, she never did. Davidson’s analysis does not provide, then, a sufficient condition for the truth of an indirect quotation. One might try to avoid the problem I raised by requiring that the reported utterance not be a sub-utterance of any structurally complex assertive utterance. But that will not help. Imagine that Alice asserted (7) rather than (6): (7) Life is difficult to understand and it stinks too. In that case, my report, ‘Alice said that life is difficult to understand’, is correct. If a speaker assertively utters a conjunction of the form ‘P and Q’, where ‘P’ and ‘Q’ represent any pair of truth-evaluable English sentences, then she will have also assertively uttered P and assertively uttered Q. Consequently, if Alice asserted (7), then she also asserted its first conjunct. In so doing, she said something that makes my indirect report true. But the proposal under consideration cannot explain why that should be the case. For Alice’s utterance of ‘Life is difficult to understand’ is a sub-utterance of her structurally complex utterance of (7), which the proposal rules out. Another possible move is to include an adverbial qualifier in Davidson’s analysis. Something along the following lines might work: ‘A said that S’ is true if and only if there is some utterance u such that A assertively uttered u and u samesaid that: [S]. This emended version of Davidson’s analysis avoids my counter-example. For although Alice’s sub-utterance of ‘Life is difficult to understand’ samesays a demonstrated utterance of the very same sentence, she did not assertively utter it. By hypothesis, no other assertive utterance that Alice ever produced samesaid a demonstrated utterance of ‘Life is difficult to understand’. So, given the revised Davidsonian analysis, the report, ‘Alice said that life is difficult to understand’, is false, which is what we want.

SAMESAYING 103

Objection (1) The paratactic analysis of ‘Alice said that life is difficult to understand’, namely (3), is false rather than true, contrary to what I have claimed. Alice did utter the disjunction (6), but, of course, her disjunctive utterance does not samesay an utterance of ‘Life is difficult to understand’. Nor in uttering (6) did she produce a different utterance that did. For, if one utterance, u, same-says another utterance, u*, then there is a proposition P such that u is an act of saying that P and u* is an act of saying that P. Alice’s sub-utterance of, (8) Life is difficult to understand. is not an act of saying something with semantic content precisely because it is a sub-utterance of her utterance of (6). Thus, given the plausible claim that utterances that samesay each other are each acts of sayings, it follows that Alice’s sub-utterance of (8) does not samesay a free-standing utterance of the same sentence. Report (3) is therefore false, assuming (as we have so stipulated) that Alice never uttered anything else that samesaid (i.e. was synonymous with) an utterance of ‘Life is difficult to understand’. Reply The objection fails because a sentential utterance that is a sub-utterance of another sentential utterance may also be an act of saying that P. Again, imagine that Alice had literally and assertively uttered the conjunction (7). Then, the report, (9) Alice said that life is difficult to understand and that life stinks. is true. But (9) holds only if the indirect quotation (1) also holds relative to the context in which she uttered the conjunction, since the ‘said that’ construction distributes over conjunction. Her sub-utterance of ‘Life is diffi-cult to understand’ in her utterance of the conjunction is, then, a case of her saying something with semantic content since (1) is true in the same context that (9) is. Therefore, the mere fact that a sentential utterance is itself a sub-utterance of some larger utterance does not automatically disqualify it from being an act of saying that P. Nor does being a sub-utterance prevent an utterance from samesaying another. Again, imagine that Alice utters the conjunction and I utter the sentence, (10 Life is hard to comprehend. )

104 REINALDO ELUGARDO

Her sub-utterance of the first conjunct in (7), namely ‘Life is difficult to understand’, samesays my utterance of (10). For the two utterances in question are synonymous, which, by hypothesis, is sufficient for them to samesay each other. But the most serious problem with Objection (1) is that its main premise is false. Contrary to Objection (1), two sentential utterances can samesay one another even though there is no proposition P such that both are acts of saying that P. Objection (2) The indirect quotation (1) in my example, is true rather than false. The reason is that an indirect quotation describes a locutionary act rather than the illocutionary act of stating. Specifically, indirect quotations report what J.L.Austin (1975) calls ‘the rhetic act’. Rhetic acts differ from two other kinds of locutionary acts (ibid., p. 95). Austin gives the following example to contrast rhetic acts from phatic acts (ibid.). The first sentence below describes a phatic act, while the second sentence describes a rhetic act: He said ‘The cat is on the mat’. He said (that) the cat is on the mat. Thus, an utterance of ‘He said that the cat was on the mat’ represents the reported speaker as possibly having referred to some cat or other and to some mat or other, and as saying that the cat in question is on the mat in question. By contrast, an utterance of ‘He said, “The cat is on the mat”’ does not represent the speaker as referring to anything or as meaning anything at all by an utterance; it instead reports the speaker as having produced an utterance, the words of which are displayed inside the quotation marks. In contrast to direct quotation, indirect quotation characterises the reported speaker’s utterance by displaying its semantic content. We can now state the second objection to my argument.14 In uttering (6), Alice also said something which means, relative to the context of utterance, life is difficult to understand. Admittedly, she did not state that. Still, Davidson’s paratactic account works because, in indirect quotation, ‘said’ has a rhetic meaning. And so, the indirect quotation, ‘Alice said that life is difficult to understand’ is true. Thus, my example is not a counter-example to the sufficiency of his analysis. Reply Even if ‘to say’ has a rhetic meaning, it does not follow that an utterance of (1) is true in my original example. Given our linguistic practice of indirect quotation,

SAMESAYING 105

(1) is false. If that is right, then the burden falls on those who think otherwise to provide a plausible reason for thinking that (1) is true. Admittedly, in Objection (2), my claim about the falsity of (1) was deflected by interpreting (1) to mean, (1a) Alice uttered something which means life is difficult to understand. which is true in my example. On Davidson’s view, (1a) is to be understood as (3), which is also true in my example. And so, (1) is true after all. However, stipulating a sense of ‘to say’ that implies that (1) is correct if and only if (3) is true is a questionable move. For, it assumes that our linguistic practice of indirect quotation would rule (1a) as the correct interpretation of what (1) expresses in my example, and thus, would justify the view that (1) is true after all. However, that assumption is precisely what is at issue, and so the objection begs the question on a very crucial point. Moreover, a speaker may utter a declarative sentence that has a definite truthcondition and yet no indirect quotation correctly reports what the speaker said. Suppose, for example, that Jones is both surprised and relieved to see his prodigal son, Jeremiah, return home after being gone for so many years. Upon seeing his son, Jones utters, ‘I’ll be damned!’ and thereby performs a rhetic act. There is no proposition P such that a sentence of the form, ‘Jones said that P’, correctly reports what he said. In particular, ‘Jones said that he will be damned’ is false. He said no such thing. Jones was expressing his joy or perhaps his surprise of his son’s return. He neither said that he will be damned, nor did he say that he was joyous or surprised. Thus, since Jones’s utterance is an expressive illocutionary act, there is no proposition P such that he asserted, stated or said that P by his utterance.15 Thus, some rhetic acts are not specifiable by any that-clause, namely, ones which are not also non-expressive illocutionary acts of any kind, as in Alice’s sub-utterance of ‘Life is difficult to understand’. An argument against the necessity of Davidson’s analysis The paratactic analysis fails for another reason: a speaker could say some-thing that makes an indirect quotation about him true even though nothing he uttered is synonymous with an utterance of the report’s complement. If that is correct, then Davidson’s analysis fails to provide a necessary condition for the truth of such a report. Imagine that Barry and Ernie are discussing the whereabouts of a mutual friend, Brian. Barry, who knows where Brian is, utters (11) to Ernie: (11 Brian went for a walk at noon. The park was beautiful. )

106 REINALDO ELUGARDO

In uttering (11), Barry intended to contribute something relevant to the conversation. For his part, Ernie assumed that Barry had intended him to take his contribution to be conversationally relevant. Let us suppose that they both mutually know these things about each other. Now if ‘the park’ in (11) is interpreted as referring to some park that Brian was not walking in, then Barry’s utterance of ‘The park was beautiful’ would be irrelevant to their conversation about Brian. In which case, on this interpretation of ‘the park’, Ernie’s assumption is false and Barry’s communicative intention is unfulfilled. Thus, if Ernie is to interpret Barry’s entire utterance of (11) as being about Brian, as Barry intended, then he has to assume that, in this context, ‘the park’ refers to the park where Brian was walking at noon that same day. Barry meant Ernie to interpret him in this way and believed that Ernie would be able to figure that out. Imagine, then, that Ernie utters (12) in reporting what Barry said: (12 Barry said that the park where Brian walked at noon was beautiful. ) Given that ‘the park’ has no anaphoric antecedent in (11), and given that Ernie issued (12), he must have ‘bridged’ the reference of ‘the park’ in the way Barry had intended him to do. Therefore, given these background contextual assumptions and Gricean considerations, (12) is a true report of what he said. One would be hard pressed to deny that (12) is a true report of what Barry said. For, in uttering (11), Barry made, in passing, a claim about Brian’s whereabouts, namely that he was walking in the park, and about the park’s aesthetic quality, namely that it was beautiful. There is no utterance, however, that Barry produced in uttering (11) that samesays an utterance of (13) (in the sense of ‘synonymy’); nor, we may suppose, has he ever uttered something in some other context that samesaid an utterance of (13): (13 The park where Brian walked at noon was beautiful. ) For, neither Barry’s utterance of (11a) or (11b), nor his utterance of their conjunction, is semantically equivalent to (13), since neither one, taken singly or jointly, entails that Brian walked in a park: (11a) Brian went for a walk at noon. (11b) The park was beautiful. Nonetheless, for the reasons presented above, (12) is a correct report of what Barry said. Consequently, the paratactic analysis fails to give a necessary condition for the truth of an indirect quotation, if samesaying is synonymy.

SAMESAYING 107

I have not argued that samesaying is not a semantic relation. Nor have I shown that Davidson’s paratactic analysis is incorrect. My arguments show at most that his analysis is problematic if samesaying is defined in terms of sameness of meaning between pairs of utterances. Fortunately, one can still accept Davidson’s analysis and hold that samesaying is a pragmatic relation, involving the employment of some interpretative scheme, rather than a semantic relation. Lepore and Loewer adopt this view, which I will now present.16 Samesaying as a pragmatic relation According to Lepore and Loewer, our linguistic practices of indirect quotation fix the extension of the samesaying relation (Lepore and Loewer 1989a, p. 76). Their proposal is attractive for several reasons. First, it saves the paratactic analysis from my two counter-examples. The examples I presented depend on the dubious assumption that sameness of meaning is both necessary and sufficient for samesaying. This assumption is denied on the pragmatic view of samesaying. Second, and more importantly, not even Davidson holds that assumption, although others do. For example, in a reply to an objection that John Foster raised against his paratactic analysis, Davidson says this: But what is this relation between utterances, of stating the same fact or proposition, that Foster has in mind?…if both reference and meaning must be preserved, it is easy to see that very few pairs of utterances can state the same fact provided the utterances contain indexical expressions…. I cannot twice state the same fact by saying ‘I’m warm’ twice. (Davidson l984a, p. 177) In Davidson’s example, the same speaker (Davidson) produces two synonymous indexical utterances that do not samesay each other if samesaying is a matter of different utterances stating the same fact. Thus, synonymy is not sufficient for samesaying in that sense. The idea that samesaying is a pragmatically fixed relation is therefore wellmotivated. However, the example I give below shows that Davidson’s analysis is either not entirely correct or it cannot be used to convert a truth-theoretic approach to linguistic meaning into a theory of linguistic understanding. Example (2) When he spoke to the multitude at Julius Caeser’s funeral, Cassius mockingly said of Brutus: (14 Brutus is an honourable man. )

108 REINALDO ELUGARDO

Nero, who was in the audience, was unaware that Cassius was being ironic. Later that evening, he reports to a friend what Cassius had said: (15 Cassius said that Brutus is an honourable man. ) If Nero’s report is interpreted as a report about what Cassias literally said by his utterance of (14), then what Nero reported is true in spite of the fact that he failed to grasp Cassius’ intended meaning. For, suppose that (15) is false. Then so is (16) since (16) entails (15): (16 Cassius said [by his words] that Brutus is an honourable man but he meant ) [by the same words] just the opposite, namely, that Brutus is a dishonourable man. But if (16) is incorrect, then we cannot explain why Cassius’ utterance of (14) is ironic, and thus non-literal, even though the words he used occur with their literal meanings. Indeed, if (16) is false, then Cassius’ utterance of (14) is not ironic. But it is since he mocked Brutus. Therefore, (15) is true. However, no utterance that Cassius ever uttered samesaid Nero’s utterance of ‘Brutus is an honourable man’. After all, Cassius never thought that Brutus was virtuous and was loath to ever say anything that would suggest that he thought otherwise. When he uttered (14), he thought that his audience would detect his sarcasm. But Nero took his remark to be a compliment. Pragmatically speaking, then, Cassius’ utterance of ‘Brutus is an honourable man’ does not samesay Nero’s own literal utterance of the same sentence. Thus, Davidson’s paratactic analysis of Nero’s utterance of (15) is false. But Nero’s report is still true. Hence, the analysis fails to give a necessary condition for the truth of (15), even if samesaying is a pragmatic relation. There is, however, some reason to think that this example fails, which I will now address. Objection (1) Cassius’ remark is better reported by (17) rather than by ‘Cassius said that Brutus is an honourable man’, i.e. by (15). (17 Cassius said that Brutus is a dishonourable man. ) It is wrong to represent Cassius as actually claiming that Brutus is honourable— indeed, it would be a gross misrepresentation. To be sure, one may directly quote him as having said the words, ‘Brutus is an honourable man’. But he was, after all, being sarcastic when he uttered them. That pragmatic fact alone rules out (15) and rules in (17) as the correct report of what Cassius said. Besides, if both

SAMESAYING 109

(15) and (17) were true, then Cassius would be guilty of having said something inconsistent by his utterance. However, not only would this be a harsh accusation but it would also be incorrect. Therefore, on grounds of charity, we ought to conclude that (15) is not true. Reply The phrase, ‘what Cassius said’, can be interpreted in several different ways.17 It can mean the pattern of the words or the sentence that Cassius uttered. It can be taken to mean the literal semantic meaning of the sentence he uttered. Or, it can be construed as referring to the content or proposition that Cassius intended to communicate to his audience, or the proposition that his audience took him to be communicating to them. Perhaps what we ought to say, then, is that (17) rather than (15) is an accurate report of what Cassius intended to convey by his utterance, whereas (15) rather than (17) is a more accurate report of what he literally said, regardless of what he intended to communicate. Both (17) and (15) are jointly compatible. We can then avoid the charge of attributing a blatant inconsistency to Cassius by distinguishing what he intended to communicate by his ironic utterance from what he literally stated by his words.18 As suggested by (16), Cassius said one thing by his words but clearly meant something else instead by them. Objection (2) In spite of this reply and my main argument for the truth of (15), the only way that (15) could be true is if it is read as an ill-formed direct quotation of Cassius’ words. In which case, Davidson’s analysis does not apply and my example becomes irrelevant. More importantly, it does seem that (17), rather than (15), is a true indirect report of what Cassius said. But, then, Davidson’s paratactic analysis of (17) comes out true. For what Cassius said by his sarcastic utterance of ‘Brutus is an honourable man’ is just what would be conveyed to someone who correctly understood what Cassius said by a literal demonstrative utterance of ‘Brutus is a dishonourable man’, assuming that the same referential semantic properties are held constant. If this is how my example should properly be understood, then it does not establish that Davidson’s analysis fails to give a necessary condition for the truth of an indirect speech report. Reply I concede this last point. However, Objection (2) now leads to the general problem for Davidson’s semantic programme that I mentioned earlier. The aim of the programme is to produce a finitely axiomatisable truth theory, knowledge of which would suffice for one’s understanding of the language in question.