THEORETICAL LINGUISTICS Vol.1
1974
W DE WALTER DE GRUYTER
G
BERLIN-NEWYORK
INSTITUT DEUTSCHS PHILOLOGIE U$VE:£r;£7 KOCHEN Inventar- 4·. r.
Archiv-Nr. 3109001089 ISSN 0301-4428 ©1974 by Verlag Walter de Gruyter & Co., vormals G. J. Göschen'sche Verlagsbuchhandlung—J. Guttentag, Verlagsbuchhandlung—Georg Reimer—Karl J. Trübner—Veit & Comp., l Berlin 30, Genthiner Straße 13.—Printed in GermanyAll rights reserved, including those of translations into foreign languages. No part of this journal may be reproduced in any form—by photoprint, microfilm or any other means—nor transmitted nor translated into a machine language without permission from the publisher. Typesetting: H. Hagedorn, Berlin—Printing: Mercedes-Druck, Berlin—Binding: T. Fuhrmann, Berlin.
CONTENTS
Artifles BELLERT, IRENA On inferences and interpretation of natural language sentences
215
GABBAY, Dov M., and MORAVCSIK, J. M. E. Branching quantifiers, English, and Montague-grammar
139
HOEPELMAN, JAN PH.
Tense logic and the* semantics of the Russian aspects
158
ISARD, STEPHAN D. What would you have done if... ?
233
KARTTUNEN, LAURI Presupposition and linguistic context
182
KASHER, ASA Mood implicatures: A logical way of doing generative pragmatics
6
KUTSCHERA, FRANZ VON Indicative conditionals
257
LIEB, HANS-HEINRICH Grammars as theories: The case for axiomatic grammar (part I)*
39
SOAMES, SCOTT Rule orderings, obligatory transformations and derivational constraints
116
Discussions and Expositions
BELLERT, IRENA A reply to H.H. Lieb
v
287
DASCAL, MARCELO and MARGALIT, A vis HAI A new 'revolution* in linguistics? 'Text-grammars' vs. 'sentence-grammars'
195
LEWIS, HARRY A. Model theory and semantics
271
Part II of Lieb's article is to appear in volume 2 No. 3 of THEORETICAL LINGUISTICS
EDITORIAL
Theoretical Linguistics is concerned with linguistic theories. This statement seems to be almost a tautology, but it is not: though that may not be immediately obvious. The formulation needs interpretation, especially if it is to be taken as a condensed statement of the editorial policy of this journal. Let me begin by stating the general tendencies determing this policy: rigour of presentation, adequacy with respect to the more complicated aspects of natural languages, and critique of the foundations of the analysis of languages. The first tendency stems from logic, the second from linguistics, and the third from philosophy of language, the three "parents" of theoretical linguistics. But let me explain this in detail. "Theoretical Linguistics" is not only the title of this journal — it is almost meant to designate a sub-dicipline of linguistics. Theoretical linguistics, in being concerned with the development of theories about general aspects of particular languages or of language and its uses in general, as well as with the discussion and analysis of the form, scope and applicability of such theories, will have to be contrasted with that part of linguistics which is either concerned with experiments, tests and field-work applied to languages or else with setting up particular empirical hypotheses about the structures of languages based on judgements about (introspective or observable) experiences. Obviously, there should be a very close interrelation between the different parts of linguistics including theoretical linguistics. In view of this interrelation, linguistic theories will have to be extensively illustrated by derivations of hypotheses about particular phenomena of languages in order to become accepted as systems of hypotheses and these special hypotheses will have to be tested to evaluate the empirical adequacy of the linguistic theories from which they are derived. Many expositions and illustrations of linguistic theories describing fragments or aspects of particular languages, dialects etc. will therefore belong to theoretical linguistics as well as empirical research trying to assess the justification of some theory in general. But the central concern of theoretical linguistics should be the construction and discussion of general theories of a high degree of methodological andy in particular, meta-linguistic sophistication.
2
Editorial
Wherever possible, the concept of an axiomatized, though not necessarily formalized, theory should be taken as a standard at which theories presented or discussed in this journal aim. Glossing over a great number of problems, I assume that (1) linguistic theories (or the central empirical propositions correlated with them), are sets of statements, some of which are empirically true or false; and that (2) the logical relations among the statements of a linguistic theory may be exhibited by an axiomatic system. The logical relations may be deductive or inductive. We may require, in particular, that (3) the logical relations of a linguistic theory should be presented either (a) as an axiomatic deductive theory with descriptive constants (which may be expressed, in its strictest form, in some formal language—the metalanguage of the language to be described); or (b) by definition of a set-theoretical predicate (or a set-theoretical structure). In empirical disciplines aiming at all at strict axiomatization method (3b) is often applied, especially because it is then possible to draw upon the huge stock of standard mathematical notions (under the assumption that their definition as set-theoretical predicates has already been accomplished in meta-mathematics). Indeed, an axiomatization along these lines may have been an implicit aim of mathematical linguistics, to the extent that the latter developed in relation to generative transformational grammars. The basic intention was apparently to define set-theoretical predicates presenting the structural properties of descriptive constants like " is a grammatical sentence in L" and ".. .is a syntactic structure of sentence _in L", defined for each natural language L. This approach was based on some fundamental notions such as "... is an elementary unit of expression" (e.g. phonetic unit) and the notion of concatenation, supplemented by definitions of auxiliary notions such as "grammatical construction", "grammatical (or syntactical) rule" etc. The basic notions could be taken to correspond to the set-theoretically defined mathematical notions "element of the generating set" and "binary operation" of a free semi-group with a generator set. In connection with these definitions particular emphasis has been placed upon the requirement that the definitions should be recursive. Since, however, the sets corresponding to the notions to be defined could not be taken as finite, methods of recursively constructing the elements had to be applied. This led to speaking of devices that generate sentences, from which some people have mistakenly supposed that the particular method of recursive definition applied might serve as a model of psychological procedures performed by speakers, and possibly also by hearers, of the language for which grammaticality of sentences and their structure had been defined. In fact, only the recursive definition of these notions
Editorial
3
and no psychological procedures were in question. (It could at most be said that these notions, if adequately defined, put constraints on possible descriptions of psychological procedures). Recently, a complex of set-theoretical predicates (or set-theoretical structures) for syntax and semantics of natural languages has been advanced by applying the methods of model-theory and intensional logic. In complexity and, perhaps, in scope, the resulting systems seem to be at least comparable to generative transformational systems. But in addition to syntax, these proposals included as an essential component an axiomatized theory of indirect semantic interpretation by defining a translation function into a language with a syntactic structure that is better suited to the definition of a direct model-theoretic interpretation than are natural languages. Recently, method (3a) has been used for axiomatization within the generative theory of grammar and the first steps have been taken towards axiomatizing grammars. Further developments in the field of axiomatic grammars will be presented in the first issues of this journal. There may be alternative ways of presenting precise linguistic theories, but they will have to be comparable in rigour to the standards of formal theories that are already well understood and analysed. Certain judgements of value may seem to be implicit in my discussion of various approaches to theoretical linguistics. Let me therefore state explicitly that these judgements are by needs subjective. I do not believe, though, that this should exclude such judgements from an editorial. Any bias that might result from my own convictions will, as a matter of editorial policy, be corrected by consultation of the members of the editorial board, who will not necessarily agree in all their views on theoretical linguistics. Admitting all this, I would still suggest that certain consequences follow concerning the style of presentation preferred in this journal: Currently established mathematical and logical conventions of formulation should be followed as far as possible. Some readers may, perhaps, object that too much emphasis has been placed upon the form in which theories are presented. It is certainly true, that the force of a theory does not derive from its rigour but from its leading ideas and from the answers that can be given to the following questions: Are these ideas revealing and fruitful? Do they offer important generalizations and insights? Are they adequate and applicable? Can they be related to neighbouring fields or do they even sustain notions that are basic in a number of neighbouring fields, etc., etc.? True enough! But it is also true that the reliability and the continuity of development of ideas does depend on the perspicuity, rigour and systematic controllability of the theories in which they are presented. It may be granted that fruitful ideas are very often not stated in a rigorous form in the
4
Editorial
first instance; a number of attempts at finding the best formulation will have to be allowed. However, the very attempt to integrate an idea into a theory will demand precise and rigorous formulation. Nevertheless, rigour of meta-linguistic presentation is only a virtue in an empirical discipline if it is not empty, i.e. providing that the proposed theory aims at empirical adequacy. Progress in logical and theoretical analyses of languages during recent years should allow for the development of theories that combine rigour with empirical adequacy. The situation has changed from that of the early days of logical analysis of languages when rigour could only be achieved at the sacrifice of adequacy in the description of natural languages. Still, whereas, because of their different training, the logician has had the advantage over the linguist with regard to meta-linguistic rigour, the position is reversed when it comes to meeting the requirements of empirical adequacy. Logicians tend to underestimate the complexities of natural languages and to ignore whatever does not fit into their framework of description. Finally, we have to acknowledge that strict adherence to the basic principles of a discipline may, at times, be detrimental to its fruitful development. It is the task of the philosopher to insist in season and out of season on the need for reflection on the basic principles and to doubt the validity of tacit assumptions. And of course, this also applies to the conception of theoretical linguistics as presented in this editorial. In particular, such points as the relation of theoretical linguistics to formalization, which are somewhat controversial, even among the members of the editorial board, should be discussed in the journal itself. It is to be hoped that logicians, linguists and philosophers of language— and at times perhaps others such as the computer scientist—will collaborate in the development of theoretical linguistics. It should be borne in mind, however, that contributions are variously assessed in different disciplines: our field will not be an exception. In order to aid mutual understanding, authors should, therefore, provide their contributions with careful introductions which present the problems succinctly, as well as the criteria for a satisfactory solution, and the methods applied in trying to attain it. This effort on the part of the authors should be met by the reader with patience, tolerance und willingness to learn even from approaches that seem to be unfamiliar at a first reading. In order to further mutual understanding as well as more thorough acquaintance with the different approaches of theoretical linguistics, the journal will have a special section called "Discussions and Expositions" containing detailed critical analyses as well as contributions designed to give an easily readable introduction and exposition of more complicated ideas and theories which are of interest. At present for example, a number of aspects of modeltheoretic grammars seem to need such an introduction. Authors who are willing and able to provide expository contributions of this type are invited to submit them to the journal. It is up to the authors and readers of this journal to contribute to the development of theoretical linguistics, the authors by providing
Editorial
5
not only original contributions of high standard, but also pedagogically perspicuous and lucid introductions to or critical discussions of the different ways of approaching a common field of interest, the readers by analyzing the results published with an appropriate combination of criticism and tolerance and by applying them fruitfully. H. Schnelle
ASA KASHER
MOOD IMPLICATURES: A LOGICAL WAY OF DOING GENERATIVE PRAGMATICS*
In this paper we present an extension of the model-theoretic framework of semantics in which some pragmatical aspects of natural language can be treated adequately. In ch.l we specify the scope of generative pragmatics. Ch. 2 outlines the formal framework of semantics. In ch. 3 we pose the problem of non-indicative sentences and in chs. 4 and 5 we reject the solutions suggested by Stenius, Aquist and Lewis. In ch. 6 we define some pragmatical concepts—preconditions and implicatures of various types—using them in ch. 7 to present a pragmatical characterization of moods in terms of preference-implicatures. Some ramifications are discussed. In ch. 8 we draw a distinction between basic preference-implicatures ("pragmemes") and derived ones. The derivations involve communication rules. In ch. 9 we outline the extended formal framework. Finally, in ch. 10, we present some open questions.
It may be that the next great advance in the study of language will require the forging of new intellectual tools that permit us to bring into consideration a variety of questions that have been cast into the waste-bin of 'pragmatics'. N. Chomsky Form and Meaning in Natural Language 1.
Introductory
Ideal speakers live in speech-communities, communicating with their fellows regularly and happily. They ask questions and answer questions, as all of us, real speakers, do; some of them issue commands, as some of us do; and a few of them give * The preparation of this paper was partly supported by the Deutsche Forschungsgemeinschaft (Bonn—Bad Crodesberg). I am grateful to John Bacon, L. Jonathan Cohen and Helmut Schnelle for pointing out some important, apparent counterexamples. I owe special thanks to Helmut Schnelle for extensive comments on an earlier version of the paper. As usual, I am grateful also to Yehoshua Bar-Hillel for some comments and remarks.
Mood implicatures: a logical way of doing generative pragmatics
7
verdicts, as a few of us do. However, unlike any of us, ideal speakers never fail to perform happily speech acts they intend to carry out. They utter an instance of sentence S in context C only if S is linguistically appropriate to 6"; in their discourse, preconditions never fail to obtain and presuppositions are always true. Pragmatics is the study of that part of linguistic knowledge and behavior which pertains to speaker-sentence-context relations. In a homogeneous community of ideal speakers the linguistic behavior matches strictly the linguistic knowledge, but this is not the case when real speakers are under consideration. In that case there is the distinction between Pragmatics in a narrow sense—the study of the pragmatical competence of real speakers, which is, so to speak, the study of the pragmatical aspects of the behavior and knowledge of ideal speakers—, and Pragmatics in a broad sense— the study of the pragmatical aspects of the linguistic behavior of real speakers. This paper is confined to Pragmatics in the narrower sense. A pragmatical theory purports to describe and explain adequately part of linguistic competence. When such a theory incorporates a system of explicitly presented postulates and rules that govern the pragmatically basic relations, it will be called a generative pragmatical theory. Pragmatical investigations are intertwined with semantical considerations. We side with Montague in the dispute about representation of meanings: logical languages and model-theoretical systems of interpretation are the backbone of adequate semantical theories of natural languages.1 On the other hand, we take sides with all those linguists who have argued to the effect that at least some transformations have no semantical bearings, and hence, pace Montague, not every linguistic rule (of a syntactical nature) should be semantically justified. Dubbing this approach "logical generativism", we may call the present work a logical way of doing generative pragmatics. In this paper we attempt to provide a framework in which the performative functions of sentences can be described and explained. Some aspects of J.L. Austin's views of speech-acts have been systematized in.J.'R. Ross* work on declarative sentences and J.F. Sadock's works on hyper-structures, but they are restricted to syntactical considerations.2 We are interested in the pragmatical point of view.
2.
Semantical Preliminaries 3
Semantics is a hydra-headed label, but we are interested here in just two heads. On the one hand it means the science of meaning and on the other hand it refers to "a discipline which^ speaking loosely, deals with certain relations between expressions of a language and the object s ... 'referred to* by those expressions ..., or possibly ... [the] 'states 1
Richard Montague's papers will be published in a collection by Yale University Press. Meanwhile, consult Montague (1970) and Montague (1973). 2 Ross (1970) and Sadock (1969). 3 The reader who is fluent in some version of model-theoretical semantics may skip this chapter. See, however, the fourth conclusion of this chapter.
8
Asa Kasher
of affairs' described by them." (Tarski 1944, 345). Since the cleavage between theories of meaning and theories of reference is endorsed both by philosophers and linguists4, let me disclose my sticker at the very outset: Unionist. A general theory of reference provides foundations of and a framework for an adequate theory of meaning for natural language. Replacing slogans by examples, consider first the phrase (1) the former Pope. (2) A*/("the former Pope", Q =/j (/c, Pc) This string of words does not refer on its own merits to anybody on earth, even when an appropriate syntactical structure is attached to it. Otherwise it could not be used to refer to different people on different occasions. Consider the following schematic description of the recent papacy: Figure 1
An utterance of (1) at /j was used to refer to Pius XI while an utterance of it at /2 was used to refer to Pius XII. Generally, where Cis the context of utterance of (1), tc is the time ofthat utterance, and P c is a succession of popes, say, from St. Peter and on, future popes not excluded. Where the actual state of the papacy is not clear we can stick, say, to the official view of the Catholic Church or to an alternative view. The function/i transfers us, so to speak, from a given interval in figure 1 to the former interval, if any. The second argument of yj (x, y) specifies the succession under consideration, the first argument specifies which interval in this succession serves as a starting point for the operation of/j, which in turn moves us from ÷ to the former interval ofj, if any. Notice that ji(/c, Pc) is sometimes undefined. Assuming that the consecration of the second pope, St. Linus, took place in 67, ji is undefined for every earlier point. Presuming that antipopes were popes, f i is undefined whenever there are two or more former popes; thus^(1122, Pc) is undefined, because both Gelasius II and Gregory VIII were then former popes. The function^ is defined at / if and only if any pope at / is an immediate successor of exactly one pope, in Pc. We can now see what the contribution of each element of (1) is to the reference made by any utterance of (1). The word 'pope' directs us to Pc, the word "former" brings forward /, and then fir and finally 'the' introduces restrictions on the domain off é, mentioned in the former paragraph. 4
Quine (1961) 21f, 47ff and 130ff. Cf. Katz (1966).
Mood implicatures: a logical way of doing generative pragmatics
9
What would be the reference made by using (1) at /^, had John XXIII been the successor of Pius XI? Obviously, the reference made would be to Pius XI rather than to Pius XII. The word "Pope" in (1) would then direct us not to Pc but to Pc, which is a possible history of the papacy, referred to in a possible context of Figure 2
utterance C at which (1) was used. One can imagine other contexts of utterance in which (1) is used to refer to popes in other possible courses of events, including different possible papal successions. However, (2) holds for each of these possible contexts of utterance, where "Pc" stands for what is the actual papal succession/row the point of view of C rather than from our point of view, or from the present Pope's point of view. Summarizing what we have found so far, we maintain that when (1) is used in a context Cto refer, the reference is a function of some aspects of the context C; one is the time-index /c, involved in the use of "former", another is the exponent (as opposes to index) eC, which is the possible world or possible history under consideration, that is involved in the use of "Pope". We take the meaning of the phrase "the former Pope" to be fully captured by a satisfactory characterization of the function Ref("the former pope", C). One way of characterizing this function is sketched in (2), which needs however some further clarification. The reference of "the former Pope" in C is a function of Pc, but how exactly should Pc be characterized? It is not the reference of "the pope" in C, because it specifies a whole succession of popes and not just the one that heads the Roman Catholic Church in C, according to some view. On the other hand, given some means GC for determining for a given time point / and a person e whether e is a pope at / or not, we can define the appropriate papal succession Pc in terms of Gc. Let us then amend (2) to read (3) A*/("the former Pope", Q = /j (/c, GC}. Again, one can imagine different C7s that determine different Fs; this is why we have GC rather than just G. Every Gc is related to somepossib/e history; the answer to the question whether e is a pope at / is relative to this possible history. Trying to get rid of this relativity we take G to be the means for determining for a given possible history b, a time point / and a person e, whether e is a pope at / in h. Again, a satisfactory characterization of G will be here considered to capture in full the meaning of the word "pope". GC is then the function G (bc, *> *)> where hc is the possible history under consideration at Cand t and e are appropriate variables. Using the ë-notation for specification of variables, we have:
10
(4)
Asa Kasher
?/("the former Pope", Q =fi (/c, ë/ ë* 6),
and, consequently, the meaning of (1), which is KC Ref ("the former pope", 6), is represented as a function of the meanings of "the", "former" and "Pope" in (5): (5) Ç = ë€ Ufa ë/ë* Gfc, /, 0). from which it is also clear that the context C in which (1) is used to refer, contributes significantly to the reference, both by index and exponent. Notice that we take a satisfactory characterization of G to be first of all a sufficient condition for capturing in full the meaning of the word "Pope". Whether it is a necessary condition or not is open to debate. We cannot do here more than deckring that for some expressions it seems to us to be a necessary condition as well. "Pope" is among these expressions. However, we admit that in many cases this could not be a necessary condition. For an extreme anti-necessitarian view, consult Schnelle (1973). Having seen how the meaning and reference of the phrase "the former pope" are rekted to semantical properties of its elements, we turn now to see how this phrase plays the role of an element in bigger phrases. (6) The former Pope was Italian. Denoting by "/" the meaning of "Italian", which is a function I(hc, /, *), similar to the function G we saw earlier and ignoring the tense of (6), it is clear that the meaning of (6) is a function of H (the meaning of "the former pope") and 7. Since both H and / have contextual arguments, vi%. bc and /c, it should come as no surprise that the meaning of (6) turns out to have contextual arguments as well. Why (6) depends on the context of utterance should be clear by now. On one occasion, given appropriate index and exponent, the proposition made is true—John XXIII was in the actual history, say, according to the official view of the Catholic Church, an Italian; on other occasions the produced proposition is false—John XXII was, according to the same official history, a Frenchman, and a possible history can be outlined in which John XXIII was not an Italian. The meaning of (6) will then be taken to be fully captured by an adequate characterization of the function that relates a context of utterance—its indexes and exponent—to the truth-value of the proposition made by using (6) at that context. An adequate characterization of the meaning of (6) should show how the meaning of (6) depends on the semantic properties of its elements. I(kc> A 0 is a function from exponents, time-points and persons to truthvalues. Given any context of utterance C9 we determine by / whether the reference of (1) in that context was an Italian, during the appropriate period of the possible history which is the exponent of C. In other words, we apply the function 7(x, y, z) to a certain triple of arguments, as follows: for ÷ we substitute 'be denoting the exponent of C which is the possible history under consideration, and for æ we substitute '//(C)', which refers to the same person, if at all, that (1) refers to in C. What we substitute for y depends on an analysis of the tense of (6) which is beyond the scope of our discussion.
Mood implicatures: a logical way of doing generative pragmatics
11
Altogether we have:
(7)
I(bc,t'c,H(C))
which is either true or false, or undefined in case the last argument fails to refer to anything. The following partial function (8) ë€É from contexts of utterance to t ruth-values represents the meaning of (6). The purpose of this section has been to hint at our semantic framework. 1) A semantic theory of a language is a theory of functions that are related to phrases of the language and their semantical properties; given a phrase x, a word or a sentence and an appropriate context of utterance C, the function attached by the theory to ÷ supplies the value of C, if any. Values of some nounphrases are individuals referred to by using ÷ in an appropriate C. 2) Values of other phrases, e.g., predicates, adjectives and prepositions, are functions that contribute to the functions attached to other phrases in which the former ones appear. Generally, a linguistic adequacy criterion for a semantic theory is that the function attached by the theory to a phrase ÷ be determined by the functions attached to the elements of ÷ and by its syntactic structure. 3) Functions may not be defined for some arguments; for example, a noun phrase ÷ may fail to refer whenever uttered in contexts in which presuppositions induced by ÷ do not obtain. 4) The contribution of any Cis two-fold: first it provides values of indexes: speaker, addressees, time and place of utterances, and so forth. Secondly, it fixes an exponent, including a possible factual background for x's functioning. An exponent may be a possible state of affairs, a possible succession of such states, or generally, any class of such states. We use the term "exponent" in order to be neutral with respect to the extent and structure of those classes. Justification and detailed development of such frameworks are not our job here and we refer the reader to works by Montague, Lewis and others.5 In what follows some problems will be considered within such a framework.
3.
The problem of other moods
We assume that the meaning of the sentence (9) He is still alive is captured satisfactorily by an adequate presentation of some partial function from contexts of utterance to truth-values. Such a presentation seems possible, 5
Lewis (1972) is an excellent introduction to Montague (1973). Consult also Thomason (1972). Notice that we have not subscribed here to any particular logical language or theory, but to the general conception of model-theoretical semantics for natural languages.
12
AsaKasher
and even natural, because whenever (9) is uttered in an appropriate context, what is said may be true or false. A context of utterance is inappropriate for (9) if, for example, it does not specify the reference of "he" in (9). Thus, lack of information may cause the function attached to (9) to be undefined for some contexts. Consider now the following sentences: (10) May he be still alive! (10') Is he still alive? (10") Who is still alive? Trying to attach to each of these sentences a function, depicting its meaning along similar lines, we have no difficulty in circumscribing the domains of these functions—the classes of elements serving as arguments of these functions—vi%. contexts of utterance determining indexes and exponents. What are the values of these functions when defined?—that is the question. Assuming that no information is lacking in a context C of uttering (10), we are still inclined to say that what was said was neither true nor false. Since we assume that no information pertaining to (10) is lacking in C, the third alternative, w%. that the truth-value is undefined, is also excluded. Our functions are defined for some £7s and their values seem not to be 'true* or 'false'. What are they, then? Without adhering to any solution of that problem, we mention two useful distinctions that have been drawn in relation to some solutions. Philosophers since Frege (1918/19) distinguish between two logical elements of a sentence, the element which the sentences (9) (10) (10') and (10") share, and what they do not share. The former is the "descriptive contenf of the sentences, indicatives and non-indicatives as well; the latter is the "mood" of the sentence— indicative, interrogative, imperative, etc. —indicated in many cases by a performative verb or by the syntactical form of the sentence. E. Stenius (1964) 168 pointed out that there are different ways in which a mood/radical distinction can be applied, and introduced the terms "grammatical mood" and "semantical mood"6. Obviously, the grammatical mood of a sentence may be different from its "deeper" mood: (11) You will pack at once and leave this house. This sentence is of the indicative mood in the grammatical sense, but in many contexts it is of the imperative mood, pragmatically. In what follows we shall confine ourselves to a discussion of semantically non-indicative sentences. Our major problem in the present paper is how to present the meanings of their moods, within the formal framework outlined earlier. However we shall have something to say about the use of indicative sentences as well. 6
Jespersen used "notional" for the same purpose (1924) 319. Better terms would be "syntactical mood" and "pragmatical mood** but we shall use Stenius' terms.
Mood implicatures: a logical way of doing generative pragmatics
4.
13
Stenius and Aquist on the other moods
One theory of meaning for (semantically) non-indicative sentences postulates the existence of a sentence-radical and a sentence-mood for every sentence, indicative and non-indicative as well. Thus, at some level of representation the suggestive notation of (16)—(20) can be used for (12)—(15), respectively: (12) Do you live here now? (13) It is obligatory for you to live here now, or live here now! (14) You live here now. (15) It is necessary that you live here now. (16) ?p (17) Op (18) ip
(19)
INp1
Notice the difference beween Ip and p. The latter does not denote the sentence (14). It is a sentence-radical which does not have a mood, and it may be expressed in English by (20) rather than by (14): (20) that you live here now. The operators ?, I and others, which combine with sentence-radicals to form representations of sentences, play a role both in the syntactical and semantical processes. The syntactical role of such operators is well known, having been used by some linguists, though under different terms. Katz and Postal (1964) 86 ff and Katz (1972) 201 ff introduced a special morpheme Q into the underlying structures of yes-no question, and Ross (1970) argues to the effect that declarative sentences are derived from deeper structures including as their main verb, a complex symbol, [ + V, + performative, + communication, + linguistic, + declarative] which is defended on syntactical grounds. Whether these detailed analyses stand up to criticisms or not is not of our concern here (cf. Fräser 1970). It suffices to notice that the postulated operators used in (16)—(19) provide syntactical anchors, that is, starting points for a transformational process accounting for the surface differences between (12) and (14). The main point of Aquist's variant of Stenius' theory is that the sentenceradical specifies what should be true somewhere and the sentence-mood indicates in a sense where (Aquist 1967). When the sentence (14) is under consideration, what is true in some possible world is (20)—that you live here now—, which is the sentence radical of (14). The sentence mood of (14) indicates in which possible world (20) is true, according to what is conveyed by (14) in its standard
7
Stenius* own notation for (19) is "Np". Cf., however, his paper (1967). We shall not discuss the differences between Stenius* two works.
14
AsaKasher
use. Since the semantical mood (14) is the indicative, the possible world indicated is usually the actual world; more accurately, if a context of utterance C has the indexes < addressee: John),
, and 0, or in other words such that the relation Rm (w0, w) holds. Rm—the /^-alternativeness relation—is the core of this analysis of /^-sentence. A mood has not been characterized, unless a theory has been provided that marks the properties of Rm which make it different from any other non-intended mood-relation Am/. A theory of moods should This observation was made by Jespersen, /0r. «'/., and also by Aquist (1967).
Mood implicatures: a logical way of doing generative pragmatics
15
provide a general framework for the representation of all moods and an explanation of the variety in moods and its internal relations. Stenius* and Aquist's suggestions provide neither a general theory of moods, nor a characterization of any single one of them. Notice that a theory of moods has two different goals. Characterization of each mood separately is not enough. Explanation of the variety is impossible where no measures for comparing moods, or sentences of different moods are available, and measures for comparison can evolve only within a common framework. Hence, any adequate theory of moods should include at least some important elements of a unified theory of moods. In one sense a cleavage of logic according to mood into logics of commands, questions, advices or what have you, is a desolation of logical theories of mood. In the sequel we adopt the basic distinction between moods and radicals and the idea of characterizing moods by specifying truth-conditions in appropriate possible worlds. But first we consider a rival theory. 5.
David Lewis on other moods
David Lewis proposes a method of "paraphrased performatives (Lewis 1972, 205—212). First, an explicit performative sentence is paired with each non-indicative sentence. For example: (23) Be late and (24) I command you to be late are associated. Non-indicative sentences are "treated as paraphrases of the corresponding performatives, having the same base structure, meaning, intension and truth-value" on an occasion (Lewis 1972,208). Secondly, the truth conditions of (24) on its performative reading are equated with its truth-conditions on its non-performative, self-descriptive reading. Doubt can be thrown about the success of this ingenious program. For some reasons, indicative sentences are not on a par with non-indicative ones, which are being treated as paraphrases of some explicit performatives. The sentences (25) and (26) are taken to have distinct truth conditions: (25) The earth is flat. (26) I declare that the earth is flat. This unwarranted discrepancy between indicatives and non-indicatives is a symptom of some deeper difficulties in both parts of Lewis' program. First of all, Lewis' proposal hinges upon the assumption that to every sentence there corresponds an explicit performative. There are, however, several kinds of sentences which resist restatement. Consider the following examples9: (27) is discussed in Culicover (1972).
16
Asa Kasher
(27) (28) (29) (30) (31)
One more can of beer and I'm leaving. Clearly, I concede that I've lost the election. Be glad that we aren't taking away your teddy-bear. Go in the back door and hope no one sees you. Give me an Akvavit, or don't you approve of drinking? The difficulty in paraphrasing (27) is due to the fact that it cannot be embedded within an explicit performative hyper-structure without modifying the first conjunct of (27) to include a verbal phrase. It has, however, been shown that the "interpretation of that conjunct is systematically unspecifiable." Since systematic paraphrasing requires systematic specifying, and the latter is impossible the former is also impossible. Another problem is exposed in (28). Any movement of the adverb will result in either a non-grammatical expression or a sentence of a different meaning. Adjoining an additional explicit performative hyper-structure will not do either. Now, the following sentence (32) is not a paraphrase of (28), because the adverb in the latter is significant, while the second conjunct in the former is pointless; and (33) is not a paraphrase of (28), because (34) is not grammatical. (32) I concede that I've lost the election and it is clear that I concede that I've lost the election. (33) It is clear that I concede that I've lost the election. (34) It is clear that I hereby concede that I've lost the election. Another problem is presented in (30), vi%. how to paraphrase the second conjunct. (35) is a near paraphrase of it: (35) You should hope that no one sees you. but in a sense it is too strong since it presupposes the existence of some particular "institution" or set of rules according to which you have to hope no one sees you, a presupposition not shared by (30) as is clear from (36): (36) You don't have to hope no one sees you, but hope no one sees you. Even in the simpler case of (29) any "You should" expression is too strong since no kind of obligation is involved in standard context of utterance of (29). One might argue that "should" is ambiguous, and that no obligation is presupposed in (37): (37) You should be glad that we aren't taking your teddy-bear., but it has been argued10 that "should" is not ambiguous, and that (37) d0*r presuppose the existence of some rules that if followed would lead you to the conclusion that you rather be glad than sorry. However, even if (37) does not presuppose any kind of obligation or rules, it is not a paraphrase of (29), because the latter presupposes that you are not glad in the context of its utterance, while the former does not: (38) You are glad and you should be. 10
Wertheimer (1972). I take this argument to mean that the difference is meant to be a difference in meaning and not in use. We return to this point later.
Mood implicatures: a logical way of doing generative pragmatics
17
Our final example, (31), is a case of one speech act affecting the force of another one. This interaction is not depicted adequately by an explicit request followed by an explicit question: (39) I hereby request that you give me an Aquavit and I hereby ask whether you don't approve of it. Our hypothesis that not every sentence has an explicit performative paraphrase is not in conflict with any principle of expressibility, as formulated by Frege, Tarski or Searle, under different titles (cf. Searle 1969, 19f). According to the latter, for example, for "any meaning ÷ and any speaker s, whenever s means (intends to convey, wishes to communicate in an utterance, etc.) ÷ then it is possible that there is some expression Å such that Å is an exact expression > of or formulation of x." (Searle 1969, 20). Indeed, what is at stake here is not even whether it is possible for natural languages to provide explicit performative paraphrases for each non-indicative, but whether it is possible to express what is expressed by non-indicatives, and this has obviously an affirmative answer. Notice also that in some cases what looks like the appropriate explicit performative paraphrase of given sentence in a speech act is non-grammatical for most speakers11: (40) *I hereby ask you what time it is. (41) *I hereby fire you. (42) *I hereby point out that most speakers find this sentence unacceptable. (43) *I plead that you will spare my life. (44) *I threaten that I will kill you. (45) *I boast that I have done that. Again, the non-grammaticality of (40)—(45) does not violate the principle of expressibility because each of these expressions has a non-indicative paraphrase, having the appropriate force under standard circumstances. The lack of strict correspondence between sentences and explicit performatives is perhaps at the root of the seeming divergence of the meanings of indicatives from their purported explicit performative paraphrases. An adequate theory of explicit performatives will undoubtedly have syntactic ingredients. The other part of Lewis' program does not seem to fare better. There are several objections to the semantic identification of explicit performatives and selfdescriptive readings of sentences like (24). Lewis maintains that different sentences which share their meanings may have different uses, but he does not outline any theory of use. It remains to be seen how such a theory is worked out to accomodate a distinction between truth- and felicity-conditions, presuppositions and conversational implicatures, and so forth. In the sequel we present what is, in a sense, a theory of use, but we do not identify the meaning of a non-indicative with that of the corresponding self-descriptive sentence, if any. 11
These examples are due to J. Sadock.
2 TLIl/2
18
Asa Kasher
We turn now to another objectionable aspect of that identification thesis.12 A sentence (type) is pragmatically self-verifying (-falsifying) if, and only if, every happy use of this sentence or any other sentence having the same meaning and performativeness status must say something that is contingently true (false). It is reasonable to assume, following Gale, that a sentence is pragmatically selfverifying if, and only if, its contradictory is pragmatically self-falsifying. 'Now, from the identification hypothesis it follows that performatives are true or false, and pragmatically self-verifying, self-falsifying or neutral. Actually, "I hereby promise you to come" and similar performatives turn out to be self-verifying: in a happy use of this sentence a promise is issued, and a corresponding descriptive sentence would have been true, had it been uttered and used happily somehow in the same context, in addition to the performative sentence. Moreover, if performatives have truth-values, they have contradictories as well; the contradictory of "I hereby promise you to come", is naturally "I do not hereby promise you to come", which is, surprisingly, selfverifying as well.13 Thus, both a performative and its contradictory are selfverifying, i.e., contingently true in each of their happy uses. If one adopts Lewis' suggestion of saying one sentence and at the same time putting down in writing another sentence, one can produce a performative and its contradictory at one and the same time. Saying of one of them that its use is unhappy would be arbitrary and saying it of both is ruling out too much—one of them is true, and consequently cannot be used unhappily. Indeed, to say that both a're used happily would involve us in a contradiction. Finally, we mention an odd consequence of the identification hypothesis: Performatives are self-verifying, but their self-descriptive paraphrases are- almost always false. If we exclude the possibility of performing two different acts simultaneously, these self-description sentences are self-falsifying. It is odd that paraphrases should differ in their truth-values to such an extent.
6.
Preconditions and implicatures
The next part of this paper will be devoted to a proposal of an alternative- theory. Moods will be characterized within a pragmatical, rather than a semantical framework, which however will include a new substantial semantical element. It is fairly obvious that besides the syntactical concept of grammaticality and the semantical concept of meaningfulness there is some linguistic concept of pragmatical appropriateness. A happy linguistic use of any given sentence is pos12 13
See Gale (1970) and Sampson (1971). Both do not apply it, however, to our problem. We ignore here some'niceties mentioned by Gale.
Mood implicatures: a logical way of doing generative pragmatics
19
sible only in some possible contexts of its utterance. It is linguistically pointless to address an English sentence to a German whom you assume not to know English at all. Similarly, if your intention is to convey information directly, and not, say, to tell a joke, it is inappropriate from a purely linguistic point of view to tell your friend anything about hairs or heirs of what you call "the present king of France" if you know that either you or your friend or both of you fail to believe, at the context of utterance under consideration, that there is a person the reference to whom by that famous expression, at that context, is appropriate. From the same point of view it seems linguistically hollow to ask the girl who lives next door what the time is when you know exactly what it is. Appropriateness is a relation between speakers, sentences ^nd contexts of utterance. We shall not dwell here on the important problems of how to characterize contexts of utterance. We mentioned earlier that contexts of utterance determine, at least partially, indexes and exponents, i.e. there are functions index(C) = <speaker a, addressee(s) â, time /, place/>/,...> and Exp(C) = eC Singling out one of the indices of these contexts—the speaker a—we abbreviate the predicate "appropriate for uttering sentence 5" by speaker", applicable to contexts of utterance, to "á-S-appropriate" or "^-appropriate for a". Every adequate theory of appropriateness of contexts for sentences (for a particular language L) will have true consequences of the form: (46) Every á-J-appropriate context of utterance, C, has the property ö. At the moment we do not impose any severe restrictions on ö except that^ it depends on C or S, i.e.,
20
Asa Kasher
themselves. Since the opposite is not true—truth-statements with regards to some indexes of the context may have nothing to do with its exponent—we replace (48) by a better, two-fold reformulation of (47); every adequate theory of appropriateness for L will have true consequences of the following forms: (49) For every á-5-appropriate context of utterance, C, the sentence
Definition I A sentence S1 is an exponent ^.-precondition of a sentence S: In every a-J-appropriate context of utterance C it is true that S' is believed by á to be true at eC. In other words, S' is believed by á to be true in the circumstances that are discussed in C.
Definition II A sentence S' is a context ^.-precondition of a sentence S: In every a-^-appropriate context of utterance C it is true that S' is believed by á to be true at C. That is to say, S' is believed by á to hold in the context of utterance. It is a necessary condition for a's performing a happy speech act by uttering S in C that all the sentences that are exponent á-preconditions of S will be believed by á at C to be true in eC, and that all context á-preconditions of S will be believed by him at C to be true then. Now the concept of exponent á-precondition, for example, is not a linguistic one. Consider a speaker a' who pretends being a French monarchist. He does not believe that there is presently a person who may be properly referred to as "the present king of France", but he is never reluctant to utter sentences like "Long live the king of France!" or "The present king of France has no heirs". Obviously, the sentence "There exists a person who is presently a king of France" is not an exponent a'-precondition
Mood implicatures: a logical way of doing generative pragmatics
21
of the former sentences. In order not to blur the issue by introducing such performance effects, we restrict the realm of speakers under consideration to that of ideal speakers, who mean exactly what they say, in each and every context of utterance, in distinction from the above mentioned fake monarchist. Later on we shall elaborate this point. Meanwhile we notice that the following concepts are purely linguistic ones.
Definition HI A sentence S' is an exponent precondition of a sentence S: For every ideal speaker á (of language L to which S and S' belong) S' is an exponent á-precondition of S.
Definition IV A sentence S is a context precondition of a sentence S: For every ideal speaker á (of language L to which S and Sf belong) S' is a context á-precondition of S. In the sequel we omit the reference to L.
Definition V A sentence S is ^precondition of a sentence S: S' is an exponent precondition of S or Sf is a context precondition of S. Whenever S' is a precondition of S, S' is (semantically) indicative, at least in the ^-appropriate contexts of utterance. Otherwise S' could not have been true in these contexts. Indeed, sentences have preconditions even if they are not indicative in any context appropriate for their utterance. Nevertheless, if a sentence S is indicative in every ^-appropriate context of utterance, then S has one distinguished precondition, vi^. S itself. This should come as no surprise: when an ideal speaker utters such a sentence he is committed to granting then that the sentence is true in the circumstances under consideration. An ideal speaker believes what he asserts. Similarly, an ideal speaker C^ who utters (50) in a context C, which is (Xi-(50)-appropriate, is committed to granting that (51) is true at eC: (50) Elsberg's present psychiatrist's former office is not in Washington. (51) There is presently exactly one psychiatrist who renders his professional service to Eisberg. Bringing into the picture the ideal speaker to whom (50) is addressed in C —the addressee â—one wonders how â is involved in his fellow's preconditions. Clearly, a belief on 's part at C that (51) is true in eC is not a necessary
22
Asa Kasher
condition for a's performing a happy speech act by uttering (50) in C. If it were necessary that â believes at C that (51) is true in eC7, then á would have been required to know it at C. Being required to be thus knowledgeable, á would have to entertain true, justified beliefs about â's own beliefs. This is a stern demand. When we scrutinize a speech act of asserting, a necessary condition for its happiness is that the (ideal) speaker involved believes what he asserts; it is to the best of his belief that what he asserts is true in eC. If when we consider the heart of an assertive speech act we find that genuine belief, rather than knowledge, is demanded, why should we expect to find a demand for knowledge, rather than justified belief, when we consider one limb of such a speech act? While the addressee is exempted from actual belief at C that (51) is true in eC, he has to be taken by the speaker to entertain this belief. It is to the best of the speaker's belief at C that it is to the best of the addressee's belief at C that (51) is true in eC. Otherwise C would not have been an d-(50)-appropriate context of utterance for the speaker a. If the speaker has somehow been assured that his addressee believes that Eisberg uses the services of no psychiatrist, presently, then he is not entitled to use (50). Similarly, if the speaker takes it for granted that according to the addressee's beliefs Eisberg is presently treated by a battery of psychiatrists, the speaker is not entitled to use what purports to be a uniquely referring expression at C—"Elsberg's present psychiatrist". We see, then, that in a (50)-appropriate context of utterance the (ideal) speaker not only believes (51) to be true in eC, but also takes it for granted that the addressee shares with him this belief. An ideal speaker does not disregard the beliefs of the addressee about the population of the possible world or worlds under consideration, because by performing a speech act he participates deliberately in a cooperative activity. The speaker intends to affect his fellow's beliefs, and one cannot affect something intentionally without having some presumptions about it. It is clear that not every belief the speaker has to hold in an ^-(SOJ-appropriate context of utterance should be taken by him to be shared with the addressee. In the case of the precondition (50) itself, we should rather expect the speaker to believe that the addressee has not yet adopted it. Otherwise it would be from a pragmatical point of view linguistically pointless to address it to him. The principle we have just applied is the maxim of precluded superfluity^ taken as well to be part of our characterization of ideal speakers. Applied to phrases and not only to sentences, this maxim provides an explanation for the different degrees of pragmatical inappropriateness of (52) and (53) in standard contexts of utterance. (52) I met yesterday an unmarried, illiterate bachelor. (53) I saw in the zoo a black raven.
14
We apply Grice's observations but introduce some new terms.
Mood implicatures: a logical way of doing generative pragmatics
23
Consider now the expression "Elsberg's present psychiatrist". According to the maxim of precluded superfluity, there is a reason for the (ideal) speaker's using this expression rather than, say, "Elsberg's psychiatrist"." The reason seems to be that the (ideal) speaker believes there to be some other psychiaterist rendering his service to Eisberg, besides the incumbent. Obviously, this other psychiatrist is not another present psychiatrist of Eisberg, but rather a former or a future psychiatrist of Elsberg's. In an ^-(SOJ-appropriate context of utterance, our ideal speaker is not required to presume that he shares this existential precondition with his addressee. Nor is he required to presume that the negation of this precondition is thus shared. Hence, there are three kinds of association of a precondition with the related belief about the addressee's beliefs. Let C be any á-J-appropriate context of utterance, in which á is the speaker and â is the addressee. Denote by *¸ (ã, C, w, X)" the statement that person ã believes at C that X is true in w. Confining ourselves to exponent preconditions, the three possibilities are as follows: (54) B(a,C,eC,JO&B(a,C,eC,B( ,C,eC,y)) (55) Â(á, C, eC, S') & B(a, C, eC, ~ Â(â, C, eC, S')) (56) B(a,C,eC,O&~B(a,C,eC,B( ,C,eC,r))& & - B (a, C, eC, - B( , C, eC, J")) Definition VI S' is a strong exponent implicatttre of S: For every ideal speaker á and for every á-J-appropriate context of utterance C, in which á is the speaker and â is the addressee, (54) is true. Strong context implicature is defined similarly, using (57) instead of (54):
(57) B(a, C, C, y) & B(a, C, C, Â(â, C, C, y»
Weak exponent and context implicatures are defined differently. We do not prohibit á from believing that a certain belief is entertained by· â. The speaker is simply not required to believe that he shares a certain precondition with â, nor is he required to believe that they do not share it. In other words, á is linguistically exempted from having certain beliefs pertaining to 's beliefs. Definition VII (i) (ii) (iii) 15
S' is a weak context implicature of S: S' is a context precondition of S; Â (â, C, C, S') is not a context precondition of S; ~Â(â, C9 C9 S') is not a context precondition of S.
Formally speaking, â is a meta-variable whose value is fixed by C in a certain way. Intuitively it denotes in each case the addressee (or class of addressees). The same holds for a, mutatis mutandis.
24
Asa Kasher
Weak exponent implicature is defined similarly. We now define two natural notions of pragmatical presuppositions.16 Definition
VIII
S' is a strong (weak) exponent pragmatical presupposition of S: S' is a strong (weak) exponent implicature of both S and its negation. Definition IX S' is a strong (weak) context pragmatical presupposition of S: S' is a strong (weak) context implicature of both S and its negation. Thus, (51) is a strong exponent pragmatical presupposition of (50) and (58) is a weak exponent pragmatical presupposition of (59): (58) These are impressions from familiar places in Charlottesville17 (59) There are places in Charlottesville which are not familiar. We explicate some pragmatical concepts in terms of beliefs. Without indulging ourselves in a discussion of how belief is related to knowledge, we would like to clear away one natural misunderstanding. When á is said deliberately to believe that sentence S is true in some possible world w, what is ascribed to him is significantly less than believing truly that S is true in w. Now we apply the maxim of precluded secrecy', according to which ideal speakers do not suppress any relevant information. When 's attitude towards S's being true in w is under consideration, an ideal speaker á will not say that â believes S to be true in w, if á believes that â knows that S is true in w. Hence, a belief-statement has among its weak exponent implicatures the negation of the corresponding knowledgestatement. However, a belief-statement does not imply the negation of that knowledge-statement. If we take knowledge to be at least true, justified belief, then the two propositional attitudes are indeed compatible. Definition I, for example, does not imply that a sentence S' is not an exponent á-precondition of a sentence S if in some a-J-appropriate context of utterance C á knows that S' is true in eC. Our definitions are stipulations of truth-conditions, and consequently "believes" should be taken in the natural sense of "at least believes".18 7.
Mood-preconditions and implicatures
Every speech act, in which a sentence S is employed under certain circumstances C, centers around a manipulation of ^Ts radical at C. Indeed, what subjects a sentence-radical to the right type of manipulation, at an appropriate context of utterance, is the accompanying sentence-mood. 16
These notions are more similar to Thomason's pragmatical presuppositions than to Stalnaker's, in his paper (1972). 17 We assume that C provides reference for "these" adequately. 18 A similar case is involved in the meaning and use of expressions like "there is one...".
Mood implicatures: a logical way of doing generative pragmatics
25
In some cases sentences show their semantical mood in their grammatical mood. Explicit performative expressions are similarly efficacious. But grammatical moods and explicit performatives do not make up a general means of putting forward semantical moods. Our contention is that every sentence carries a class of weak context implicatures that characterises its semantical mood. To be sure, what we propose is a pragmatical characterization of moods. Roughly speaking, any performance of a speech act comprises two acts of choosing: the speaker can be thought of as picking out a semantically characterizable radical, and indicating by way of choosing a semantical mood to what use he is putting that radical. One way of using a radical is to import somehow that a certain predicate applies to it or that it participates in certain relations. Now, a semantical mood signifies some standard relations in which an accompanying radical participates. By choosing a certain mood the speaker shows that he believes these relations to obtain at the context of utterance. Which are these standard relations? The claim that we would like to put forward is that the standard relations are preference relations exclusively. When an ideal speaker á commands â by saying "Give me the file marked confidential", he conveys his preference of a state of affairs in which he has been given the file by â over a state of affairs in which he has not been given it by â, assuming that these two states of affairs are different only in this respect. If á asks â by uttering "Where were you born?", he makes it known to â that everything else being equal he prefers knowing where â was born to not knowing it. Similarly, upon asserting by using the sentence "The notion that the end justifies the means proved contagious", á manifests the preference on his part of one state of affairs, Wj, over another state of affairs, w2, if they differ from each other exactly to the extent that in Wx â knows that á believes that the notion that the end justifies the means proved contagious, while in w2 â does not know it. We proceed now to an elaboration of one of these observations by way of illustration. For a certain sentence S we shall show that an expression of an appropriate preference relation is a context precondition of S and a weak context implicature of it, but not a pragmatical presupposition of that sentence. Then we generalize. Consider, then, a context of utterance Cl9 in which an ideal speaker á addresses to â the following sentence: (60) Give me the file marked confidential. Let CQ be a context immediately preceding C\\ both á and â are present, and á is about to utter (60). Now, there are many possible courses of events extending somewhat beyond C0 and C\. We divide them into two disjoint classes: in the first class there are all those in which not much later than C^ â gives á the file marked confidential, and in the second there are all the rest, i.e., all those in which â does not give á that file a short while after Q. (In both cases
26
Asa Kashcr
the relevant time-interval stretches from the endpoint of Q.) Let us pick out two possible courses of events, Wj from the first class and w2 from the other one. Although we know that á wants â to give him the file marked confidential, and in Wj but not in w2 â does it, we cannot generally be assured that á prefers wx over w2. They might be different in too many respects. If in wj â gives á the file and then shoots him to death, while in w2 â just leaves á and disappears, it is reasonable to assume that á prefers w2 to Wj, rather than the other way round. However, knowing that á is an ideal speaker who utters (60) in Cj , what we can be assured of is that given w3 and w4 that are different from each other exactly to the extent that in w3 â gives á the file marked confidential a short while after Cj and in w4 he does not, á prefers w3 over w4 . Introducing a piece of notation, a preference relation that is involved in an appropriate use of (60) by á is :
(61) where "p" stands for "â gives the file marked confidential to á a short while after C±\ Assuming that (61) and similar formulae are expressible in natural languages, we shall ignore in the sequel the due distinction between (61) and its expression in any L. When referring to (60) as a precondition or an implicature of some sentence S, we shall mean that the corresponding expression in L is such a precondition or an implicature of S, respectively. Notice that p represents the radical of (61). If á uttered in C^ (62) rather than (60), what would we say about his linguistic behaviour? (62)
I don't want you to give me the file marked confidential. Give me the file marked confidential.
Two ways are open for explaining (62). Either á changed his mind immediately after uttering the first sentence of (62), in which case there is no conflict between the two parts of (62) because they reflect a's state of mind in two different moments; or, else, the first part of (62) is addressed to one person, while the second part is directed to another one. In this case there is also no conflict between the different parts of (62). If the speaker is, however, of the same mind during the utterance of (62) and he directs it all over to one and the same hearer, a conflict does arise, simply because the two parts of (62) show divergent preferences on the part of the speaker. The first part expresses one preference explicitly, while the second part, which is (60), conveys implicitly the opposite preference. (61) is a context precondition of (60)'. This should seem by now self-evident, except for the observation that not a preference on the part of the speaker is required by the definition of context preconditions, but rather his belief that he entertains that preference at the context of utterance. In many cases such preferences are accompanied with beliefs that they obtain, and vice versa, but theoretically this is not the case. At the bottom of the distinction between cases of preferencebeliefs and of actual preferences are two important observations.
Mood implicatures: a logical way of doing generative pragmatics
27
First, a distinction should be drawn between extrinsic and intrinsic preferences.19 A preference of one state of affairs over another is extrinsic when it can be adequately explained; in such cases the question "Why is this preferred to that?" does make sense, and an answer can be provided, at least in principle. When an explanation is not sought for and just a liking is involved the preference is intrinsic. Indeed, in the most natural instances of using (60) to command, the preference involved is not a's mere fondness of 's giving him the marked file. Usually á tries to adhere to his own general plans. The command he issues by uttering (60) in C^ is meant to achieve some minor goal, thus supporting a major plan or purpose. Preference relations like (61) are often consequences of more general preference relations that constitute a's general purposes and plans. Now, á is an ideal speaker and not necessarily a rational person or an ideal logician. Hence, his general preferences may fail to entail some particular preference he entertains. He believes that (61) obtains, but he would concede that he has been mistaken, if he were to be shown that (61) does not follow from, or is not even compatible with his more basic preferences. However, from a linguistic point of view it is sufficient that when á utters (60) he believes (61) to obtain. Whether (61) is congruent with d's more general patterns of conduct and belief is of no linguistic significance. (61) is a weak context implicature of (60). We have just shown that (61) is a context precondition of (60). To prove that it is also a weak context implicature, we have to show that neither B( , C, Ct (61)) nor its negation are context preconditions of (60). This can be done by pointing out two a-(60)-appropriate contexts of utterance Q and C2, such that B ( , Cj, Q, (61)) is not believed by the speaker in Q to hold in Q, while  (â, C2, C2, (61)) is believed by him in C2 to obtain then. The first case is obvious. Usually, CLhas no reason to assume that â is already aware of a* s current particular preferences. Indeed, Ct would not utter (60) in C\, had he seen that â is bringing him the file marked confidential. On the other hand, there are cases in which á has been given some reasons to believe that â knows exactly what á wishes him to do. Consider the case of tt's uttering (60) repeatedly, just because â is openly refusing to bring that file to a. After all, â is an ideal hearer, not an ideal friend, collaborator or disciple. (61) is not a pragmatical presupposition of (60), of any kind. It is not clear what one takes or should take the negation of a command or of a sentence like (60) to be. Assuming that negation applies to the radical of (60), the resulting sentence will invoke in its appropriate contexts of utterance, not (61) but rather its negation. Nothing in our arguments hinges upon the particular example we discussed. A consideration of any other sentence S and a class of ^-appropriate contexts of utterance in all of which S is semantically imperative, will lead us to
von Wright (1972) 142ff.
28
Asa Kasher
similar conclusions—a preference relation of the same structure is involved and it is a weak context implicature of S. Now, what about sentences of other semantical moods ? For our proposal to have at least some theoretical significance, we have to show that every sentence, of whatever semantical mood, carries such preference relations, and that each semantical mood bears a characteristic class of such relations. We present now without discussion some preference relations that are weakly implicated by sentences of certain different semantical moods. These relations will all be amended in the sequel, but the following examples do already show one important property of the ultimate class of these preference relations, vi^. that sentences of different semantical moods weakly implicate structurally different preference relations. This property will remain intact throughout all the amendments. (63) Did you give me the file marked confidential? (64) ÑÊÅÑ â , €é ( íé ,í 2 ) where v, is the set{Be, Cl (K . Cl p v K , Cl <-/>)> K«, C2P v K«, Cz -/>}, and v2 is the set {Be, c (Kp, Clp í Êâ, Cl ~p)9 -K«, C2p v K«, C2 -/>} where '¸ã^ c" and "K^ c" stand, respectively, for ã believes or knows about C, at C, and where C2 is a context immediately following the context of utterance Ci. p represents the radical of (63). Under similar notation: (65) She gave me the file marked confidential. (66) PREF^v,^) where vi is the set {Be,C|/>, K t C2Ba> Clñ], and v2 is the set {Be, Clp, - Kp, C2Be, Clp}. (67) If you don't give me the file marked confidential I shall dismiss you. (68) (In the sense of threatening) PREF., c (r, ~ r) & PREFa> c ({- r, q}, {- r, ~ ?})20 where ^ r is the antecedent of the radical of (67) is represented by p: if ~ r, q. ( 69) I advise you to give me the file marked confidential. (70) PREF^c/ivLVa) where vx is the formula PREFp> Cz (p, ~p)9 and v2 is the formula PREF > Cz \~p, p). (71) I hereby apologize for not giving you the file marked confidential. 20
When the arguments of a preference relation are unit sets, we don't use the set notation, but specify only the element.
Mood implicatures: a logical way of doing generative pragmatics
29
(72)
ÑÊÅÑ á ,ï( í é ,í 2 ) where vj is the formula PREFat eC (/>, ~ p), and v2 is the formula ~ PREFaf & (/>, ~/>). In a variety of semantical moods, the mutual distinctions are structurally displayed without any recourse to the internal structures of the radicals themselves. However, this is not generally so. The basic preference relation weakly implicated by (73) seems structurally similar to (61) (73) I promise to give you the file marked confidential. The obvious distinction between commands and promises is revealed in the preference relations implicated, when their structures are specified beyond the level of radicals. If the subject of p points to á and the predicate of p described a future act of a, (61) is a preference relation characterizing promises21, but when the subject of p points to â and the predicate of it to a future act of â, (61) characterizes commands.
Thus we see that the semantical moods put the accompanying radicals to characteristic uses in forming weakly implicated preference relations·, possibly restricted by properties of the radicals themselves.
Still, there are different semantical moods that are indiscernible. Even if the properties of the radicals are taken into account we have not yet provided means for telling demands from requests, and both from commands, for example, There are two ways to tackle this problem, and for reasons that will become clear soon, we are going to use both. Both a skyjacker demanding ransom and a refugee asking for political asylum are not in a capacity of issuing orders as to their particular preferences. Both convey preferences of the same structure, but obviously demands seem to be more intense than requests, which in turn may take different degrees of intensity, as shown by the possible modification of the explicit performative involved by "please", "very much", "very very much", etc. So one may try to adopt either of two approaches: demands and requests are of the same semantical mood, but employ it in different degrees of intensity, or alternatively, they are different kinds of speech acts, involving different semantical moods, which however share their structure but not their intensities. Without delving into a comparative study of the two approaches, we here adopt the second one, to distinguish between requests of different kinds, and between demands and requests. Consequently, the characterizing implicatures include a variety of preference relations, sharing their formal properties, except those pertaining to particular intensities.
21
There are necessary conditions for the happiness of an utterance of (73) but we ignore them for the moment. The same holds for the commands under consideration.
30
AsaKasher
Now, when á issues a command he presumes that both he (or she) and â believe that á has the xight to be obeyed, being in an appropriate position of authority in some mutual hierarchy, at least in the context of utterance. That á has at the context of uttering St by which he orders â to do something, or to avoid doing something, a certain authority over â, is a strong context implicature of S. If captain á did not believe that â is a soldier under his command, he would perhaps prefer â's not bringing him the file marked confidential to his jdoing it, by reasons of regulations, caprice, or what have you. Thus, what is weakly implicated by (60) is (74): (74) PREF e , Cl ( Vl ,v 2 ) where vx is the set {/>,?}» v2 is the set {~/>, q}, p is the proposition that â gives the file marked confidential to á a short while after C\ and q is the proposition that á is in a position to command â at C\. Put differently, at C± á prefers one state of affairs, w x , in which â brings him the file marked confidential a short while after Cly over another state of affairs, w2, in which â does not do it, if Wj and w2 are different just to that extent and in both of them á has command over â at Q. We shall generalize this observation later. Meanwhile we notice that this amendment of (61) enables us now to discern between commands and requests, and between commands and demands, by recourse to the structures of the preference relations weakly implicated, namely (61) and (74). Similarly, we have to include in both arguments of the characteristic preference relations of promises and warns, their appropriate positive and negative strong implicatures, respectively (cf. Searle 1969, _58f). This will render these semantical moods descernible from each other, and different from other ones in still another respect. One may take objection to our approach by pointing at a context of utterance C which is appropriate for uttering a sentence S by a speaker a, although the preference relation we claim to be implicated is not believed by á to hold at C orateC. Consider for example the case of á who is obliged to command â, at C, to give him the file marked confidential. He might have uttered the sentence (60) even though his own preferences had actually been in favour of 's not giving him that file at all. (A similar apparent counterexample concerning promises is suggested by Carter (1973). In such a case, it might be argued, (61) does not hold in C and it is not believed by á to hold then; consequently, (61) is not (weakly) implicated by (60). Now, the interpretation of a PREF-formula is done by means of certain possible worlds, which differ from each other to a certain, specified extent. We
Mood implicatures: a logical way of doing generative pragmatics
31
admit that (61) is not implicated by (60), if no restriction are imposed on the possible worlds involved, except those concerning the extent of differences between compared possible worlds. But there is nothing remarkable in this observation. We have seen already an additional restriction imposed on all these possible worlds, vi%. that the strong context implicatufes of (60) hold in each of them. When á issues a command by uttering (60) he exercises his authority over â. Even if á is obliged to issue that command he is not absolutely compelled to do-it: he may resign, he can violate a regulation, he is even in a position to commit suicide if he wants to avoid by all means uttering (60) to command â to give him the file marked confidential. By uttering (60) á shows that he is not willing to resign, to violate a regulation, or to commit suicide, in order to avoid the production of that command. Given that he deliberately avoids taking any step to avoid uttering (60), á prefers â's giving him the file marked confidential to â's not doing it, everything else-being equal. Notice that since á is an ideal speaker he cannot pretend commanding â. If he utters (60) in an appropriate context, he cannot fail to produce a command. Assuming that â believes that á has some authority over him at C9 we conclude that â believes also that if á issues a command, then á prefers its being obeyed by â over a's resignation, violation of law, or deathr Hence, these preferences are strong context implicatures. All the possible worlds involved in the interpretation of (61) are required to be such that these implicatures hold in them. When (61) is thus interpreted it is true. 8.
Pragmemes
Given that the sentence (60) (Give me the file marked confidential) weakly implicates appropriate English expressions of (74), let us compare twa possible states of affairs, w5 and w6, in both of which (75) á is in a position to command â at û\, and (76) At Ci, â knows that á believes that (74) holds, and which differ from each other exactly to the extent that in w5 (77) â does not give the file marked confidential to á a short while after Q holds, while in w6 (78) â gives the file marked confidential to á a short while after C± is the case. Obviously it follows from our ceteris paribus interpretation of "PREF" in formulae like (74) that á prefers at C± w6 over Wj. Compare now two other possible states of affairs, w7 in which (77) holds and w8 in which (78) holds, in both of which (75) and the negation of (76) are true. Again it follows from (74) that at Q á prefers w8 over w7. Thus not only (70) is weakly implicated by (60) but obviously also (79) and (80): (79) PREF., Cl ({(75), (76), (78)}, {(75), (76), (77)}) (80) PREF., Cl({(75), non-(76), (78)}, {(75), non-(76), (77)})
32
Asa Kasher
These two weak implicatures of (60) are, so to speak, parts of the meaning of the weak implicature (74); no special non-deductive rules are needed for deriving them from it. Consider however two other possible states of affairs, W9 and WJQ, that differ from each other exactly to the extent that (75), (76) and (78) hold in w9, but (75), non-(76) and fion-(78) (i.e., (77)) hold in w10. Which if any of these states of affairs does á prefer at Q over the other? Would you prefer such a w9 to such a w10 ? Which if any of these states of affairs does á prefer at Q over the other? Would you prefer such a w9 to such a w10, playing the role of á at Cj? Certainly this is the question of balancing the merits of â's bringing that file to á on time and the possible demerits of disclosing a's preference (74) to â. Clearly, by uttering (60) at C\ Ct tilted the balance towards w9. If he had prefered w1(> over w9, he would have kept silent, at least with regard to his preference (74). (81) PREFe, Cl ({(75), (76), (78)}, {(75), non-(76), non-(78)}) Since á in his capacity as an ideal speaker is neither required to believe that â takes (81) to obtain in an a-(60)-appropriate context of utterance nor required to believe that â takes the negation of (81) to obtain then, it seems that (81) is weakly implicated by (60). Notice that this seeming weak implicature does not follow from (74) logically, and in this respect it is interestingly different from the above mentioned weak implicatures of (60), vi%. (79) and (80). Actually, (81) is not a weak implicature of (60) at all, for one nicety. In uttering (60) at 6\, á shows his willingness to "pay" â for the prospective rendered service, by revealing some of his current particular preferences to â. The latter has his own purpose and plans, and he is apt to use information about á to advance his own interests, whether á would li ke it or not. In this sense, á pays â. But this payment is significantly restricted. An ideal speaker is aware of what he is saying; we assume that he knows what he is doing by uttering (60). The payment is, so to speak, under control. Consequently, á would prefer w9 over w10 if in the former á knows that â knows that á believes (74) to be true, at C^. There is no reason to assume that in his capacity as an ideal speaker á would prefer, say, wu in which he has command over â, â brings him the file on time, â knows that á believes (74) to hold, at 6\, but á does not know that â knows //, over w12 in which he has the same command over â, â does not bring him the file on time, and â does not know that á believes (74) to be true, at C\. It is imaginable that á is willing to take the risk of letting â know that he is interested in a file marked confidential,, but only on a condition that he knows what â knows about it and can take the appropriate protective steps, if necessary. A surprising disclosure of the fact that á prefers some state of affairs to another might be too risky for a. (82)
Consequently, a weak implicature of (60) is (82) rather than (81): PREF e , C l ( V l ,v 2 ) where Vl = {p, q, K , Cl (74), Ke, Cl K p , Cl (74)}
Mood implicatures: a logical way of doing generative pragmatics
33
and í 2 -{Á*~Ê,. Ãé (74)}; p and q are as in (74).
One might object to our using the knowledge operator and not the belief operator in (76) and (82). Notice, however, that our discussion has been restricted to a linguistic community of ideal speakers. Thus á means exactly what he says, and conveys exactly what is implicated by the sentences he utters, and â understands exactly what á says; á never misleads â and â never misunderstands a. When a manifests a certain belief r on his part, he entertains it at the context of utterance, and â grasps it precisely and completely, â is, therefore, in a position to know that á believes r at that context. To be sure, we assume that the community is linguistically homogeneous, and that every ideal member of it knows that he and the others are ideal and that they all know it. Obviously, (82) is logically independent of (74), but they are not prgamatically so. A rule that we dubbed elsewhere22 "the rule of the communication price" enables us to derive (82) from (74). Where "w >" stands for the relation of weak implication, we have: (R)
Sw>PREF e , c (v 1 ,v 2 ) Sw>PREF t t f C (u 1 ,u 2 ) where
and
Applying (R) to (82), one finds still another weak implicature of (60). This process can of course be reiterated indefinitely many times, yielding new weak implicatures of (60). The set of preference relations involved, characterizing the semantical mood of (60), includes this infinity of weak implicatures of (60) as well as (79) and (80) and, indeed, similar consequences. But this infinite class is finitely definable: it includes the basic preference relation (74) and is closed under (R) and under logical deduction rules. All the characteristic classes of preference relations are similarly definable. The basic preference relations of these classes we call "pragmemes"?* 22
Kasher (1973). See also Kasher (forthcoming). We do not develop here the due distinction between uninterpreted pragmemes and interpreted ones. See chapter 9, below. We introduced the term "pragmemes" in our lecture at the IVth Congress of Logic, Philosophy and Methodology of Science (Bucharest 1971). Since then it has been put to another use by Konrad Ehlich and Jochen Rehbein. 23
3 TLIl/2
34
9.
Asa Kasher
A formal theoretical framework, outlined
Earlier we have made informal use of a variety of preference relations between possible worlds, at some contexts of utterance. We shall now show how to formalize this conception within a theoretical framework, meant to be a basic, pragmatical component of an adequate linguistic theory. Let L be a natural language. A linguistic theory for L is a system M of postulates and rules, formulated in a language Ë serving as a meta-language of L. First, we outline some elements of this system. /. Sentences. Ì specifies recursively the set of sentences of L. By "sentences" we do not mean just strings of phonological units or of letters, but pairs each consisting of such a string and some syntactical surface parsing and lexical specifications of it, in a way that makes sentences to be syntactically24 and lexically unambiguous.25 In this we deviate from some common ways of understanding the vague term "sentence" but this deviation will make our exposition simpler. 2. Semantical representations. Ë includes a set of uninterpreted formulae by means of which semantic analyses of sentences of L will eventually be represented. We call these formuke "creative". (This term is the pragmatical counterpart of "generative".) These might be the well-formed formulae of a second-order predicate language with intentional operators of various kinds, with many-sorted and restricted variables, branching quantifiers and other logical devices which will be found useful.26 Ì-includes also general means for considering interpretations of the creative formulae. Adopting Thomason's terminology, a valuation consists of an interpretation assignment which associates semantical values—intensions—to atomic expressions, and a method of projection (or better: a method of combination) which determines the semantical values of molecular expressions, recursively. Where intensions are taken to be functions whose domains consist of classes of possible worlds or other semantical exponents (see section 2 above) a valuation includes also an interpretation structure which specifies a class of semantical exponents. In order to capture some linguistic phenomena, e.g. semantical presuppositions, we admit valuations that are partial in the sense that the functions which serve as semantical values of formulae or other expressions might be left undefined for some elements in their domains.
24
Notice that only surface syntax is involved here. For a similar conception of "sentence" see my paper (1972). 26 For the usefulness of intentional operators, consult Montague, loc. fit.; concerning many-sorted and restricted variables, see Kasher (1973a); a discussion of branching quantifiers is included in some unpublished papers of Hintikka; see also Gabbay and Moravcsik, this issue. 25
Mood implicatures: a logical way of doing generative pragmatics
35
3. Contexts of utterance. Another facility of M is the means it provides for considering classes of possible contexts of utterance. A major duty of M is to characterize adequately the notion of appropriateness of contexts for uttering sentences. At the moment we assume that a context of utterance C determines the values of the meta-variables á (the speaker) and â (the addressee or addressees), and the exponent eC. 4. Translation. Ì is meant to render ultimately adequate descriptions and explanations of meaning and use. For that purpose, Ì characterizes a translation function that maps pairs <^, O °f sentences and contexts appropriate for their utterance, onto triplets *, r, v) where r is a creative formula and v is a valuation; r is related to the radical of S at C2/; and m is a representation of the mood of S at C. Indeed eC should be admissible according to the interpretation structure of v. A problem that arises here is that of sentences which are semantically ambiguous, e.g. those having both a specific and a non-specific reading for one of their phrases. Our view is that this should be taken care of by the context: a context of utterance is not appropriate for uttering S if it does not resolve all the semantical ambiguities of S. Indeed, people produce and understand sentences even if used in some contexts that do not supply all the information needed for resolving the ambiguities, but then people also produce quite often and understand quite well utterances of non-grammatical expressions. These two facts should, to our mind, be dealt with similarly, in a performance-extension of the theory we outline here. Pragmatical inappropriateness should be on a par with syntactical inappropriateness. We mention in passing that we hold v fixed in most of the cases. The characterization oiuses of sentences is done within M in the following way. A given pair <5, £T> is evaluated by the translation function. If C is not J'-appropriate, then the translation function is undefined for <j\ CT>. If C is ^-appropriate, then (S, Cy corresponds to some (my r, v). The pair (r, v) is actually an interpreted creative formula which represents as a function of eC the radical of S at C. Now it is the mood's turn to be staged. Following the general conception of this paper, M ascribes to each semantical mood m a class of characteristic pragmemes. These are presented as special formulae of Ë whose main relational expression is "PREF1", where "i" refers to a degree of intensity.28 Among the
27
r is not the radical of S at C, since it is not interpreted yet. The valuation y, the exponent of Cy eC, and r, define the radical of S at C, which is an interpreted creative formula. 28 We deliberately ignore again the due distinction between PREF-formulae and their English expressions.
36
Asa Kasher
arguments of these relations are always functions of the interpreted formulae. M also includes logical rules and rules of communication (like our [R]) that are iteratively applied to the pragmemes and to the derived preference formulae. These formulae are interpreted, with the aid of the valuation v and the exponent cC. The class of these interpreted formulae represents in M the contribution of the semantical mood of S to its use at C. Another component of M, yet unspecified, will provide a representation for that part of the contribution of r to the use of S at C, which is independent of».29 We conclude this outline by showing how to formalize the interpreted preference relations. Generally speaking we adopt von Wright's logic of preference, somewhat revised according to our purposes.30 Preference relations are relations between possible states of affairs. We formalize the later notion by employing van Fraassen's concept of supervaluations.31 These are partial functions from formulae to truth-values, i.e. to {"True", "False"}. Some restrictions are imposed on these functions within M: (i) (ii)
If a supervaluation s ascribes the value "True" to x, and y follows logically from x, then s(y) = "True" ; If y is a semantical presupposition of ÷ and s is defined for x, then
The third restriction is imposed only on those supervaluations that take part in characterizing the semantical mood of S at C by preference relations. In this case we also require: (iii) If j is a strong implicature of S, s(y) is "True", if s(S) is "True". Thus, the preference relations involved in uttering S at C, are relations between possible states of affairs in which the strong implicatures of S hold. It seems that eventually we shall have to impose a further restriction, namely that the participating supervaluations render "True" what is weakly implicated by the radical of S at C9 independently of the accompanying mood^ but since we don't have even an approximation to a theoretical framework for this purpose, we do no more than mention it here. Naturally we require also that the supervaluations participating in the interpretation of a preference formula related to the utterance of S at C will all be admissible according to the valuation v that the translation function relates to 29
See however our paper (1973a) for a discussion of one aspect of the problem. Unfortunately we used there "presupposition" for few different relations, one of which is the weak context implicature, defined here. 30 von Wright, loc. «'/., and the bibliography thereof. 31 van Fraassen (1971: pp. 153—163). 31 van Fraassen (1971) 153—163.
Mood implicatures: a logical way of doing generative pragmatics
37
Finally, given two finite sets of formulae, íë and v2, we say that "the states of affairs Wi and w2 are different from each other exactly to the extent that vj holds in Wi and v2 holds in w2," just in case there are three supervaluations Si9 S2, and s3 such that st is an extension of J3 defined where s3 is defined, on Vj and nowhere else and ascribes to all the elements of Vj the value "True", and s2 is also an extension of J3 defined where s3 is defined, on v2 and nowhere else and ascribes to all the elements of v2 the value "True" and ^ defines wt and s2 defines w2. 10.
Open problems
In conclusion we would like to mention two problems that have not been discussed here. Both seem to deserve close consideration. The first problem pertains to some semantical theories as well as to our pragmatical theory. The closed set of logical consequences of a proposition produced by uttering S at C, is sometimes taken to characterize the meaning of S (at C)32. Similarly, we define a set of preference relations, which includes certain pragmemes and is closed under logical rules and under (R). This is somewhat doubtful in the sense that it takes the ideal speaker to be an ideal logician with respect to meaning and use. If one rejects the conception of logical capacities being perfect in one area of application but not in another, which is admittedly somewhat strange, it is possible for him to tackle the problem by using Hintikka's distinction between surface semantics and depth semantics.33 The second problem is deeper: What is it in the pragmemes that makes them, rather than other preference relations or other relations in general, basic elements of language use? And being so basic, why do not natural languages reflect them perspicuously in the surface structures of the sentences?
References AQUIST, L. (1967) Semantic and pragmatic characterizability of linguistic usage, Synthese 17, 281—291. BELLERT, I. (1969) On the problem of a semantic interpretation of subject and predicate in sentences of particular reference, pp.9—26 in: Bierwisch, M. and K.E. Heidolph (Eds.), Progress in Linguistics, The Hague: Mouton. CARTER, W.R. (1973) On promising the unwanted, Analysis 33, 88—92. CULICOVER, P. W. (1972) OM-sentences, Foundations of Language 8, 199—236. FR SER, B. (1970) A reply to "On declarative sentences", ML AT, NSF-24, Cambridge, Mass. FREGE, G. (1918/19) The thought: A logical inquiry, pp. 17—38: in Strawson, P.F. (Ed.), Philosophical Logic, London: Oxford University Press. 32 33
E.g. Hiz (1969), and Bellert (1969). Hintikka (1973).
38
AsaKasher
GALE, R.M. (1970) Do performative utterances have any constative function? The Journal of Philosophy 67,117—121. HINTIKKA, J. (1973) Surface semantics: Definition and its motivation, pp. 128—147 in: Leblanc, H. (Ed.), Truth, Syntax and Modality, Amsterdam: North-Holland Publ. Com. Hiz, H. (1969) Aletheic semantic theory, The Philosophical Forum 1, 438—451. JESPERSEN, O. (1924) The Philosophy of Grammar, London: George Allen & Unwin. K AS HER, A. (1972) Sentences and utterances reconsidered, Foundations of Language 8, 313—345. — (1973) Worlds, games and pragmemes: A unified theory of speech-acts, pp. 201—207 in: Bogdan, R.J. and I. Niiniluoto (Eds.), Logic, Language and Probability, Dordrecht/ Boston ;D. Reidel. — (1973a) Logical forms in context: Presuppositions and other preconditions, The Monist 57,371—395. — (forthcoming) Aspects of Pragmatical Theory, Frankfurt: Atheneum. KATZ, J.J. (1966) Mr. Pfeifer on questions of reference, Foundations of Language 3, 241—244. — (1972) Semantic Theory, New York: Harper & Row. —, P.M. POSTAL (1964) An Integrated Theory of Linguistic Description, Cambridge, Mass.: MIT Press. LEWIS, D. (1972) General semantics, pp. 169—218, in: Davidson, D. and G. Harman (Eds.), Semantics of Natural Language, Dordrecht: D. Reidel. MONTAGUE, R. (1970) Universal grammar, Theoria 36, 373—398. — (1973) The proper treatment of quantification in ordinary English, pp. 221—242, in: Hintikka, J., J.M. E. Moravcsik, and P. Suppes (Eds.), Approaches to Natural Language, Dordrecht/Boston: D. Reidel. QUINE, W.V.O. (1961) From a Logical Point of View, Cambridge, Mass.: Harvard University Press. Ross, J.R. (1970) On declarative sentences, pp. 222—272, in: Jacobs R.A. and P.S. Rosenbaum (Eds.), Readings in English Transformational Grammar, Waltham: Ginn and Co. SADOCK, J.M. (1969) Hypersentences, Papers in Linguistics 1, 283—370. SAMPSON, G. (1971) Pragmatic self-verification and performatives, Foundations of Language 7, 300—302. SCHNELLE, H. (1973) Meaning constraints, Synthese 26,13—37. SEARLE, J. (1969) Speech Acts, London: Cambridge University Press. STALNAKER, R.C. (1972) Pragmatics, pp. 380—397, in: Davidson, D. and G. Harman (Eds.) Semantics of Natural Language, Dordrecht: D. Reidel. STENIUS, E. (1964) Wittgenstein's 'Tractatus', Oxford: Basil Blackwell. — (1967) Mood and language-games, Synthese 17, 254—274. TARSKI, A. (1944) The semantic conception of truth, Philosophy and Phenomenological Research 4; reprinted in Feigl, H. and W. Sellars (Eds.), Readings in Philosophical Analysis, New York: Appleton-Century-Crofts, 1949, pp. 52—84. THOMASON, R.H. (1972) Extensions of Montague Grammars, MS, University of Pittsburgh. VAN FRAASSEN, B.C. (1971) Formal Semantics and Logic, Macmillan: New York. VON WRIGHT, G.H. (1972) The logic of preference reconsidered, Theory and Decision 3, 140—169. WERTHEIMER, R. (1972) The Significance of Sense: Meaning, Modality and Morality, Ithaca and London: Cornell University Press.
HANS-HEINRICH-LIEB1
GRAMMARS AS THEORIES: THE CASE FOR AXIOMATIC GRAMMAR (PART I)
Introduction Part I: Axiomatic grammar writing: A program
1. 1.1. 1.2. 1.3. 1.4. 1.5. 1.6. 2.
Historical setting: The status of generative grammars as theories The thesis Classical TGs: The Earlier Model Classical TGs: The Later Model (excluding semantics) Classical semantics. Later developments Re-interpreting generative grammars Correlating grammars and axiomatic theories A framework for axiomatic theories (1): Formalized systems and abstract theories 2.1. Introduction 2.2. Formalized systems 2.3. Interpreted formalized systems 2.4. Abstract axiomatic theories 3. A framework for axiomatic theories (2): Realized theories and theory integration 3.1. Realized axiomatic theories: Interpreted and applied 3.2. On the definition of "applied axiomatic theory" 3.3. Uninterpreted constants in applied theories 3.4. Integration of theories: Conflation and use 4. Grammars as axiomatic theories 4.1. The program and its linguistic background 4.2. Questions of logic 4.3. Problems of feasability 4.4. Advantages of axiomatic grammar writing 4.5. Historical notes : Index of technical terms 1
This essay is part of the work currently undertaken by a research group on Theory of Language and Theory of Grammar, directed by the author, at the Freie Universität Berlin. Basic research for the essay was begun as early as in 1970, when the author was at the Universität Kolny and was then supported by a grant from the Deutsche Forschungsgemeinschaft.
40
Hans-Heinrich Lieb
Introduction In the present essay I shall propose and develop in detail a new idea that has important consequences for the scientific study of natural knguages. (In making such a claim, I do not wish to be included among the busy revolutionaries of the linguistic scene: When time has restored perspective, today's revolution may appear as the deposing of the city mayor.) The idea is simply that a scientific grammar (as opposed, say, to a pedagogic one) of a natural language or language variety can and should be formulated as an interpreted axiomatic theory. As most new ideas, this one, too, has a history, and this one, too, may be difficult to distinguish from more familiar ones if viewed from the wrong angle. For those who might be prone to take such a view, § 1 may have a corrective function: There it is argued in detail that generative grammars as offered by the more influential schools of that province of linguistics are not theories in any ordinary sense of the term. There are indeed two methods, one developed by Wang, the other by myself, by which generative grammars can be rekted to axiomatic theories; but, so the argument goes, as the latter have advantages over the former but the former have none over the latter, generative grammarians may well be left to their scholastic squibs: Both for future work in the theory of grammar and for future grammar writing the more fruitful framework should be assumed. Given the half-life of linguistic schools, § 1 may be of historic rather than immediate interest by the time this paper appears; let me emphasize, therefore, that my main points do not depend on either accepting or rejecting any tenet of transfof mational grammar. In Part I of the essay, discussion of generative grammars is followed by an outline of a framework for axiomatic theories (§§2 and 3) that contains various improvements over traditional conceptions; in particular, there is a new account of how different axiomatic theories may be combined. Combination of theories is important because a theory of language must be combinable with a theory of communication and with psychological and sociological theories; a grammar must be combinable with a theory of language; and various partial grammars (or parts of a grammar) with each other (§ 3.4). Further comments on the framework may be found in § 2.1. The main thesis of the essay is put forward, explained, and defended in § 4: "A complete or partial linguistic grammar can and should be formulated as an applied axiomatic theory in the language of predicate logic or set theory"; in a second, stronger version the stipulation "and in terms of a theory of knguage that presupposes a theory of communication" is added (§4.1). First, those parts of the thesis which cannot be understood by reference to the general framework for axiomatic theories are expkined. Assuming a distinction between theory of knguage and theory of linguistic description (in particular, theory of grammar) as drawn in Lieb (1970), it is seen that the thesis belongs within the theory of grammar but presupposes a theory of language (which has to be used for explaining the phrase "complete or partial linguistic grammar" and for evaluation of the thesis in general). The theory of language
Grammars as theories: the case for axiomatic grammar (part I)
41
(and communication) developed in Lieb (1970) is chosen as basic, together with various extensions that are outlined in the essay itself. The pragmatic aspects of languages are explicitly recognized by the theory of language; correspondingly, "linguistic grammar" is understood in an extended sense so as to allow for semantic and pragmatic (parts of) grammars. In § 4.2, some logical aspects of the presupposed theory of language are discussed. This is followed (§ 4.3) by an attempt to counter several more or less obvious objections that can be raised against the program; in particular, it may be doubted that an extensional language of logic is sufficient for grammars, and axiomatkability should not naively be taken for granted. This discussion is followed by an account of the advantages that the proposed format for grammars will have for linguistics (§ 4.4), and by a sketch of previous work that seems to recognize or imply an axiomatic approach to grammar writing (§ 4.5).2 In Part II—"Axiomatic idiolect grammars: Testing the program"— the main proposal of Part I is systematically studied as to its feasability. Assuming that systems of idiolects and systems of languages and language varieties are similarly structured, we concentrate on idiolects and their systems, treating separately each major subsystem: the phonological, morpho-syntactic, and semantic subsystems. The presupposed theory of language is extended by additional assumptions on the systems and subsystems and on their relation to speakers and actual speech events. The structure of a grammar is outlined for each subsystem such that the grammar is an axiomatic theory of the subsystem whose 'application* (interpretative part) rektes the system to a given speaker during a given time and to given speech events. In this way, grammars for the 'syntactical' (the phonological and morphosyntactic) subsystems are discussed, and 'semantic idiolect grammars' are given detailed consideration (pragmatic ones are considered only briefly). For morphosyntactic grammars a new conception of morpho-syntactic structure is introduced that gives a prominent place to the notion of paradigm, thus resuming a central concept of traditional European grammar. Two methods of specifying the semantic subsystem are distinguished, the translation method and the direct approach; the structure of corresponding grammars is outlined separately for each approach. We conclude by discussing questions of 'integration': It is shown how partial idiolect grammars may be integrated into a complete grammar; how idiolect grammars, grammars of language varieties, and grammars of languages may be related to each other; and how non-linguistic theories such as psychological and sociological ones can be linked to linguistic ones. The present paper might be called abstract to the second degree if we 2
Linguists may find Part I of the present essay more accessible by reading § 1 and subsections 4.4f before turning to the more technical parts.—Explicit recognition of the axiomatic approach is quite recent; relevant work is only beginning to appear. I may perhaps claim priority—for whatever that may be worth (cf.fn.l). I am happy to find my position strenghtened by other authors taking a similar stand.
42
Hans-Heinrich Lieb
consider as concrete a theory of an individual language or idiolect, as abstract a general theory of language, and as even more abstract an investigation not into languages or language but into the properties of theories whose subject matter are idiolects or languages. In order to persuade the reader that such an undertaking is important for the scientific study of natural languages, I conclude by indicating three of the benefits that may be obtained from axiomatic grammar writing (for a fuller picture, cf. § 4.4, below): The discussion of theories in the philosophy of science will directly apply to grammars; in particular, questions of interpretation, confirmation, and evaluation of grammars can be discussed directly within the framework of the philosophy of science. Secondly, it can be stated clearly what a theory of language contributes to grammars of individual languages, or language varieties, or idiolects. Thirdly, it can be shown in which way a grammar may be formally linked to theories on psychological, sociological, or other aspects of communication so as to give a fuller picture of actual language behavior—which goes a long way to explicating the concept of pragmatics for natural languages, a task in which generative grammar so far has failed completely. PARTI AXIOMATIC GRAMMAR WRITING: A PROGRAM
1. /./.
Historical setting: The status of generative grammars as theories The thesis It has been asserted repeatedly by Chomsky and his followers that a transformational generative grammar of a natural language is a 'theory'.3 On closer inspection this turns out to be a questionable claim. It has never been clear what the subject matter of such a theory should be; in Chomsky (1965) Ch. 1, the reader is confronted with a series of vague and possibly inconsistent characterizations, which remain vague even in the light of Chomsky's later work.4 Disregarding problems of subject matter, a more fundamental objection remains: A theory in any customary sense of the term says something about something; more specifically, it contains expressions that can be understood as statements on the intended subject matter of the theory. (This formulation is still naive, deliberately so at the present stage of discussion.) A transformational generative grammar does not contain any expressions that may be understood in this way if a Chomskyan framework is assumed: A grammar is an algorithm for specifying expressions of a 3
E.g., Chomsky (1955) Ch. 1; (1957) 49; (1965) Ch. 1 pass.; Katz and Postal (1964) 2; Bach (1964) 9. 4 -Juxtaposing some of the well-known formulations in Chomsky (1965) brings out their possible inconsistency: "A grammar of a language purports to be a description of the ideal speaker-hearer's intrinsic competence" (3). "...by a generative grammar I mean simply a system of rules that in some explicit and well-defined way assigns structural descriptions to sentences" (8). "Any interesting generative grammar will be dealing, for the
Grammars as theories: the case for axiomatic grammar (part I)
43
'formal system'; as such, it does not say anything about the language of which it is to be a grammar or a theory. This objection can be raised not only against any Chomsky-type grammar, whether based on the 1955 theory or the latest version of the Standard Theory: Any generative semantic grammar is confronted with the same difficulty. Such grammars, it is true, have not been claimed to be theories; even so, we are left with the problem of how to relate them to their intended subject matter in a more than intuitive way. The interpretative problem for Chomsky grammars was first analyzed in Lieb (1967), (1968£) and a solution suggested. Using a different approach Wang (beginning in 1968) was eventually led to establish a direct link between transformational generative grammars and axiomatic theories. To give full historical perspective to the present essay I shall first show in detail why a generative grammar cannot be a theory of its intended subject matter, and then discuss the attempts made independently by Wang and myself to overcome the ensuing difficulties. It will be concluded that these attempts, whatever their value, are misguided: Writing a grammar as an axiomatic theory is superior to choosing an algorithmic format and does not depend on it. For my arguments concerning generative grammars I will concentrate on Chomsky-type grammars up to 1965: During this period the formal properties of grammars are reasonably well defined, and the relevant arguments carry over to later developments in generative grammar, which as a rule are rather sketchy in their specification of grammatical form. Let us understand by "classical TG" any transformational generative grammar (of a natural language) that conforms to the theory of transformational generative grammars up to and including the versions proposed in or based on Chomsky (1965). Let us call this theory "the classical theory". I will argue for the following thesis: (1) If understood within the framework of the classical theory, neither a classical TG nor its levels (as far as they are specified) contain expressions that can be understood as statements on the intended subject matter of the grammar.5 most part, with mental processes that are far beyond the level of actual or even potential consciousness" (8). "A grammar can be regarded as a theory of a language [my emphasis, H.L]; it is descriptively adequate to the extent that it correctly describes the intrinsic competence [my emphasis] of the idealized native speaker" (24). Among the host of more or less competent discussions of Chomsky's "competence", Botha (1968) § 3.5, may be mentioned as an early ambitious attempt; in Lieb (1970) Chs. 10 and 11, I have investigated a number of relevant problems in a new theoretical framework. 5 I am using "statement" informally but assuming that any statement is a sentence of some natural or constructed language that can be assigned a truth-value. In arguing for (1) I shall disregard formulations such as, "We can consider a grammar L to be a function mapping the integers onto .L...." (Chomsky (1959) fn. 1): Under this interpretation of "grammar", (1) would be trivially true.—A position similar to.(l) is also taken in Hermanns (1971).
44
Hans-Heinrich Lieb
The earlier and the later forms of the classical theory will be kept apart, and the proposals for semantic description are treated separately. 1.2.
Classical TGs: The Earlier Model By "the Earlier Model" we mean the classical theory from 1955 to 1963, excluding the Katz-Fodor proposal for a semantic component (Katz and Fodor 1963). For this period the original presentation in Chomsky (1955) remains fundamental for grammars and their levels. This means, in particular, that all 'constituents' of a grammatical component (e.g. of the phrase structure component) are 'strings' of some level of the grammar or set-theoretical entities (e.g. relations) obtained from such strings.6 Given a grammatical level, the only entities that can reasonably be considered as interpretable expressions are the strings of that level. In Lieb (1968£) § 5, I investigated for all levels except the transformational one how the strings may be interpreted if the framework of Chomsky (1955) is assumed. As it turned out, those strings which are interpretable at all can be understood only as names of entities of the object language (of the grammar), or as names of sets of strings. Hence, none of these levels contains expressions that may be understood as statements on the intended subject matter of the grammar, for the simple reason that it does not contain expressions that may be understood as statements. Consider, in particular, the set of phrase markers which is a constituent of the phrase structure level. According to Chomsky (1955) Ch. 6, a phrase marker is a set of strings of symbols of the phrase structure level, vi%. the set of all strings occurring in the elements of a set of derivations which are equivalent in a certain respect (which can be represented by the same 'tree'). Neither the individual strings nor the set of strings (or, for that matter, the corresponding 'tree') is or can be understood as a statement. Hence, it is misleading to adopt the usual terminology and call a phrase marker a 'structural description' of a sentence: A phrase marker is not a statement and therefore does not 'describe' anything in any ordinary sense of the word. Now it is indeed possible to obtain statements on the basis of a phrase marker when terms such as "is" or "dominates" are defined for substrings (including individual symbols) of strings in a phrase marker (cf. the '6-relations' in Chomsky (1955) e.g. 202). Any such statement, however, will be a statement on strings to the effect that a certain relation holds between substrings of a phrase 6
The interpretation of "constituent", above, depends on the logical structure assumed for a grammatical component (or level); e.g. in Chomsky's formulation of the phrase structure component (1959, Section 2) the constituents are the members of an ordered #-tuple. A component of a grammar must be distinguished from the corresponding level; thus, the phrase structure component is (or has as a constituent, or contains in some other way) a set of 'rules'; the phrase structure level has as one of its constituents the set of phrase markers specified by those rules. The set of rules belongs only to the component, the set of phrase markers only to the level.
Grammars as theories: the case for axiomatic grammar (part I)
45
marker. As these strings do not form part of the intended subject matter of the grammar (which is a 'natural language', in some sense of the term), the derivative statements do not refer to it, hence, they do not justify a terminology in which a phrase marker may be called a 'structural description' of a sentence (of the natural language).7 For the grammatical levels studied in Lieb (1968£) it was found that they do not contain expressions which may be understood as statements. This result carries over to the corresponding components of the grammar since their only interpretable expressions are strings of the levels. To be more specific, consider the phrase structure component of a grammar as characterized in Chomsky (1959) Section 2. This is an ordered septuple (or would be, in a stricter formulation). One of its constituents is a two-place relation, denoted by "—»", whose members might be considered as strings of a phrase structure level in the sense of Chomsky (1955). It is assumed that there is a finite subrelation of —>, /'. e. a finite set of ordered pairs of strings, with certain specified properties; the elements of this set are called the rules of the grammar (i.e. of the phrase structure component of a grammar). The 'rules' are the only entities of the grammar that it might seem possible to interpret as statements. Now a rule is an ordered pair of strings, and however the two strings are interpreted, the pair of strings is not (and can hardly be) considered an interpretable expression. On the other hand, an expression of the form ö—* ÷9 where "ö" and "÷" stand for arbitrary strings, is easily interpreted as a sentence of symbolic logic stating that the relation —» holds between ö and ÷. We might redefine "rule" so as to refer to such expressions. But then a rule would be a statement only on two strings of a grammatical level, which do not belong to the intended subject matter of the grammar but can at best be interpreted to refer to it. Hence, (1) would remain unaffected: A rule could be understood as a statement but not as one on the intended subject matter of the grammar. The transformational level and component have not been considered. Because of well-known insufficiencies of the early conception of transformations I will take them into account only in connection with the later model of transformational generative grammars for which more adequate treatments are available. Our later results carry over to the transformational component of the Earlier Model; they do not say much concerning the earlier transformational level because such a level was left unspecified in the later model. 1.3. 7
Classical TGs: The Later Model (excluding semantics) By "the Later Model" we understand the classical theory from 1964 (when
In Chomsky (1955), and frequently in later versions of the ckssical theory, the distinction between object language and metalanguage is blurred; the above statement is, therefore, somewhat questionable. For a detailed discussion, see Lieb (1968) §§ 1.3 and 6. For a method of relating a phrase marker to a true 'description* cf. § 1.6, below.
46
Hans-Heinrich Lieb
essential points of Chomsky 1965 were first presented) to, roughly, 1968, including the 1963 Katz-Fodor proposal for semantics and the treatment of transformations in Ginsburg and Partee (1969). Thus, most of the Later Model corresponds to the 'standard theory', to the exclusion of 'the extended standard theory' as developed, in particular, in Chomsky (1972£) (a collection of articles published or prepublished years earlier) and Jackendoff (1972).8 ' We first consider the transformational component as conceived in Ginsburg and Partee (1969).9 Within this framework the transformational rules cannot be understood as statements on the intended subject matter of the grammar. Ginsburg and Partee define a transformational rule "as an ordered pair (£>, C), where D is a domain statement and C is a structural change statement on D" (310). "A domain statement is any expression of the form Dv or DI Ë Z?0" (309), where DI is a finite string of symbols from a certain set and D0 is a 'Boolean domain statement'; the latter concept is defined recursively as either (a) a finite string of symbols from the given set or (b) any expression of one of the forms (Di v £>2), (D^ Ë D2),~ Diy where DI and D2 are Boolean domain statements (309). A structural change statement on D (D a. domain statement) is again a string of symbols. The expressions "domain statement" and "structural change statement" seem to suggest that the members D and C of a transformational rule are intended to be interpreted as statements. There is, however, not the slightest hint how this might be achieved. In the case of the domain statement only a single interpretation suggests itself: Given a certain 'graph assignment', the domain statement might be said to denote the set of 'trees' that 'satisfy' the statement for that assignment (cf. the definitions I.e. 311; "tree" is not defined). For the 'structural change statement', I do not see any interpretation at all. The transformational rule as a whole remains uninterpreted. If an interpretation is sought, there is only one that would seem defensible: The rule (Z>, C) denotes a two-place relation (more specifically, a function) between 'trees', vi%. between trees satisfying D and their 'transforms'.10 Even so, a transformational rule is not understood as a statement. There is an informal way of specifying transformational rules which makes use of an arrow notation. This notation reappears in Ginsburg and Partee (1969) 313 as "ô=>7\^ô'", which, on the basis of a similar formulation I.e., may be read as "T and/change ô into ô'" (Ã a transformational rule,/a graph assignment, ô 8
For the two {errns, see Chomsky (1972*) § 3.2. Ginsburg and Partee characterize the scope of their paper as follows (297 f): "As yet, no mathematical model has been given which encompasses most of these different versions of a T-grammar. The purpose of this paper is to propose one such mathematical model". (198, fn.:) "One important exception is the notion of "syntactic features" described in Chomsky (1965) and included in most subsequent T-grammars." 10 Again, denotation would have to be relativized to a graph assigment.
9
Grammars as theories: the case for axiomatic grammar (part I)
47
and ô' trees).11 By binding the variables or replacing them by constants, we would obtain an expression that could indeed be interpreted as a statement. However, this would be a statement on transformational rules, graph assignments, and trees. From Ginsburg-Partee it is fairly clear that the rules belong to a grammar of a language, not to the intended subject matter of the grammar, however defined. Therefore, any such expression, although interpreted as a statement, would not (or not completely) be interpreted as a statement on the intended subject matter of the grammar. Given the definition of "transformational grammar" I.e. 315, the transformational rules and their two components are the only entities of the transformational component for which an interpretation as statements might be considered.12 Ginsburg and Partee formalize the transformational rules of syntax only. In the Later Model the rules of the phonological component came eventually to be interpreted as being, in a sense, transformational too. The most explicit formulation of this view is found in Chomsky and Halle (1968) 20: The phonological rules are said to be 'local transformations' in the sense of Chomsky (1965).13 Such transformations are apparently covered by the Ginsburg-Partee formalization if it can be extended to allow for the 'features' of phonology, which seems possible.14 In this case, our previous results carry over to the phonological rules. The situation is, however, complicated by the fact that Chomsky and Halle develop a formalization of their own (I.e. 390—399) which makes the phonological rules very different formal objects from the transformational rules of syntax as usually conceived and as formalized by Ginsburg-Partee.15 I will not try to reconcile the two versions.16 Suppose that the Chomsky-Halle formalism is adopted. No explicit interpretation is given but it is clear that rules are the only expressions that could possibly be interpreted as statements (besides rule schemata, which can be reduced 11
Ginsburg and Partee omit reference to/, which is, however, presupposed in the context introducing the arrow notation. 12 It is not clear if and how on the basis of Ginsburg and Partee a transformational level could be constructed that would be compatible with the theory of levels in Chomsky (1955) or an appropriate modification of that theory. In the present context, such a level must be considered as unspecified. 13 "By a local transformation (with respect to A) I mean one that effects only a substring dominated by the single category symbol A" (/. c. 215, fn. 18). 14 Ginsburg-Partee did not consider grammars with 'syntactic features', see above, fn. 9. 15 The phonological rules as specified I.e. 391, (6), are isomorphic to context-sensitive phrase structure rules in the following sense: If the so-called 'units' in a rule are replaced by single symbols, then a phonological rule has the same form as a context-sensitive phrase structure rule, understood as an expression of the form ö—* ÷. The formalism is eventually expanded to include 'rule schemata' (393—396). 16 The situation is further complicated by the fact that Chomsky and Halle consider extending their formalism to cover certain rules "which are rather similar to transformations in their formal properties" (I.e. 398), i.e. similar to syntactic transformational rules in their usual format.
48
Hans-Heinrich Lieb
to rules, cf. 394, (13)): In a rule ö—»÷, the arrow is taken as denoting a two-place relation whose members are somehow 'given* by ö and ÷, where "ö" and "÷" are to be replaced by strings of certain primitive symbols (specified 390, (1)). As a reading of the arrow "rewrite as" is given; furthermore, the rules are 'applied* so as to allow the transition from strings to strings (391 f, (8) and (9)). Hence, if taken as a statement, a rule should be taken as a statement on strings. The following formulation may be roughly adequate: "For any string x corresponding to ö and any string y corresponding to ÷, x is rewritten as y" (where "corresponding to" can be defined on the basis of 391, (8), and 392, (9)). The strings are not understood as belonging to the intended subject matter of the grammar; they are apparently considered as expressions that should be interpreted with respect to it. Hence, even if considered as a statement, a rule is not to be taken as a statement on the intended subject matter of the grammar. Let us now briefly consider grammars according to Chomsky (1965), disregarding the semantic component. The categorial component of the base and the phonological component are covered by our previous discussion. The same holds for the (syntactic) transformational component if we assume that taking into account 'syntactic features' will leave unaffected our previous statements on transformational rules. Although I have not made a detailed investigation, this seems to be true with the following qualification: If the restated transformational rules involve complete 'lexical entries', those entries are among the entities which might be considered as statements. There is, however, no explicit interpretation to this effect, and an appropriate interpretation would not invalidate our thesis (I).17 Hence, the lexicon of the grammar (consisting of lexical entries and 'redundancy rules', which are not problematic with respect to our thesis (1)) also contains no expressions that can be understood as statements on the intended subject matter of the grammar. Thus, thesis (1) holds for all the non-semantic components of a grammar conforming to Chomsky (1965), which is the key formulation of the Later Model.l8 1:4.
Classical semantics. Later Developments
As conceived in Katz and Fodor (1963), the semantic component of a grammar consists of a dictionary and a projection rule component. (The revisions 17
A lexical entry is a pair (Dy C) where D is a phonological distinctive feature matrix, and C is a set of specified syntactic and other features (Chomsky (1965) 84, 87). These features are expressions such as " + N" (corresponding to the phonological specified features in Chomsky and Halle (1968) 391, (6a)). It might be proposed that "(£>, Q" is to be interpreted as synonymous with the conjunction of statements "D is F! and ... and D is Fn", where Ft, ..., Fn are the features in C. This is a statement on a phonological distinctive feature matrix (such matrices are the values of the variable "D"\ the expressions substitutable for "/}" are names of matrices). The distinctive feature matrices do not, however, belong to the intended subject matter of the grammar (see our fn. 4); they are at best expressions that can be interpreted with respect to it. 18 The concept of level must again be regarded as unspecified.,
Grammars as theories: the case for axiomatic grammar (part I)
49
in Katz and Postal (1964) are irrelevant to our problem; hence, we use the earlier conception). The projection rules are indeed formulated as statements. They are, however, statements on paths in phrase markers and on their 'amalgamation'. The paths and their amalgams consist (in some sense of the word) of 'lexical strings' and strings of 'syntactic' and 'semantic markers' and 'distinguishes'. The markers and distinguishers (and strings and paths?) are regarded as expressions that should be interpreted with respect to the intended subject matter of the grammar—they do not belong to it.19 Hence, the projection rules are not statements on the intended subject matter of the grammar (which apparently consists of certain 'abilities' of speakers, cf. Katz and Fodor (1963) [(1964) 484, 493]). This should also follow from our § 1.3, if Katz's claim is correct that projection rules are transformational rules (Katz (1971)). In the case of dictionary entries, be it in the original form of Katz and Fodor (1963) or as modified in Katz (1967) 144f, 149, we may argue in a similar way as for the lexical entries according to Chomsky (1965).20 The 'semantic interpretation of a sentence' contains statements on that sentence (Katz and Fodor (1963) [(1964) 503]), and the interpretation might appear among the entities of a 'semantic level' (left unspecified) which would correspond to the semantic component. The 'sentence' is, however, not a sentence of the natural language but a corresponding string of symbols of a grammatical level (cf. Lieb (1968£) § 3) and for this reason not part of the intended subject matter of the grammar. Hence, thesis (1) remains unaffected. This concludes our argumentation in support of (1), which by now should be well established. The thesis is restricted to classical TGs; our original claims in § 1.1 were more general, covering most generative grammars that have been suggested. We shall briefly discuss whether (1) can be extended to later developments in generative grammar. It is natural to do so in the present context since in those developments it was treatment of semantic phenomena that came to be the key issue. Katz (1972) is an updated and expanded version of 'classical' semantics and might have been subsumed under it (as the most elaborate version of the standard 19
There are a few passages in Katz and Fodor (1963) where "marker" does not seem to refer to an expression [(1964) 518]. But cf. Katz (1967) 129: "A semantic marker is a theoretical term that designates a class of equivalent concepts or ideas." 20 There are, however, certain problems with this: (1) So-called complex semantic markers (as introduced by Katz) may perhaps be reconstructed as open formulas of predicate logic (cf. Bierwisch (1969); accepted in principle in Katz (1972) 166, fn. 21; criticized in Bartsch (1971) 43—48). (2) Whereas markers in Katz and Fodor seem to denote classes of lexical items that share a factor in their meanings, they are later interpreted by Katz as denoting a class of 'ideas' (see previous footnote). As the phonological part of a dictionary entry continues to be understood as a phonological distinctive feature matrix (Katz (1967) 144)— i.e., as an expression—an interpretation on the lines of our fn. 17 would require some adjustment. 4 TLIl/2
50
Hans-Heinrich Lieb
theory in the sense of Chomsky). The following differences are the most important ones in the present context: (a) Introduction of additional notational devices, in particular, of 'categorized variables'.21 (b) Replacement of different projection rules by a single one and assignment of this rule to 'general semantic theory* rather than to individual grammars (Ch. 3, § 10). (c) Introduction of a new interpretative component of grammars to take care of 'rhetorical' phenomena such as topic and comment. The new notation is covered by our previous discussion22. The single projection rule is again of a metalinguistic nature; moreover, its status is no longer relevant because the rule is no longer part of individual grammars. Of the 'rhetorical interpretations' it is explicitly stated (I.e. 433) that they "will employ the same vocabulary of representation as semantic interpretation does." Thus, thesis (1) seems to hold for Katz (1972) too. Let us briefly consider 'the extended standard theory' as proposed in Chomsky (1970£), (1972*) (including the 'lexicalist hypothesis' of Chomsky (1970*)) and developed most fully in Jackendoff (1972). Outside semantics, the only innovation that need concern us is the use of syntactic features instead of non-lexical category symbols such as "NP", both in Chomsky (1970*) and Jackendoff (1972).23 This extension of the feature notation is easily covered by an extension of our earlier discussion (above, fn. 17). As for the semantic rules, Chomsky does not make any proposals of his own. Jackendoffs 'projection rules' (e.g. (1972) 107, 293) are statements or pseudo-statements (using imperatives) on formal objects of the grammar (readings, semantic markers, syntactic phrase markers, and the like). In the case of his 'semantic structures', in particular, his 'functional structures', I am at a loss of how to understand them because of poorly explained notation. With this proviso, then, thesis (1) may be extended to the extended theory as formulated by its main proponents, and thus to the more important recent developments in *interpret(at)ive' semantics as a whole. An extension of thesis (1) should also be correct for most or all of generative semantics. Concerning the conceptions developed by McCawley and Lakoff, we argue as follows.24 As far as a systematic account of grammars has 21
Ch. 3, § 9. Also: various parentheses p. 165f; a number of 'abbreviatory* symbols (e.g*> p. 314). 22 No demonstration will be attempted. Of a semantic marker it is now stated that it "is a theoretical construct which is intended to represent a concept" (Katz (1972) 38); the question of what the ontological status of a concept is "will be left here without a final answer" (39). 23 In Chomsky, there are additional changes in the form of the base rules, which, however, are of no consequence with respect to thesis (1). 24 I had worked out a detailed argument on the basis of Lakoff s unpublished book [1969] only a small part of which was published as Lakoff (1971). In 1972 I learned from the author that his book is being completely rewritten. I therefore decided not to publish my analysis, which, however, led to basically the same results as the following argument.
Grammars as theories: the case for axiomatic grammar (part I)
51
been given, the only expressions of a grammar that could possibly be understood as statements are the rules of the grammar. All rules (whether 'descendants' of the earlier phrase structure rules or of the transformational rules) are understood as specifying 'constraints' on 'derivations' that are sequences of phrase markers. The phrase markers do not belong to the intended subject matter of the grammar (a natural language, in some sense) because they are partly constructed from metalinguistic symbols such as "S", "N" etc. Hence, the rules cannot be understood as statements on the intended subject matter of the grammar.25 The work of Keenan (e.g. Keenan (1972), Keenan to appear) should also be covered by our previous arguments: There are rules for specifying an interpreted formal system certain expressions of which are the 'logical forms' or 'logical structures' of sentences of a natural language. Neither the rules nor the expressions are statements on the intended subject matter of the grammar. The logical forms are to be related to sentences of the natural language (rather, to names of such sentences) by transformations. To the transformational rules our previous arguments (§ 1.3 and this paragraph) apply.26 In a way similar to Keenan, Bartsch (1972) and Bartsch and Vennemann (1972) are aiming at greater logical explicitness within a generative-semantics-type framework. Thesis (1) can also be extended to cover their work although this is somewhat difficult to establish due to imperfections of their formalization (e.g. Bartsch (1972) Ch. XX).27 25
LakofPs account of the concept of phrase marker in [1969] is completely inadequate if taken literally, though similar in intent to McCawley's definition of tree ((1968) 244£), which, allowing for minor modifications, may be taken as basic for later work (McCawley (1972) assumes the 1968 account: 543, fn. 4). Our above statement on phrase markers is based on the following feature of McCawley's conception: McCawley includes a 'labeling relation' among the constituents of a tree and speaks explicitly of trees "whose non-terminal nodes are labeled by syntactic category names" (246). It should be noted, though, that McCawley introduces "tree" as a purely set-theoretical term. One might try to apply his definition to trees that are entities of the natural language: The set of 'nodes' would have to be identified with a sentence of the natural language and certain of its parts, the set of 'labels' with a set of syntactic categories (not category names). But certain conditions in the definition would now exclude many trees of the usual sort. (The difficulties are with multiple occurrence of the same part of the sentence, and with assigning a single part to several categories simultaneously, e.g. to N and NP).—For different but related criticism, cf. Bellert (1972£) 294f, Bellert (1973). 26 For Keenan's own view of his work, cf. the following quotations ((1972) 430, 453): "Using the naturalness condition presented in Part I we shall argue here that the L-structures we have proposed are natural transformational sources for NL-structures." ("L": "logical"; "ML": "natural language".) "Our work is akin in spirit to that in generative semantics (G. Lakoff 1969). The difference here is more one of emphasis: we have been more concerned to formulate rigorously certain logical properties of NL and less concerned to define the functions which derive surface structures from logical structures." 27 The question was discussed with the authors in winter 1972/73 in the Berlin research group on Theory of Language and Theory of Grammar.
52
Hans-Heinrich Lieb
Our sketch of recent developments in generative grammar should be sufficient; it would be rather useless to try and aim at exhaustiveness. Most forms of generative grammars that were not considered are systematically compared in Hirschman (1971); apparently they fall in line with the forms that were studied. Thesis (1) or an appropriate extension may now be considered as wellestablished for the majority of all work undertaken in generative grammar since its inception. Hence, we have also established the thesis that most generative grammars cannot be considered as theories of their intended subject matter in any ordinary sense of "theory (of)"; moreover, we are confronted with the general problem of how those grammars can be interpreted so as to 'refer', in a reasonable sense, to their intended subject matter. It is these questions that were taken up independently by Wang and myself and studied with respect to the syntactic components of classical TG grammars. 1.5.
Re-interpreting generative grammars In my own work I attacked the problems connected with (1) by what may be called the re-interpretation approach, as opposed to the correlation approach used by Wang. Briefly, the first method attempts to re-interpret the rules of a generative grammar so that they may be understood as statements on the intended subject matter of the grammar; the second method attempts to correlate an axiomatic theory to a given generative grammar.28 In Lieb (1967), an idea formulated in Lieb (1968£) 379, is applied to (one version of) the categorial subcomponent of a grammar as understood in Chomsky (1965)29: The rules are re-interpreted so as to become statements on the system of the natural language. For this purpose, all rules are considered as expressions containing "—»", and are first interpreted in a standard way as statements on strings of a grammatical level. The lexical and grammatical formatives occurring in the terminal strings of the component are then interpreted as names of certain entities called "-L-constructs"; the basic morphological elements of the system of the object language are assumed to be Æ,-constructs. The terminal strings are interpreted to denote appropriate sequences of L-constructs. A category symbol 28
The possibility of re-interpretation seems to be naively assumed in much work within the generative framework (see also the quotations from Bach (1964) in Lieb (1967) 369 f). For an approach similar to Lieb (1967) though less formal and questionable in detail, cf. Hermanns (1971). The conception of syntax developed in Bellert (1972 £) may be considered as a first step to either the re-interpretation or the correlation approach; cf. below, § 4.5. (Both Hermanns and Bellert are unaware of the work done by Wang or myself.) In a number of publications Marcus distinguishes and compares 'generative' and 'analytic' models of language (esp. Marcus (1967 *), (1969)). Although this parallels a distinction between algorithmic and axiomatic grammars, Marcus does not recognize the problems presented by our thesis (1) and considers the distinction between la linguistique analytique and la linguistique synthetique ou generative as a distinction essentielle^ irreductible a une autrey plus elementaire ((1969) 320). 29 Lieb (1967) was written later than Lieb (1968£) and actually published only in 1969.
Grammars as theories: the case for axiomatic grammar (part I)
53
denotes "a class of sequences of /.-constructs, roughly, the class of sequences with the following property: Each sequence is denoted by a certain part of the terminal string of some generalized phrase-marker, where the phrase-marker is a deep structure underlying a surface structure, and the part of the terminal string is dominated by the category symbol." (Lieb (1967) 371). Now a second interpretation of the rules is given, which makes them statements on the system of the object language, formulated either in predicate logic or in set theory. To give an example: Suppose that "S—*NP^VP" is one of the rules. "S", "A//5", and "KP" denote the same classes as the corresponding category symbols "S", "NP" and "VP".30 The arrow is interpreted as £=, and " " as denoting the product operation on classes of sequences, e.g. "NP VP" denotes the class of sequences obtained by concatenating every sequence in NP with every sequence in VP. On this interpretation, the rule is equivalent with "Every element of S consists of an NP followed by a VP" (understood as "Every element of S is the concatenation of some NP and some VP"), which corresponds to an intuitive reading given to such rules by many linguists (as could be easily documented from the literature). A way is suggested as to how the rules, understood as statements on Æ,-constructs and classes of Æ,-constructs, can be understood as statements on units or categories of the object language at least in the case of a 'correct grammar'. Any such possibility presupposes the second interpretation of the rules.31 Lieb (1967) does not yet formulate definite results but outlines a program for research. In subsequent work (unpublished) I carried the program through in complete detail, but eventually discontinued the entire line of research for two reasons: (a) When the classical theory was supplemented by the interpretative part I had worked out, the whole system seemed to be unnecessarily complex, especially if compared to the conception of theories and their interpretation as developed in the philosophy of science, (b) After 1965, the classical theory dissolved. In Lieb (1967) the question of how to understand a classical TG as an interpreted axiomatic theory was not treated explicitly; rather, I confined myself to the immediate problems raised by thesis (1). The correlation approach developed by Wang is concerned with the former question. It has grown out of work with a somewhat different orientation, which we will also characterize. /. 6.
Correlating grammars and axiomatic theories
In Wang (1968), a grammar of a natural language is conceived as a system of rules by which 'grammatical statements* (einfache grammatische Aussagen, I.e. 21) 30
For reasons explained in Lieb (1968£) § 2.2, the letter symbols in rules should be italicized and thus distinguished from the corresponding symbols of the grammatical levels. 31 That interpretation is not a psychological one; hence, it may be questioned whether we are establishing a direct relation to the intended subject matter, cf. the quotations above, § 1.1, fn. 4. The same objection may be raised against Wang's solution.
54
Hans-Heinrich Lieb
can be derived. These statements have the form of sentences of first-order predicate logic consisting of a one-place predicate and a closed individual expression. (For formulating the rules Wang also introduces expressions analogous to individual variables.) Wang's 'statements' can indeed be interpreted as statements, and they are obviously intended as statements on the natural language.32 It is proposed (Lc. 40) that a 'structural description' (strukturelle Beschreibung) be identified with a construction (Konstruktion, Lc. 23) of a grammatical statement on the basis of the rules, the construction being a sequence of grammatical statements ending with the statement in question (say, "S he came"). Wang's general approach consists in developing formal systems that correspond to a syntactic system for first-order predicate logic (in the sense of Carnap (1958) Ch. B); a grammar of a natural language is the system of rules of one of the formal systems; grammatical statements on the natural language are derived by those rules, which thus correspond to the rules of inference in a syntactic system for logic.33 It is then shown that the resulting grammars correspond (in a specific way and allowing for a number of deviations) to transformational grammars and structural descriptions as characterized in Chomsky (1965). Further developments of this approach are found in Wang (1971^). For structural descriptions in the sense of Wang it is indeed easier to see how they could be directly interpreted with respect to the object language of the grammar. On the other hand, the rules of a grammar cannot be interpreted in this way since they are rules for the derivation of grammatical statements. E.g., "NPu, VPv—»Suv" (1968) 24, if interpreted as a statement, might be read as: "From "NPu" and "VPv", "Suv" is derivable" (where "u" and "v" are variables whose values are expressions of the natural language, and juxtaposition of variables denotes concatenation). The rule in this example is easily replaced by a sentence of predicate logic: "(u)(v)(NPu & VPv->Suv)"—"for every u and v, if u is an NP and v a VP, then u concatenated with v is an S." This suggests that the language of first-order predicate logic might have been used in the first place: Instead of giving a formal system whose rules of derivation coincide with the rules of the grammar, the latter are introduced as axioms in an axiomatic theory formulated completely within first-order .predicate logic itself. The axioms may then be understood as statements on the object language. The usual rules of inference are used to derive theorems which can be interpreted as statements on the object language, such as "S (he comes)".
32
As constants denoting expressions of the natural language Wang uses those expressions autonymously (cf. Lc. 21). If this is taken literally, he allows only for written forms of natural languages. Cf. "N'man*" (or "Nman") /. c. 20, to be interpreted as meaning that "man" is an N. 33 Basically the same approach is used in the appendix of Smaby (1971), where the concept of a formal system as put forward in Smullyan (1961) is applied.
Grammars as theories: the case for axiomatic grammar (part I)
55
This possibility is indeed explored in Wang (1971£) (cf. also Wang (1972*) Section IV, (1973*)), where the rule quoted from Wang (1968) 24, is replaced as suggested above. Wang indicates a general way of replacing a context-free grammar by an axiom system formulated within first-order predicate logic (with identity) such that the axioms can be understood as statements on the object language. He formulates a special case of a theorem (proved in Wang (1971*)) stating that a sentence of the form Kx can be derived from the axioms (where K is a category symbol and ÷ a closed term to be interpreted as the name of a syntactic constituent) if ÷ is a terminal string derivable from K in the grammar ((1971£) Section 4).34 An attempt is made to replace transformational rules by axioms, too (Section 5). Emphasizing that his results agree with Chomsky's conception of a grammar as a theory (p. 273), Wang finally applies the concept of deductive-nomological explanation to grammars reformulated as axiom systems. A more detailed application of that concept is attempted in Wang (1972*) (to be discussed below, §4.4). In recent work Wang has been trying to combine ideas from Katz and Fodor (1963), Knuth (1968) and Montague (1970*, £), (1973) in order to arrive at a semantic component that could be combined with an axiomatic syntax (Wang (1972*) Section V; (1972£); (1973£)); this attempt does not provide an axiomatic reformulation of semantic symbolisms as developed within the generative framework. Wang's work may well represent the most promising attempt to understand the syntactic component of a classical TG as an interpreted axiomatic theory (by mechanically constructing an uninterpreted axiomatic theory from the grammar). Still, his account is problematic because of (A) and inadequate because of (B) and (C): A. Consider phrase structure rules such as "S—>NP VP". The sentence correlated with this rule—"(x)OO(NPx & VPy—>Sx"j)"—is equivalent with "NP + VP^S" (assuming an appropriate language of logic and a definition of " + " so that NP + VP = {^: (3*)(3j)(NPx & VPj & æ = *">)}. But why not correlate "(æ)(S^;—>(3x)(3j) (NP* & VPj & î = x"jO)", which would be equivalent with "S£ NP + VP"? Indeed, it is this latter sentence that is equivalent with the rule in the re-interpretation approach of Lieb (1967), and it is only this sentence that corresponds to the immediate constituent analysis which phrase structure grammars were meant to capture. Wang naively assumes that it is only a difference of formulation ("anders ausgedr ckt" (1971h) 277) whether constituents are combined or subdivided. But the two axiomatic theories that could be correlated with a phrase structure grammar differ in a non-trivial way. How are they related? Would it be possible to read " = " in each axiom of the two theories 34
In Section 3, it is shown that from any semi-Thue system a corresponding axiom system in first-order predicate logic can be constructed. The resulting system is, however, quite different from the systems obtained in Section 4, and it is left open how an interpretation with respect to natural languages might be effected.
56
Hans-Heinrich Lieb
("S = NP + VP"), which collapses the difference and gives us a grammar very much like the ones proposed in Cooper (1964) ? B. In many cases, the theory correlated with the complete syntactical component of a classical TG will be inadequate for lack of correspondence and may even be inconsistent. According to Chomsky (1965) 174, "rules of agreement clearly belong to the transformational component." We may surely assume a classical TG of English where the strings the boy is running and the boys are running (or analogous strings that would also serve the purpose of the demonstration) are terminal strings of the syntactic component and have (surface or shallow) phrase markers in which the boy is dominated by "NP" and are running by "VP". By Wang's method we should have "NP the boy39 and " VP are running as theorems of the correlated theory.35 Since agreement is handled by transformational rules, we may assume a base rule "S—»NP + VP" (or some other rule that would serve for the rest of the argument). The correlated theory then contains the axiom "(x)(y)(NPx & VPy—>Sx y)". We immediately obtain the theorem "S the boy are running*. On the other hand, the boy are running cannot be derived from "S" in the grammar, assuming an appropriate rule of agreement that blocks the derivation. Hence the theory is inadequate, for lack of correspondence with the grammar. It may even be the case that we have a theorem of the form — (Sx) whenever the derivation of from "S" by means of the grammar is transformationally blocked. In this case, we have "—(S the boy are running)", which makes the theory inconsistent. C. Wang's informal indications of how to interpret a theory correlated with a classical TG grammar are inadequate.36 Consider the theory correlated with the categorial subcomponent of a classical TG. Suppose the grammar is meant to be a grammar of a certain natural knguage L (in a sense where L is not identified with a set of strings derived from "S" by means of the grammar, to avoid circularity; cf. Lieb (1968£) §3). In the theory assockted with the categorial subcomponent, all category symbols —"S", "NP" etc. —are primitive (undefined) axiomatic constants. Following Wang's interpretative hints we would interpret them to refer to categories of the system of L by sentences such as "'S' denotes the set of sentences of L" etc. Assuming that denotation for category symbols can be understood as a finite two-place relation and thus be given by enumeration we are still left with the problem of making 35
"Should", because Wang makes only suggestions for treating the transformational component ((1971 b) §5). If those theorems cannot be obtained, the method is inadequate for that very reason. 36 Wang (1971 b) 277: "Ein formales System ist erst für die Linguistik von Interesse, wenn es eine linguistische Deutung besitzt. Wir sagen zum Beispiel im Fall der Regel wie NP—»T N, daß NP für eine Nominalphrase, N für ein Nomen und T für einen Artikel steht." Wang probably wants to say that "NP" denotes the set of noun phrases, etc., which is then implicitly assumed also for "NP" in the axioms of the correlated grammar. Wang's interpretative hints are extremely scanty.
Grammars as theories: the case for axiomatic grammar (part I)
57
sense of the expressions "sentences of, "noun phrases of, etc., and of the constant "L". Wang does not even see the problem. Thus, an interpretation along his lines is empty. Instead of supplementing the theory by an interpretation as above we may suggest to add axioms which are identities: "S = Sentence-of-L", "NP = Nounphrase-of-L", etc. This would bring out even more clearly the emptiness of Wang's interpretative hints: They do not contain anything about how to interpret the new theory. What are the consequences to be drawn from our investigation of the two approaches that were meant to overcome the difficulties created by thesis (1)? If either the re-interpretation or the correlation approach is successful, a generative grammar (or rather, the syntactic component of a classical TG) may be taken as equivalent, in a certain sense, with an interpreted axiomatic theory. I do not wish to maintain that the problems and inadequacies of Wang's method could not be overcome, just as my own re-interpretation program can be carried through in detail. My argument is that the two attempts are equally misguided: To make sense of a generative grammar (to make it say something about its intended subject matter) we either end up with an interpretative machinery inferior to the interpretation concept that has been or could be developed for axiomatic theories (Lieb), or we make use of an axiomatic theory anyhow (Wang). In order to make sense of an axiomatic theory of a language, we certainly do not need a generative grammar. If it can be shown that such a theory is not otherwise inferior to a corresponding generative grammar, there is no reason why one should bother with such a grammar in the first place. And indeed, this can be shown (§ 4.4, below). Hence, the attempts made by Wang and myself are misguided from the point of view of developing an optimal format for scientific grammars. Their main value consists in showing that the enormous amount of generative research could be made avaikble, or partly available, for use in a better framework. In turning to our main subject—grammars as axiomatic theories—we have to admit right away that the accepted view of interpreted axiomatic theories is not quite sufficient for our purposes. Thus, a general discussion of axiomatic theories is required. 2.
A framework for axiomatic theories (1): Formalized systems and abstract theories
2.1.
Introduction
The aim of this and the next section is an account of axiomatic theories— preliminary in many ways—that may be adequate for treating the general problems of axiomatic grammar writing. The two sections have grown out of an attempt to handle those problems; from a systematic point of view they are basic to the discussion of grammars in Part II. At the same time they try to develop a coherent picture that may be of value independently. At the beginning even elementary
58
Hans-Heinrich Lieb
points may be stated explicitly (also for the possible benefit of some readers); soon enough we shall have to modify customary conceptions. Such departures will not always be made explicit; the rich literature on axiomatization must to a large extent be presupposed. Our way of presentation will be relatively informal; a rigorous discussion of all relevant problems in the present context would be both impossible and out of place. Even so the two sections assume a wider scope than is customary in the field and may come closer to giving a synthetic picture than most existing work. They also contain some new results, especially on the combination of different theories into new ones (§ 3.4). An important feature of the framework is separation of formalized systems (logical calculi) (§ 2.2f) from axiomatic theories (§§ 2.4 and 3): An axiomatic theory is formulated by means of a calculus without being a mere extension of the latter, hence, another calculus. We also assume a richer structure for axiomatic theories than is customary: An axiomatic theory is not identified with a set of sentences of some calculus—quite apart from questions of 'interpretation'. The basic distinction usually made between 'uninterpreted' and 'interpreted' axiomatic theories is reconstructed as a distinction between 'abstract' and 'realized theories'. Two types of abstract axiomatic theories are distinguished (§ 2.4), one using axiomatic constants, the other axiomatic variables and an 'explicite predicate' (e.g. a set-theoretical predicate). The realized theories are subdivided into 'interpreted' and 'applied' (§3.1); the latter correspond to 'partially interpreted (empirical) theories' as discussed by Carnap and others. They are given special attention (§ 3.2f) because our later account of axiomatic grammars (Part II) is based on them. The realized theories present a more differentiated picture of 'interpretation' than is usually assumed; in particular, there is no 'partial interpretation' in a Carnapian sense although there may be partial interpretation in a literal sense in the case of an interpreted theory, and there must be such interpretation in the case of an applied one (the 'theoretical terms' are //«interpreted). Interpretation of the logical calculus used in a theory is discussed in connection with 'interpreted formalized systems' (§ 2.3) where a concept of pre-interpreted formalized system is introduced; it is such pre-interpreted systems that are used for formulating axiomatic theories.37 The subsections on formalized systems (2.2£) are in my own view less satisfactory than the ones on axiomatic theories. Not only did I have to be sketchy in my presentation of the ordinary conception of calculi (choosing, moreover, a fairly conservative version that may be too narrow to include recent developments in logic); I also had to introduce some new concepts without being able to develop their consequences. It may be worthwhile for the logician to pursue any nontrivial problem my account may present since it is completely motivated by the actual needs of linguistic theory construction. 37
Stegmüller has emphasized the ambiguity of "interpretation" as applied to theories ((1969/70) II, 340f); our conception takes his criticisms into account.
Grammars as theories: the case for axiomatic grammar (part I)
59
These needs led to elaboration of the kst subsection (§ 3.4) where a new conception of theory integration is developed. I distinguish two main types of joining given theories into new ones and formulate two theorems which are important for combining linguistic theories with each other and with non-linguistic ones. My account of formalized systems will at the beginning be based on Fraenkel et at. (1973)280—293. 2.2.
Formalized systems
In Fraenkel et al. (1973) a formal system** is considered as "an ordered quintuple of sets fulfilling certain requirements" (282): "(1) A set of primitive symbols, the (primitive) vocabulary, divided into various kinds, such as variables, constants and auxiliary symbols" (280). "(2) A set of terms as a subset of the set of expressions", an expression being any string (usually finite) .of symbols (281). "(3) A set of formulae as a subset of the set of expressions" (281). "(4) A set of axioms as a subset of the set of formulae" (282). (5) A set of "rules of inference according to which a formula is immediately derivable as conclusion" from an appropriate "set of formuke as premises" (282). If those sets satisfy certain requirements of 'effectiveness' (mentioned in their specification pp. 280—282), such as the terms being effectively specifiable, the formal system is said to be a logistic system (285). The notion of theorem is introduced in the usual way (282): A sequence (finite in the case of a logistic system) of one or more formulae is called a derivation from the set of premises if each formula in the sequence is either an axiom or,a member of For immediately derivable from a set of formulae preceding it in the sequence. The last formula is called derivablefrom . A derivation from the empty set of premises is called a proof of its last formula, hence, of each of its formulae. A formula is called provable or a formal theorem if there exists a proof of which it is the last formula.
In addition to formal systems, Fraenkel et al. introduce what they call 'formalized theories', each being (287) an ordered sixtuple of sets: a set of symbols, a set of terms, a set of formulae, a set of logical axioms, a set of rules of inference, and a set of valid sentences, with various relations obtaining between, and various conditions fulfilled by, these sets.
The essential features are exclusion of non-logical axioms and introduction of a special set of 'valid' sentences (I.e.): 8
On terminological equivalents, cf. op. cit. 281. For technical terms, variables, assumptions, and theorems introduced in §§2-4, cf. the index pp. 106fF(end of Part I of this essay). I gratefully acknowledge some valuable suggestions that were made to me for the present subsection by Franz von Kutschera (Universität Regensburg) and Wolfgang Stegmüller (Universität München).
60
Hans-Heinrich Lieb The exact extension of this term has to be defined from case to case, the only general condition which such a definition will have to fulfil being that the set of valid sentences should be closed with respect to derivability^ i. e. that every sentence derivable from valid sentences by the rules of inference should be valid itself . . .
It is maintained that any formal system is a formalized theory (287). There is an inconsistency in this, since a formal theory is a quintuple not a sixtuple. Obviously, the authors think of a sixtuple obtained from a formal system by taking the set of theorems as a sixth member. But even then not all formal systems are formalized theories because a formal system may contain non-logical axioms. Let us introduce "formalized system" as a new term to cover both formal systems and formalized theories. A formalized system, then, would be a sixtuple of a certain kind. This conception is still not adequate for our purposes: Fraenkel et al. "wished to dodge the problem of the status of definitions in formal systems" (284). We shall explicitly recognize definitions by including the set of definitions among the members of the formalized system; thus, a formalized system becomes an ordered septuple. A final change concerns the 'rules of inference'. By the conception that seems to be adopted in Fraenkel et al. (cf. also Carnap (1942) 22, 157), rules are sentences of some language that refer to formulas of the formalized system; thus, they are 'metalinguistic' entities. To avoid such entities, we take the relation of immediate derivability itself, i.e. a certain relation between formulas and sets of formulas, following the suggestions in Carnap (1958) § 26b. (The rules of inference themselves could also be reconstructed as entities that are no longer 'metalinguistic' in nature; cf. Hatcher (1968) 12.) We thus arrive at the following conception. A FORMALIZED SYSTEM (FS) is an ordered septuple of sets satisfying certain conditions (which are partly spelled out in Fraenkel et al. (1973) and cannot be reproduced here). The first member of the FS is called the set of symbols or the vocabulary of the FS; the second is the set of its terms (including the variables and certain constants from the vocabulary); the third the set of its formulas·; the fourth the set of its axioms. These four members are related to each other and to the expressions of the FS (strings of symbols of the FS) as previously indicated. The fifth member is the (possibly empty) set of definitions of the FS, all of them taken as formulas. The sixth member is a two-place relation between formulas and sets of formuks of the FS, called (immediate) derivability in the FS. The last member is the set of valid sentences of the FS, again a subset of the set of formulas, more specifically, a subset of the set of sentences·, /'. e. the closed formulas of F (the ones without 'free variables').39 The provable formulas on formal theorems of the FS are defined as before but 39
This assumes that the expression "free variable" can be defined for arbitrary FSs. Note that validity here is a non-semantic notion that must not be confused with truth.
Grammars as theories: the case for axiomatic grammar (part I)
61
the concept of derivation is changed by allowing not only axioms but definitions in the derivation sequence. All sentences that are provable and all sentences that are derivable from a set of valid sentences are valid sentences. A FORMAL SYSTEM (F1S) is now defined as an FS whose valid sentences are exactly the provable sentences or theorems of the FS. The terms "logistic system" and "formalized theory" could be re-introduced but we shall have no use for them.40 The following term refers to an important property of FSs: An FS is called axiomati^able if there is an F1S with the same set of valid sentences. Discovery of non-axiomatizable FSs was one of the main results of foundational research in mathematics. Since we explicitly recognize definitions, a few explanatory remarks may be in place. The definitions are formulas of a specific form that satisfy certain conditions relative to each other and relative to the other constitutive sets of the system. Thus, the definitions and axioms form disjoint sets; and there are four functions from the set of definitions: the defined term of, the deßning terms of, the definiendum of, and the definiens of. The first is a function into the set of constants of the system; the second into the set of non-empty sets whose elements are either constants or variables; the third and fourth are functions into the set whose elements are either terms or formulas. The defined term is the only constant occurring in the definiendum; it does not occur in the definiens. The defining terms are exactly the constants and variables occurring in the definiens. The defined terms of all the definitions are the defined terms of the system. The set of constants of the system can be partitioned into the set of defined and the set of undefined terms.41 Using "F", "F!", ... as variables ranging over septuples such as formalized systems, and "C", "Q", ... as variables over sets of constants of FSs, we now introduce the following notions. Let C be a set of constants of the FS F. The C-defined terms of F = the defined terms of those definitions of F among whose defining terms there is an element of C or a C-defined term of F. (This should be replaced by a proper recursive definition.) The C-terms of F= those terms of F in which an element of C or a C-defined term of F occurs;42 analogously, the C-formulas, C-definitions, C-axioms, C-theorems, valid C-sentences and C-expressions ofF. The following concept (which I have not found in the literature) is fundamental for later definitions: Let F be an FS and C a set of constants of F. F-MINUS-C = the septuple consisting of the following sets: the vocabulary of 40
Also, "formalized theory" will be used later (§ 2.4) for different purposes. For a systematic account of the conditions that must be satisfied by the set of definitions, see a sufficiently comprehensive standard introduction to logic such as Suppes (1957) Ch. 8; for a specialized treatise, cf. Essler (1970). For treating definitions not as formulas but as rules of inference, cf. e.g. Carnap (1942) 157f. For replacing definitions by axioms of a special type cf. Shoenfield 1967, §4.6. 42 Occurrence does, of course, allow for identity.
41
62
Hans-Heinrich Lieb
F-(C u the C-defined terms of F);43 the terms of F-the C-terms of F; and the five sets obtained analogously from the formulas, axioms, definitions, the derivability relation, and the valid sentences of F (in the case of derivability we take that subrelation of immediate derivability in F for which no first member and no element of a second member is a C-formula of F). We shall say that of two FSs F and Pl9 F is contained in Fj if there is a set C constants of F! such that F = Fj-minus-C. The union of F and Fa is the septuple obtained by uniting corresponding members of F and Fi9 except that the fourth member is to be the set of axioms of F and Fl that are not definitions Under certain conditions, here left unspecified, an FS may be called a FORMALIZED SYSTEM OF LOGIC (FS of logic, logical calculus), where "logic" may be replaced by more specific terms such as "predicate logic", "firstorder predicate logic" etc.44 It is assumed that in any logical calculus the set of constants (similarly, of axioms) may be partitioned into a non-empty set of logical constants and a set, possibly empty, of non-logcal constants** If there are no nonlogical constants, the calculus will be called pure, otherwise, non-pure?6 If F is an FS of logic, C will be called a harm/ess set of constants of F if C is a set of constants of F such that some logical axioms of F are not C-axioms of F. We make the following ASSUMPTION ON MINUS: For any FS of logic F and any harmless set C of constants of F, F-minus-C is an FS of logic.47 We finally extend the concept of FS as follows. Most standard logical calculi, at least the pure ones, have 'readings' in natural languages, i.e. some or all of their constants, terms, and formulas are systematically correlated with expressions of the natural language (thus, "All ÷ are P", "Every ÷ is P", "For every ÷, *· is P"9 "All ÷ are elements of P" etc. may all correspond to "(,*)/V). Frequently axiomatic theories are formulated not directly by means of an FS of logic but by using 43
Where "— " stands for difference and "u" for union of sets. For a recent introductory survey, see Rogers (1971); cf. also Hatcher (1968). 45 In a sense where this distinction does not depend on an interpretation of the calculus (cf. the -remarks in Carnap (1958) § 25a). Logical connectives and operator symbols like "3" are counted among the logical constants. We shall also assume that in any logical calculus the (possibly empty) set of mathematical constants can be specified by purely formal means. The mathematical constants are assumed to be non-logical. If our conception of FS of logic is to cover systems of natural deduction (cf. Hatcher (1968) § 1.6), we may have to allow for the set of logical axioms to be empty (unless we keep the tautologies as logical axioms). 'Dummy constants' (as introduced by Hatcher I.e., or 'ambiguous names' in Suppes 1957, § 4.3) must be excluded as constants. Concepts such as derivation and proof would have to be redefined. 46 Reflecting Church's "pure functional calculus" and "applied functional calculus", used in an analogous context ((1956) 173f). 47 A complete explication of "FS of logic" might show that the above assumption is too strong. In that case we would have to strengthen the concept of harmlessness by including additional requirements in its definition. 44
Grammars as theories: the case for axiomatic grammar (part I)
63
a reading of such a system. I allow for readings from the very beginning, suggesting an explication along the following lines. Let us consider regimented forms of natural languages as a special type of FS i8 Such expressions as "All ÷ are P" would belong to a regimented form of English. Assuming this concept, we define as follows.49 F is a natural language reading (NLR) of Fj iff: (1) F is an FS. (2) Ft is an FS of logic whose expressions are not expressions of a regimented part of a natural language. (3) There is an F2 and a function r such that: (a) F2 is a regimented form of a natural knguage. (b) The domain of r is the set of constants, terms, and formuks of Fj. (c) The values of r are sets of expressions of F2. (d) For all e, ej (e ö ej) in the domain of r, r(e) Ð r(ej) = ^.50 (e) The vocabulary (the constants, the variables, terms, formuks, axioms, definitions, valid sentences, respectively) of F = the set of all e such that, for some e ls e 6 r(ej) and ej e the vocabulary (the constants, variables, terms, formuks, axioms, definitions, valid sentences) of Fj. (f) Immediate derivability in F = «e, E>: (Be^ee r(ej) & <e!,{e2: (3e3)(e3 e Enr(e2))}> e Immediate derivability in Fj)}.51 If F is an NLR of Fj, any r that for some F2 satisfies condition (3) is called a rendering of F± as F. We now make the following ASSUMPTION ON NLRs: If F is an NLR of F!, F is an FS of logic (and if Fj is an FS of predicate logic etc., so is F). This assumption is, on the whole, unproblematic.52 48
In the light of Montague's work (esp. Montague 1970£), this seems justifiable. As of now, there does not seem to exist an explication of the concept of reading that would really do justice to the domain the explicatum should cover, i.e. the controlled use of natural languages (extended by symbolisms) in the rigorous formulation of theories. The following definition may still be simplistic, i.e. may apply to only some of the entities that should be covered. I believe, though, that the rest can be obtained by formal operations from readings in the defined sense. 50 With "n" for set-theoretical intersection and "0" for the empty set. 51 I.e. e is immediately derivable in F from E if there is an ej such that: ee rfo) and % is immediately derivable in Ft from the set of all e2 for which there is an element of E that is an element of r(e2). "<( ,X* is the ordered-pair notation and "{. ·: —} ' denotes the set of all .. such that—. "(3..)" is to be read as "for some ..". 52 Note that for each term or formula of Fj there may be a multiplicity of terms or formulas of F; e.g., for one axiom of Fj, say, "(x-)(Px—»Rx)", we might have "Every P is R", "All P are R", "For all x, if ÷ is Ñ, ÷ is R", etc. Also, r(e) may be empty, for a given r and e. Neither possibility should raise problems. It may, however, be advantageous not to introduce r and F2 by existential quantification but take an NLR as a triple
64
Hans-Heinrich Lieb
In discussing interpreted formalized systems we shall have to keep in mind that NLRs have been allowed as formalized systems. 2.3.
Interpreted formalized system s
The standard way of interpreting a formalized theory is characterized in Fraenkel et al. (1973) by stating how a 'first-order theory', i.e. a non-pure calculus of first-order predicate logic, is interpreted; this may be quoted in full (288£) because it also applies in the case of formalized systems: Such a theory is interpreted by providing the logical symbols with their usual signification, by fixing the universe (of discourse) U over which the variables range, and by assigning, through rules of designation, to each individual constant some member of U, to each unary predicate a certain subset of U and in general to each »-ary predicate a certain /7-ary relation whose field is a subset of U, finally to each w-ary operation symbol an #-ary function from ordered »-tuples of U to members of U. Calling the sequence consisting of U and these individuals, sets, relations, and functions ordered in a way similar to that of the constants to which they are assigned, a structure, semi-model, or (possible) realisation of T, rules of truth finally determine under what conditions a formula of T is true in a given structure, relative to a given valueassignment to its free variables if it contains such. (These rules determine, of course, also truth conditions for all the sentences.) An interpretation of a given theory T is called sound if under it all the valid sentences of T become true in the structure determined in the interpretation; this structure itself—and often also the assigning function—is then called a model of the theory.53
Accepting this account, how can we formally represent an interpreted formali^ed system? There is no answer to this question in Fraenkel et al. (1973). An interpreted system is specified by adding 'semantic rules' to an FS F. The rules are metalinguistic sentences on the 'meaning' or the 'reference' of expressions of F (the above 'rules of designation' belong to the semantic rules). We might take an interpreted system as a couple where D is a 'designation' relation for expressions of F (to be taken as a relation between expressions of F and 'extensions' such as classes and truth values, or between expressions and 'intensions' such as properties and propositions). To make this more precise, we first introduce an auxiliary concept. Let "B", "B!", ..., "C", "Q", ..., "D", "D^', ..., "U", "IV, ..., "v", "e", "f', "x" stand for any sets. Let F be an FS of logic. B is called a MINUS-C BASIS FOR if D is a two-place relation and C a harmless set of constants of F and B a triple «Ui> ..., Un>, v, f> whose first member 53
The above account only indicates the starting-point for model-theoretic work; a more detailed presentation would be impractical in the present context.
Grammars as theories: the case for axiomatic grammar (part I)
65
is a non-repetitious n-tupel of non-empty sets, with n = the number of different types of individual variables of F, such that: (1)
(2)
(3) (4)
f is a one-place function whose domain is the set of constants, terms, and formulas of F, and there is exactly one m > 1 such that the values of f are 'set-theoretical constructs in «Uj,..., Un>, {!,..., m})'.54 v is a one-place function whose domain is the (possibly empty) set of undefined non-logical constants of F-minus-C such that, for any argument e ofv,v(e)ef(e).55 The domain of v £ the domain of D £ the domain of f.56 For all <e, x> e D, ÷ e f(e), and if e 6 the domain of í, ÷ = v(e).
The term "interpreted formaUzed system" will not be defined; we only give necessary conditions for its application that are indispensable for subsequent discussion. From the very beginning we will relativize the notion to "interpreted except for a set of constants C (and all C-expressions)". This will eventually enable us to deal with so-called partial interpretation of axiomatic theories. ASSUMPTIONS ON MINUS-C INTERPRETED FORMALIZED SYSTEMS OF LOGIC (C-IFS of logic). If is a C-IFS of logic, then: (1) F is an FS of logic. (2) C is a set of constants of F. (3) D is a one-place function. (4) The domain of D is the set of constants (including logical ones), closed terms (i.e. terms without free variables), and closed formulas (i.e. sentences) of F-minus-C. (5) There is a minus-C basis for . An interpreted formalized system of kgic (IFS of logic) is a couple that is a C-IFS of logic, for some C. 54
f is the function assigning to each 'well-formed' expression of L the set of its 'possible values relative to ({Ui» · · ·> Un), {l,..., ðé}}'. Thus, in the above quotation from Fraenk'el et al. (1973) it is specified that the set of 'possible values'Ifor any individual constant is U, the set of 'possible values' of a unary predicate is the power set (the set of subsets) of U, etc.; the set of possible values for formulas could be taken as {1,2}, the set of 'truth-values'. The term "set-theoretical construct in" will not be defined, but set identity, power set, and Cartesian product may be taken as examples. Following recent advances in intensional logic, intensional entities such as properties of elements of Uj could also be treated via set-theoretical constructs in an appropriate tuple of sets that would include Uj-, e.g. we might add among others a set J of 'possible worlds' as a fourth member of B and take the values of f as set-theoretical constructs in «Uj,.. .,Un>, {!,.. .,m}, J>. Thus, if e is a formula, f(e) could be the set of propositions understood as functions from J into {!,.. .,m}, and properties etc. would be treated by also using the Uj. The above formulation is to be understood as generalized on these lines whenever necessary. We did not use a more general fomulation in order to leave open the specific form it might take. 55 /. e. v(e) is a 'possible value of e relative to «Uj,..., Un>, {l,..., m}>'. 56 With "c" for set inclusion. 5 TLIl/2
66
Hans-Heinrich Lieb
A minus-C interpreted formalized system of predicate logic (etc.) is a C-IFS of logic such that F is an FS of predicate logic (etc.)?1 An IPS of predicate logic (etc.) is a C-IFS of predicate logic (etc.), for some C. We further make the following ASSUMPTION 1 ON IFSs OF LOGIC: For all F, B, C, if there is a D such that is a minus-C IPS of logic (of predicate logic etc.) and B is a minus-C basis for , then there is exactly one D such that is a minus-C IPS of logic (of predicate logic etc.) and B is a minus-C basis for . This assumption is justified as follows. A minus-C basis for an interpreted formalized system of logic would be specified by rules of designation, including rules for .possible values (these would specify a class of functions to which f belongs). On this basis, the rules of evaluation (including the rules of truth) uniquely determine D if an interpretation of the logical constants of F is presupposed. The assumption is needed for some important considerations on axiomatic theories in § 3.2, Ad (W)y and § 3.3. An IPS of logic is called extensional if every sentence e of F is extensional in in the following sense: Let ej be in the domain of D and occur in e; let e2 be 'equivalent* to ^ in ;58 let e3 be the expression obtained from e by replacing an occurence of ^ in e by e2: Then e is equivalent to e3 in .59 In constructing an empirical theory we do not use a formalized system of logic that is completely unspecified semantically.60 Rather, we use systems that have been specified 'up to designation': The semantic rules have been formulated up to the point where indication of a set C of constants and an appropriate triple «Uj, ..., Un>, v, f> is sufficient to characterize a relation D that would make the system (minus-C) interpreted. Let us call such systems 'pre-interpreted* and treat them as pairs (where "Ä" stands for sets of relations D), as follows: is a PRE-INTERPRETED FORMALIZED SYSTEM OF LOGIC (Pre-IFS of logic) iff: (1) F is an FS of logic. (2) For any D 6 Ë: (a) The domain of D £Î the set of constants, terms, and formulas of F. (b) For any C, if there is a minus-C basis for , is a minus-C IPS of logic. (Again, "logic" as part of the defined term may be repkced by a more specific expression like "predicate logic" if the same is done in the definiens.) A Pre-IFS of logic
These definitions may have to be replaced by different ones that contain special conditions on D. The term in question should then be taken as undefined in the present essay, assuming, of course, that a C-IFS of predicate logic etc. is a C-IFS system of logic. 58 I.e. the sentence e2<-»e1 is true in (where "«->" is defined by material equivalence between sentences and "true in" is defined by reference to truth-values in the usual way). 59 This is strictly analogous to Carnap's definition of "extensional" for 'semantical systems' ((1956) 48). Note that this definition does not assume anything about the values of D (whether they are 'extensions' or 'intensions'). 60 This has been emphasized by Stegm ller ((1969/70) II, 303—305, 340).
Grammars as theories: the case for axiomatic grammar (part I)
67
61
k
if is an IPS of logic, is extensional (as defined above). Making a certain assumption on natural language readings we may obtain a Pre-IFS from Ft if Fj is an NLR of F and is a Pre-IFS. We first define, for any two-place relation D, any NLR Fx of F and any r that renders F as F! : The r-form of D = «e, x>: (3e1)«e1, x> 6 D & ej 6 domain of r & e 6 rfo))}. We now make the following ASSUMPTION ON NLRs AND PRE-IFSs: Let be a Pre-IFS of logic (predicate logic etc.), F1 an NLR of F, r a rendering of F as Fj, and D 6 Ä. Then, for all C, if there is a minus-C basis for , then is a Pre-IFS of logic (predicate logic etc.). From this we obtain the theorem: If <Ñ,Ä > is a Pre-IFS of logic (etc.), F1 an NLR of F, and r a rendering of F as Fj, then <Ñé,Ä *> is a Pre-IFS of logic (etc.), where Ä * = {D: (30! e Ä & D = the r-form of Dx)} (proof omitted). If (F,A y is a Pre-IFS of logic and C a set of constants of F, we understand by EXCLUDING C that couple for which F! = F-minus-C andAj = {Dx : (3D)(D e Ä & Dj = D-the CF-part of D)}, where the CF-partofD is defined as follows, for any FS F, set C of constants of F, and any relation D: that subrelation of D whose first-place members are C-expressions of F. Making an additional assumption to be formulated immediately, we obtain the following theorem (proof omitted): If is a Pre-IFS of logic and C a harmless set of constants of F, excluding C is a Pre-IFS of logic. For use in the assumption we first define, for any FS of logic F, any set C of constants of F, and any relation D: the C-reduction of % Z>> = < F-minus-C, D-the CF-part of D>. ASSUMPTION 2 ON IFSs OF LOGIC then reads: Let F be an FS of logic and C a set of constants of F. If existence of a minus-Q basis for is sufficient for to be a Q-IFS of logic, then existence of a minus- Q basis for the C-reduction of is sufficient for the C-reduction of (F, D^ to be a Q-IFS of logic.—The following three concepts are all based on exclusion. ^Fi> A I ) is contained in a Pre-IFS of logic if, for some set C of constants ofF, ^F ls Á÷) = excluding C. If is a Pre-IFS of logic, the underlying logic of < F, Ä > = excluding C*(F), where C*(F) = the set of undefined non-logical constants of F. A Pre-IFS <^F, Ä) is logically compatible with another, (Fi> Ë^, if the underlying logic of is contained in the underlying logic of A!> ; every expression e of F that is an expression of Fj 'has the same logical status' in F and PI ; and, for all e, x, Xj, D e A and ¼1 6Ä ,, if <e, x> 6 D and <e, x^ e D l9 ÷ = x^62 61
This concept will be needed for discussion in § 4.2. "Has the same logical status in... and..." cannot be defined by the concepts introduced in § 2.2. Intuitively, we must prevent cases like a constant e being an individual constant in F and a two-place predicate in Fj; or a logical constant in F and a non-logical one in Fj. For explicating "logical status" we would have to specify the subclassifications assumed for the vocabulary, terms, and formulas of a given FS of logic. This goes beyond the limits of the present essay.
6
68
Hans-Heinrich Lieb
The following concept and assumption will be needed for the Conflation Theorem in § 3.4, below: If and are Pre-IFSs, the amalgamation of and is the couple or is the vocabulary of F, the terms are the terms of F, etc. Finally, a minus-C interpreted formal system of logic is a C-IFS of logic
Abstract axiomatic theories
For developing our notion of abstract ('uninterpreted') axiomatic theory (AT) we may start from Carnap's account of 'axiom systems' ((1958) § 42): For the formulation of an AS we need to choose or construct a language L, the so-called basic language of the AS. Usually this basic language contains only logical signs. The axioms and theorems of the AS contain certain constants not occurring in language L, called the axiomatic constants of the AS. Some of them are given without definitions; they are called the axiomatic primitive constants of the AS. All other axiomatic constants are introduced'by definitions on the basis of the primitives. The language L' obtained from the basic language L by adding the axiomatic constants is called the axiomatic language.
By "axiom system" (AS) Carnap seems to understand the set of all axioms, theorems, and definitions.63 For relating an AT to an AS in the sense of Carnap, different proposals may be considered. First, we could simply identify AT and AS. But this is inadequate because ATs with different sets of axioms could still be identical (there may be more than one way to 'axiomatize' the same set of sentences). Next we could identify the AT with the 'axiomatic language' of the AS. The axiomatic language would be taken as a formal system of logic in the sense of §2.2, which agrees well with Carnap's conception ((1958) §§26b, 42a, b). Choosing the axiomatic language as the AT would correspond to the approach in Fraenkel et al. (1973) (cf. e.g. 283). However, the difference made by Carnap between those constants and sentences of the axiomatic language which belong to the AS and those which don't is not formally reconstructed. This may seem unimportant if the 'basic language' of the AS contains only logical constants; in 63
In Carnap (1960), "Axiomensystem" refers to the set of axioms only.
Grammars as theories: the case for axiomatic grammar (part I)
69
this case, it can be identified with the axiomatic language minus its non-logical constants, in the technical sense of "minus" (§ 2.2). For the conception to l>e developed here it will be important, though, that some non-logical constants of the 'language' of the AT may not be axiomatic constants of the AT. Therefore, in defining the notion of an AT, the axioms, definitions, and constants of the AT should be given independent status. Also, the 'language' of the AT should not be a formalized system but a pre-interpreted formalized system, as noted above (§ 2.3). The 'basic language' can be identified with the axiomatic knguage excluding the axiomatic constants, in the technical sense of "excluding" (§ 2.3)—we start from the axiomatic rather than the basic language.64 This leads us to a third proposal: An AT is considered as a quadruple consisting of the axiomatic language, the axioms, definitions, and axiomatic constants of an AS in the sense of Carnap.65 We will have to distinguish two forms of ATs only one of which is closely related to ASs. In order to have a second handy term for theories of this form we will re-introduce "axiom system" as a technical term into our framework. For the following definitions, "L", "IV, ...are introduced as variables over pairs-(F, Ä>; "S", "S^',... stand for arbitrary sets of formulas of formalized systems; "C", "Q" for any sets of elements of the vocabulary of any formalized system. An ABSTRACT AXIOMATIC THEORY OF TYPE 1 (AT1) or AXIOM SYSTEM (AS) is an ordered quadruple such that: (1) L is a pre-interpreted formalized system of logic. (2) C is a set of undefined constants of L. (3) S is the (non-empty) set of C-axioms of L. (4) St is the set of C-definitions of L. (5) Every element of C occurs in an element of S or Si. Note that C is non-empty because of (3), and that Sj may be empty. Both the C-axioms and C-definitions may contain non-logical constants of L that a*e neither in C nor C-defmed. (Cf. the definitions in §2.2, p. 61.) By (5) 'unused' constants are excluded, which would have been too stringent a requirement on formalized or formal systems in general. S and Sj are disjoint by (1), (3), and (4). Using "T", "Tj",... to stand for quadruples : L is the axiomatic language of T; C the set of axiomatic primitives of T; S the set of axioms, and S1 64
The appealing idea that the axiomatic language is obtained from the basic language by 'adding constants' seems to cause insuperable difficulties when a precise formulation is attempted. Similarly, Stegm ller ((1969/70) II, 305) notes for the Observation language* and the 'theoretical language* that they can be specified only as parts of the total language of a theory. 65 A similar approach is occasionally found in the literature; thus, in Mates (1965) Ch. 11, § 1, a first-order theory is identified with a pair <Ä, Ã× where Ä is the set of non-logical constants occurring in the formulas of the set à (the set of theorems); similarly, Rogers (1971)54.
70
Hans-Heinrich Lieb
the set of definitions of T. The defined constants of T = the C-defined terms of the axiomatic language of T. The axiomatic constants of T = the axiomatic primitives and defined constants of T. The axiomatic terms of T = the C-terms of the axiomatic language of T (i.e. any term in which an axiomatic constant occurs) and the axiomatic constants of T. The basic language of T = the axiomatic language of T excluding C. The non-logical constants ot T = the non-logical constants of the axiomatic language of T. The mathematical constants of T = the mathematical constants of the axiomatic language of T (cf. above, fn. 45). The basic non-logical (non-mathematical) constants of T = the non-logical (non-mathematical) constants of the basic language of T. The defined (undefined) terms of T = the defined (undefined) terms of the axiomatic language of T. The theorems of T = the sentences of the axiomatic language of T that are not axioms or definitions of the axiomatic language of T; are derivable in the axiomatic language of T from a set of sentences each element of which is an axiom of T, a definition of T, or a valid sentence of the basic language of T; and are not derivable from a set of valid sentences of the basic language of T. The sentences of T = the axioms, definitions and theorems of T. It is unusual that we should have defined "theorem" by using validity rather than just derivability. This enables us to speak of axiomatic theories even if the basic language (or rather, its first component) is non-axiomatizable in the sense of § 2.2, p. 61, which would make the axiomatic language non-axiomatizable, too (note that the latter has not been required to be a formal Pre-IFS of logic). In this way we make sure that all true mathematical sentences can actually be used in an empirical theory in the derivation of theorems; if we had required that only formal theorems of the axiomatic language of T be allowed in the derivation of theorems of T, we would run into trouble because of non-axiomatizability results obtained in the foundations of mathematics.66 Our conception of "theorem" also allows us to introduce a concept of axiomatizability for theories that does not presuppose axiomatizability of the language of the theory. We first define: An abstract formalized theory of type 1 (AFT1) is an ordered quadruple
Grammars as theories: the case for axiomatic grammar (part I)
71
The second type of abstract axiomatic theories is obtained by defining an 'explicite predicate' (Carnap (1958) § 42d); best-known examples are 'axiomatizations within set theory', where a 'set-theoretical predicate' is defined (Suppes (1957) Ch. 12, is one of the best introductions). By an ABSTRACT AXIOMATIC THEORY OF TYPE 2 (AT2) we understand a couple such that: (1) L is a pre-interpreted formalized system of logic that contains a Pre-IFS of predicate logic. (2) e is a definition of L. (3) The defined term of e is a predicate of L. (4) The definiendum of e is an atomic formula of L. (5) All {the defined term of e}-defmitions of L except e are conditional definitions whose antecedent is the definiendum of e. (6) The axioms of L = the axioms of: L excluding {the defined term of e}. We have the following auxiliary terms for an AT2 (L, e^: L is the language^ e the (main) definition; and the defined term of e is the predicate of . The axioms of are the conjuncts forming the definiens of e if the definiens is a conjunction, otherwise, the definiens is the only axiom. The definitions of are the {the defined term of e}-definitions of L. The theorems of e> are the universal implications having the definiendum of e as their antecedent which are derivable in L from a set each element of which is a definition of or a valid sentence of L and which are not derivable in L from a set of sentences of L that does not contain a definition of are the non-logical (non-mathematical) constants of L (there may be such although there are no 'axiomatic constants'). The following two concepts apply to both types of ATs in an analogous way: An AT is formulated in the language of predicate logic if the axiomatic language (the language) of the AT is a Pre-IFS of predicate logic. The AT is formulated in the language of set theory if there is an AT1 that is a set theory and whose axiomatic language = the axiomatic language (the language) of the AT excluding the set of undefined non-logical constants of the AT that are not axiomatic primitives of the set theory.68 AT*s and AT2s can be related as follows (cf. also Carnap (1958) § 42d). Assume an AT1 T such that: The axiomatic language of T is a pre-interpreted formal system of predicate logic; the axioms, definitions and axiomatic primitives of T are finite in number; and all definitions satisfy the criterion of eliminability for the defined term.69 Transform T as follows: First eliminate all defined axiomatic constants from the axioms on the basis of their definitions. In the sentences so obtained, replace the n axiomatic primitives, in a standard way, by n distinct variables of the axiomatic language of T. Let e be a non-repetitious conjunction of the new formulas: e is an open formula of the basic language of T. Now consider any Lj such that, for some et and e2: ej is an ç-place predicate of LJ ; e2 is a definition 67
Valid sentences are included for the same reason as above in the case of ATAs. The concepts of abstract theory and axiomatizability could also be generalized. 68 The qualification is not needed if those primitives are counted as logical. 69 Some but not all of these restrictions could be removed.
72
Hans-Heinrich Lieb
of LI; the definiendum of ^ — .the open atomic formula consisting of ej and a non-repetitious ç-tuple of the n free variables of e (thus, et = the defined term of €2); the definiens of e2 = e; the axioms of Lj = the axioms of the basic language of T; and the basic language of T = Lj excluding {ej}. Now consider any expression e3 that has been obtained from a theorem of T by treating it in the same way as the axioms of T: Obviously, the universal implication whose antecedent is the definiendum of e2 and whose consequent is e3, is provable in Lj. T is, in a sense, represented by
3.
A framework for axiomatic theories (2): Realized theories and theory integration
3.1.
Realised axiomatic theories: Interpreted and applied
As mentioned in § 2.1, the realized axiomatic theories (RTs) are subdivided into interpreted and applied. The former are obtained from AT1 s as would be expected, by 'interpretation', i.e. by adding a triple (^Ui>..., UnX v, f^ (where n is the number of different types of individual variables of the axiomatic language). Certain constants may be left uninterpreted, for reasons that will become clear immediately. A MINUS-C INTERPRETED AXIOMATIC THEORY (C-INT) is a couple such that: (1) T is an AT1. (2) Every element of C is an undefined non-logical, non-mathematical constant of T. (3) There is a D such that, if <(F, A>=the axiomatic language of T, then O€\ and B is a minus-C basis for .71 An interpreted axiomatic theory (INT) is a couple that is. a C-INT, for some C. A completely interpreted axiomatic theory (cINT) is a 0-INT (â = the empty set). An incompletely interpreted axiomatic theory (ilNT) is an INT that is nota^-INT. The concept of minus-C interpreted formalized theory (C-INTF) may be defined in strict analogy to C-INT, by using AFT1 s and making obvious substitutions. 70 The above account has to be modified when an AT1 is related to a theory *axiomatized within set theory* in the customary sense: The predicate of the latter usually has m places more than would be expected, by introduction an m-tupel of sets such that the i-th set corresponds, in the AT1, to the set of possible values for individual variables of type i. 71 As is a Pre-IFS, it follows from the relevant definitions in § 2.3 that is an interpreted formalized system of logic.
Grammars as theories: the case for axiomatic grammars (part I)
73
One would expect a subdivision of RTs corresponding to the distinction between AT1 s and AT2s. On the usual conception an AT2 is interpreted' by adding a sentence to the effect that a certain ç-tuple of entities satisfies the predicate of the theory. Technically, this corresponds to constructing an abstract axiomatic theory (of type 1) from the AT2 by introducing an axiom and axiomatic primitives. From this theory we may obtain an interpreted theory in the previous sense. Hence, instead of introducing a new type of RT tp correspond to AT2s, we characterize a special type of AT1 s: An interpreted axiomatic theory with predicate (INTP) is an INT for which there is an AT2 and an ej such that: (1) The basic language of T = L. (2) ej is an axiom of T. (3) et is an atomic sentence such that: (a) The predicate of et = the predicate of . (b) The argument expression of ej consists of the axiomatic primitives of T.72 The concept of an incompletely interpreted axiomatic theory was introduced to deal with so-called 'partial interpretation' of empirical theories. It should be inadequate, though, simply to equate a 'partially interpreted' empirical theory with an iINT: The essential distinction in the former between the 'theoretical' part and the 'correspondence' part (relating 'theoretical terms' and Observation terms') would not be reconstructed. A 'partially interpreted' empirical theory should have a richer structure than an INT. This leads us to introduce applied theories as a second type of realized axiomatic theories. The following definition will be explained in the next subsection (3.2): An APPLIED AXIOMATIC THEORY (APT) is an ordered quadruple such that: (1) T is an AT1. (2) Tj is an AT1. (3) The axiomatic primitives of T cz the axiomatic primitives of Tt. (4) The axiomatic primitives of Tj -c the constants of Tj that are non-logical and non-mathematical. (5) The axiomatic language of T= the axiomatic language of Tj exluding the set of axiomatic primitives of Tj that are not axiomatic constants of T. (6) The definitions of T! = the definitions of T. (7) For every axiom e of Tj that is not an axiom of T there is an el5 e2, e3 such that: (a) ej is an axiomatic constant of ¾ß that is not an axiomatic constant of T. (b) ej occurs in e. (c) e2 is an axiom of ¾1. (d) ej occurs in 62. (e) €3 is an axiomatic term of T. (f) 63 occurs in %?* (8) L is a pre-interpreted formalized system of logic. (9) The axiomatic language of Tj is contained in L. (10) If is the axiomatic language of Tt and C the set of axiomatic primitives 72
On this account AT1 s are more basic than AT2s in the following respect: Even the concept of an INTP presupposes the concept of AT1; the attempt to interpret an AT2 automatically creates an AT1 that is then interpreted.—Until Sneed (1971), interpretation problems of AT2s had been given little consideration. The above conception corresponds most closely to the first of several approaches discussed by Sneed. (His book came to my attention too late to be carefully considered in the present paragraph.) 73 In addition, e must satisfy certain formal conditions, such as not being the conjunction of a sentence of T and a sentence of Tj that is not a sentence of T. For relevant discussion, see Stegm ller (1969/70) II, Ch. V, esp. § 9.
74
Hans-Heinrich Lieb
of T and undefined basic non-logical and non-mathematical constants of Tj, then for some D 6 , is a minus-C basis for . Corresponding to an INTP we have an applied axiomatic theory with predicate (APTP), i.e. an APT for which there is an AT2 and an ej such that: (1) The basic language of T = LJ. (2) ea is an axiom of T. (3) ej is an atomic sentence such that: (a) The predicate of ej = the predicate of (Li> e)>. (b) The argument expression of ej consists of the axiomatic primitives of T. Notice that we have obtained a complete parallelism between INTs and APTs: In either form there are theories with and without predicate. The realised axiomatic theories may now be identified with the INTs and APTs. In defining "applied theory" we were trying to reconstruct the concept of 'partially interpreted (empirical) theory' as developed by Carnap and others and exhaustively discussed in Stegmüller (1969/70), II, 181—437,74 There are, roughly, the following correspondences. T: the theoretical part of the theory. Tj: the theoretical part supplemented by correspondence rules. L: the total language of the theory, possibly including sentences that are not expressions of the theory. B: an 'interpretation* specifying domains of individuals for all individual variables of the theory and assigning values only to the non-theoretical primitives of the theory. The axiomatic constants of T correspond to the 'theoretical terms', the axioms of T to the 'theoretical axioms'; the axiomatic constants of Tx that are 74
For further discussion see, for instance, Suppe (1971) (who does not really take into account Carnap's later views), and Lewis (1970); partly relevant also Sneed (1971). Lewis presents an important proposal that takes Carnap's final views as a starting-point but "has the advantage that it permits theoretical terms to be fully interpreted and explicitly defined" (439 f). It consists, roughly, in positing 'meaning postulates' to the effect that the theoretical terms name the components of the unique realization of the theory (the model that satisfies the conjunction of its axioms) in case there is such a realization, and do not name anything otherwise (434). I find this quite unsatisfactory for a reason which I can only hint at: If we allow the theoretical terms to be first members of a naming relation, then we should also allow entities to be named by them regardless of whether those entities form a model that satisfies the axioms: We should admit a unique realization of the theory and an interpretation of its theoretical terms such that the latter do not name the components of the former. In other words, I am arguing that Lewis' 'meaning postulates' are inadequate for explicating "interpretation" for theoretical terms as intended by Lewis, vi%. in a 'realistic' not an 'instrumentalist' way. There are other problems with Lewis' proposal (thus, his crucial 'empirical hypothesis' p. 439 does seem to lead to a negative result, at least in my case) but it has to be admitted that accepting his proposal should simplify some of the problems raised by our own treatment (cf. below, § 3.3). In Stegmüller (1973), which appeared only after the present essay was finished, a new conception of theories is proposed that is a modified and expanded version of Sneed's approach (1971). As I shall show in another paper, most of the criticisms that Stegmüller directs against the 'received view' of theories do not apply to the conception of realized axiomatic theories as developed above, and valuable insights gained by Sneed and Stegmüller can be incorporated into my own framework.
Grammars as theories: the case for axiomatic grammar (part I)
75
not constants of T correspond to the Observation terms', the axioms of Tj that are not axioms of T to the Correspondence rules'. The axiomatic language of T corresponds to the 'theoretical language'; L excluding the non-logical constants of T to the Observation language'. Before turning to some more detailed comments on applied theories, we introduce the auxiliary terminology that will be needed for APTs later on in this paper. Let "È", "©j",... stand for arbitrary quadruples . If è = is an APT, T will be called the core, Tj the applied core, L the total language, and B the semantic basis of È. The semantic basis (an ordered triple) consists of the domain sequence, the primitive interpretation, and the possiblevalue function of 0. In connection with T, we have the core language of è = the axiomatic language of T; the core constants of è (divided into axiomatic and basic) — the non-logical constants of T. The core axioms, definitions, theorems, and sentences of è are the corresponding sets of T. The application language of è= the total language excluding the core constants of È. The axiomatic language of Tx is called the applied-core language of è . The axiomatic constants, axiomatic terms, axioms, definitions, theorems, ana sentences of è are the corresponding sets of Tj. Those axiomatic primitives of Tj which are not axiomatic constants of T will be called application constants of è and those axioms of T! which are not axioms of T, application axioms of È. The triple <$!, L1} B> such that Sj is the set of application axioms, Lx the application language and  the semantic basis of è is the application of È. The concept of formalized theory can be extended so as to allow for applied theories, as follows. An applied formalized theory (APFT) is a quadruple such that: (1) T is an AFT1. (2) Tx is an AFT1. (3) The second component of T c the second component of Tj. (4) The second component of T! £Î the constants of the first component of Tj that are non-logical and non-mathematical. (5) The first component of T = the first component of Tj excluding the difference of the second components of T! and T. (6) The fourth component of Tj = the fourth component of T. (7) L is a Pre-IFS of logic. (8) The first component of Tj. is contained in L. (9) If is the first component of Tj and G = the second component of Tu the set of undefined non-logical and non-mathematical constants of T! that are not elements of the second component of Tj, then for some D e Ä,  is a minus-C basis for . The concepts of axiomatization and axiomatizability can now also be extended: è is an axiomatization of È! iff: (1) è is an APT. (2) èé is an APFT. (3) The core of È is a (type-1) axiomatization of the first component of 0j. (4) The applied core of È is a (type-1) axiomatization of the second component of È! . (5) The total language of È = the third component of èé · (6) The semantic basis of è = the fourth component of è j. è is axiomati^able if there is an axiomatization of è.
76
Hans-Heinrich Lieb
We also generalize the concept of PART as defined for AT1 s and AFT1 s (5 2.4, p. 70). Let T be any AT1 or AFT1 and è and è ß any APTs or APFTs. Ô is a part of È if Ô is a part of the second component of è; È is a part of Ô if the first component of è is a part of T; and è is a part of S1 if the first component of è is a part of the second component of è^.75 The (homogeneous) part relations are obviously reflexive; they are also antisymmetric and transitive if the same is assumed for the relation of containing in § 2.3, p. 67. An APT or APFT è is FORMULATED IN THE LANGUAGE OF PREDICATE LOGIC if the third component of è is a Pre-IFS of predicate logic, and FORMULATED IN THE LANGUAGE OF SET THEORY if there is an AT1 that is a set theory and whose axiomatic language = the third component of è excluding the set of undefined non-logical and non-mathematical constants of the third component of è that are not axiomatic primitives of the set theory. We now turn to a more detailed discussion of APTs, beginning by comments on the definition.
3.2.
On the definition of "applied axiomatic theory"
Ëá (1) to (6). The set of primitives of Ô is non-empty by (1) and the definition of "AT1", (3). As it is a proper subset of the set of primitives of T x , the set of axioms of Ô is also properly included in the set of axioms of T! , by (5), (6), and the definition of "AT1", (5); in other words, there are 'correspondence rules', and the Observation terms' occurring in them are not-defined in Tj. It also follows that all theorems of Ô are theorems of Tj. In (4) we explicitly recognize the possibility of non-logical constants other than the axiomatic ones and exclude purely logical or mathematical theories from the applied ones. Ad (7). This is a liberalization of the usual requirement on 'correspondence rules' which can be partly restated as: Every axiom of Tx that is not an axiom of Ô contains an axiomatic constant of Ô and an axiomatic constant of Tx that is not a constant of T. The liberalization is motivated by the fact that we allow for non-logical constants of the basic language. An application axiom may relate an application term and a basic non-logical constant of T on the condition that the application term also occurs in an application axiom where it is related to an axiomatic term of T. By adding condition (7) to the previous ones we are able to prove the following theorem: The set of application axioms of è (the 'cor75
For the last definition a stronger requirement could also be considered: È is a part of 0j iff: (1) The first component of È is part of the first component of 0j. (2) The second component of È is part of the second component of 0le (3) The third component of È is contained in the third component of 0j. (4) The fourth component «U^,.. -,U ln >, Vj,^) of È is related to the fourth component «U 21> .. -»U^Xv^f^) of È1 as follows: U2i — Ujj, for all i = 1,.. .,n and some j = 1,.. .,m; Vj ^ v2; and fj ^ f 2 .
Grammars as theories: the case for axiomatic grammar (part I)
77
respondence rules') = the set of C*-axioms of the applied-core language of è (the axiomatic language of Tj), where C* = the application constants of è (the Observation terms' occurring in axioms of È). Ad (8) and (9). It has been emphasized by Stegm ller (1969/70), II, 305 that the 'theoretical language' and the Observation language' can be distinguished only within the total language of the theory (the latter also contains 'mixed sentences' in which both theoretical and observation terms occur, most notably the correspondence rules). Correspondingly, the total language L is introduced independently (8), and two abstract axiomatic theories are distinguished (1, 2) such that the axiomatic language of one is contained in the other (5) whose language in turn is contained in L (9). Ad (10). As T! is an AT1 by (2) and C a set of undefined non-logical, non-mathematical constants of Tj by (3), (4), and (10), it follows from (10) and the definition of "minus-C interpreted theory" that <Ô1? Â> is a minus-C interpreted axiomatic theory. Moreover, there is a Bj and Q such that be the axiomatic language of Tj, and is a minus-Q IPS of logic (obviously, Q = the set of axiomatic primitives of Tu the set of undefined basic non-logical, non-mathematical constants of Tj). Hence, by Assumption 4 on C-IFSs, the domain of D* = the set of constants and closed terms and formulas of F-minus-Q. Let D1 = D* — the QF-part of D*. By (5) and the definition of "excluding" in § 2.3, ¿1€^1 and Fx = F-minus-Q. It now follows that B! is a minus-Q basis for (F-minus-Q, I>i> (cf. the definition of "basis" in § 2.3). Hence, by the definition of "C-INT", is a minus-C! interpreted axiomatic theory. The axiomatic language of Tj is a Pre-IFS of logic (§2.4, definition of "AT1"); hence, it follows from (10), the definition of "Pre-IFS" (2b) and Assumption 1 on IFSs of logic (§ 2.3, p. 66) that there is exactly one D such that is a C-IFS of logic and  a C-basis for . Let us call this D the B-interpretation of Tit or, for short, int(B, T±) (remember that this is a function with expressions of T! as arguments). All expressions of Tt that contain an axiomatic constant of T are not in the domain of int(B, Tj): they are not 'interpreted' at all. This amounts to saying, in customary terms, that no 'theoretical term' and no expression in which it occurs
78
Hans-Heinrich Lieb
receives an interpretation; only the Observational terms* and corresponding terms and formulas do. I am here taking the position that there is no 'partial interpretation* of theoretical terms by supplying 'correspondence rules': Those very rules remain uninterpreted if they contain theoretical terms.76 Moreover there may be basic non-logical and non-mathematical constants of Tj (that are neither axiomatic constants of Tj nor of T). These, too, would not be in the domain of int(B, Tt) nor would the expressions in which they occur. This view raises a number of problems which may throw into doubt our whole concept of applied axiomatic theory. 3.3.
Uninterpreted constants in applied theories
We shall formulate and discuss four objections to condition (10) in the definition of "applied axiomatic theory", which are all concerned with the problems raised by allowing uninterpreted constants. First objection to (10).
Leaving the axiomatic constants of T uninterpreted has the consequence that no sentence of the axiomatic language of T^ containing such constants is assigned a truth-value on the basis of int(B, Tx) (understood as in the preceding subsection); for this the sentence would have to be in the domain of int(B, Tj). Thus, no theorem or axiom of Tj with such constants can be said to be either true or false: It cannot be understood as a statement. I suggest the following solution for the special case that the axioms of T! are finite in number. Let T2 be any AT1 that is a part of Tj (in the sense of §2.4, p. 70; T2 may be identical with Tj). Consider the Ramsey sentence of T^77 Assuming that the axiomatic language of Tj has bindable variables for each nonlogical constant, the Ramsey sentence of T2 (RT2) is obviously a sentence of the axiomatic language of Tj excluding the set of basic non-logical and non-mathematical constants of Tx and axiomatic constants of T; thus, RT2 is in the domain of int(B, Ti) and is assigned a truth-value. Assuming, in particular, that the values of int(B, Tj) are classes etc. arid truth-values, the truth-value of RT2 = int(B, Tj) (RT2), /. e. a certain positive integer. For all sentences of the axiomatic language of Tx that are in the domain of int(B, Tt) we can define the concept of ÂÔ×-truth in the usual way by means of truth-values. The concept of BTX-truth is then extended to theorems and axioms of T! as follows: A theorem or axiom of Tj is BTj-true if it is an axiom or theorem of a part of Tj whose Ramsey sentence is BT^true. The two definitions can be combined in the following way: 76
For a similar view on 'partial interpretation' cf. Kutschera (1972), I, §3.1 f. This is the (more accurately: a) sentence obtained by eliminating all defined non-logical, non-mathematical constants from the axioms of T2; forming a non-repetitious conjunction of the axioms; replacing all non-logical, non-mathematical constants of T occurring in the conjunction by appropriate variables; and by preposing existential quantifiers which bind those variables. On the special role of the Ramsey sentence, cf. Carnap (1963), (1966), and the detailed discussion in Stegm ller (1969/70), II, 400—437.
77
Grammars as theories: the case for axiomatic grammar (part I)
79
Let e be any sentence of the axiomatic language of Tj. e is BT^-true iff either (a) or (b): (a) e is in the domain of mt(B, 1^), and int(B, T1)(e) = l; (b) e is an axiom or theorem of a part T2 °f ºÉ such that RT2 is BTj-true. e is ÂÔé-false iff either (a) or (b): (a) e is in the domain of int(B, 1^), and e is not BT^-true. (b) e is an axiom or theorem of 1^ and not BTj-true. As all axioms and theorems of T are axioms and theorems of T ls the two concepts apply in all important cases. Given the special status of the Ramsey sentence (which may be taken as representing the factual content of the theory) the above solution should be acceptable. True enough, it contains a number of limiting assumptions but these are easily removed except for the assumption of a finite number of axioms. That requirement may be dropped only if the concept of Ramsey sentence is generalized to at least sentence schemata (unless we allow infinitely long expressions in the axiomatic language of Tx). Assuming the above solution we may feel less reluctant to accept the following hypothesis: No sufficiently complex theory in the empirical sciences can be developed or reconstructed as a minus-C INT so that some 'theoretical terms' are not in C. This is analogous to what Stegm ller calls the Braithwaite-Ramsey Vermutung ((1969/70) II, Ch. IV, § 5) and can be justified by similar arguments. It is this hypothesis which makes us choose APTs rather than INTs as a format for linguistic theories.78 Assuming that the first objection to (10) has been refuted there are still other properties of (10) that may seem objectionable. Second objection to (10).
There may be non-logical constants of the total language L that are not constants of the axiomatic language of Tj and thus remain uninterpreted. This is inadequate because it destroys the correspondence between the application language of the APT È and a customary Observation language'. Anybody who is convinced by this argument may simply replace "the axiomatic language of Tj" in (10) by "L". I decided on the weaker formulation for the following reason: An empirical theory will always be applied in a context where the understandability of certain terms and sentences is not questioned. These terms may be anything but Observation terms' as conceived by an empiricist, and the sentences may be far from being Protokolls t^e. Hence, they should not be covered by the theory's semantic basis. The application terms must receive a basic interpretation unless one wants to give up all hope of calling a sentence of the theory either true or false (no great sacrifice to an instrumentalist' philosopher of science); the only general requirements on the terms of the application language 78
The real contrast, then, is between INTs and APTs and not between AT1 s and AT2s as some formulations in Suppes (1957), Ch. 12, or Sneed (1971), Ch. 1, might suggest; on our account the AT1—AT2 dichotomy is not reproduced for realized theories, where we have both INTps and APTps,
80
Hans-Heinrich Lieb
should be pragmatic ones such as, perhaps, understandability in a standard context of use.79 Let me emphasize, though, that this semantic open-endedness of the application language is not meant as a loop-hole for terms from other theories to slip in: For dealing with such terms, non-logical constants other than axiomatic ones have been allowed into the applied-core language (their role will become apparent in the next subsection on theory integration). Further requirements on the application language of an empirical theory may still be in place—admitting that the concept of 'empirical' has become dubious after the Zusammenbruch der Signifikan^idee—the sceptical belief that there is no important concept of empirical significance to be explicated.80 Third objection to (10).
Leaving the axiomatic core constants uninterr>reted has the consequence that no APT is a theory of something or, conversely, has a subject matter: It is by such constants that the subject matter of the theory would have to be denoted (e.g., "theory of English"}. This is a serious objection that seems to have passed unnoticed in the literature. I see two possible ways of countering it: One is a recourse to pragmatics by understanding "theory oP' as referring to a relation involving the theoretician, e.g. it might be proposed to understand "theory of" and "subject matter oP' always in the sense of "intended theory of (subject matter of)". The term after "theory of" could then be characterized independently of the theory as to its reference. The second counter-argument, which I will here adopt, is a semantic one. Although the core constants of the theory è are not in the domain of the basic interpretation í of È, they are in the domain of the possible-value function f and are thus assigned a set-theoretical construct in the domains of è as their set of possible values. We may use this fact and propose definitions along the following lines. Let ÷ be any set. is a theory of ÷ iff, for some C and n: is a minus-C interpreted theory, and there is a non-repetitious n-tupel ^,..., e,,) and an n-tupel <÷ ë ,..., ÷«) = ÷ such that {ej,..., en} = the set of axiomatic primitives of Ô and xj 6 f(e$) if Cj 6 C, and xj = v(Cj) if ej ö C, where f = the third member of  and í = the second member of B.81 Conversely, ÷ is a subject matter (an object} of , if is a theory of x. We generalize to an APT è s follows: è is an abstract theory of ÷ if is a theory of x, where B* = (the domain sequence of 0, the empty relation, that subfunction of the possible-value function of è whose domain does not contain any expressions with application constants of è>. è is an applied theory of x if
For a similar position, cf. Kutschera (1972) I, Ch. 3, § 2. See also Lewis (1971) 428, who even includes all non-logical terms that are not axiomatic terms of Tj. 80 Cf. Stegm ller (1969/70) II, Ch. 5, § 13. 81 This presupposes a finite number of axiomatic primitives. If we want to consider a non-finite number, we might choose an appropriate function instead of an n-tupel.
Grammars as theories: the case for axiomatic grammar (part I)
81
of È, the semantic basis of è> is a theory of x. Conversely, ÷ is an abstract subject matter (object) of è if è is an abstract theory of x, and a specific subject matter (object) of è if è is an applied theory of x.82 As the most striking consequence of the definitions we have an indeterminacy of subject matter except for completely interpreted theories.83 Hence, all APTs are indeterminate even for their specific subject matters. But this is quite different from not having a subject matter. Thus, the third objection is rejected if we do not accept the presupposed uniqueness of subject matters.84 The fourth and last objection is related to the third. Fourth objection to (10). Leaving the axiomatic core constants uninterpreted may have the consequence that no sufficiently complex APT in the empirical sciences is either a true theory or a false theory of something. This claim is indeed correct. We might propose to call an APT è a true theory of x if it is a theory of x and, for Ô its applied core and  its semantic basis, the Ramsey sentence of Ô is BT-true. But this would be inadequate because the Ramsey sentence is true in case there is some appropriate ç-tuple, which might not be x. The only acceptable explication may be as follows: è is a true (abstract) theory of x if è is an (abstract) theory of x and there is a Bj such that, for  = the semantic basis of è and Tt = the applied core of è: The first and third members of Bj = the first and third members of Â; the second member of  £ the second member of Bj; is a completely INT; and the Ramsey sentence of T! is T1B1-true; if it is false, è isfo/se. In discussing the first objection to (10) we already formulated a hypothesis for sufficiently complex empirical theories that would exclude existence of a Bj such that is a cINT. Hence, no such theory would be either true or false. I find this consequence acceptable. On our conception, a theory is not a sentence; in particular, it is not the conjunction of its axioms. Whereas "theory of" may be needed as a semantic concept, "true theory of" may well be replaced by pragmatic ones as far as the component of truth is concerned. Let me emphasize, though, that in this subsection we have been moving on notoriously unstable ground and should not feel certain to have gained a tenable position. We finally turn to the question of how different theories may be combined into new ones.
82
All these definitions generalize to interpreted and applied formalized theories. The trivial multiplicity of subject matters due to different arrangements of the axiomatic primitives may be disregarded. 84 There may, of course, be exactly one intended subject matter; this is a relevant pragmatic feature of a given theory, see above. 83
6 TLll/2
82 3.4.
Hans-Heinrich Lieb Integration of theories: Conflation and use
The relations between different theories that have so far been studied in logic and the philosophy of science are mainly of two types: In logic, 'extensions' or 'subtheories' of a given theory have been considered; in the philosophy of science problems of 'reducibility' of one theory to another have received special attention, due to the specific problems of the natural sciences.85 It is perhaps not surprising that there are important relations between linguistic theories which are not covered by studies of this kind. Take the following examples: (a) A psychological theory is used in a theory of language to derive theorems of that theory, (b) A theory of language is used in a grammar of an individual language to derive theorems on that language, (c) Axiomatic constants of a theory of language are combined with axiomatic constants of a grammar to form axiomatic terms of the grammar ("is a word in", "English", "is a word in English"), (d) The phonological theory of a language is combined with a morpho-syntactic theory of the same language to form a grammar (in the traditional narrow sense). Whereas a 'subtheory' approach could be considered for (d), it would be misleading to attempt either a subtheory or a reduction approach in cases (a) and (b): A theory of language is not 'reduced* to the psychological theory that it may invoke in the derivation of theorems, and is certainly no 'subtheory' of it; and the same holds for grammars and theories of language. Rather, we are confronted with a relation of 'presupposition*. In (c) the situation is still different; we may .say that the grammar is (partly) formulated 'in terms oP the theory of language (which should also be presupposed). In (d) we may speak of 'conflation' of theories. Whereas the sentences of a presupposed theory should not appear among the sentences (axioms, theorems, and definitions) of the given theory, all sentences of the conflated theories should also be sentences of the conflation. In explicating the various concepts we must distinguish between abstract and realized theories. In the special case of formulation-in-terms-of it will turn out that an applied axiomatic theory formulated in terms of an abstract one provides a 'completion' for that theory, i.e. another applied theory 'based' on the abstract one. This result is of special importance because it suggests that a theory of language may be taken as an abstract axiomatic theory in terms of which certain applied theories, vi%. grammars, are formulated, which then provide completions for the theory of language. For the following definitions we first consider AT1 s and then generalize to 86 APTs Let T and Tt be any AT*s and è and 0X any APTs. We first define "conflation" as follows. 85
For a recent study of deducibility' and 'equivalence* of theories, cf. Sneed (1971) Ch. 7. Cf. also Kutschera (1972) II, Ch. 4, § 6. 86 I have not attempted to include AT2s and INTs (which seems possible) because the linguistic theories to be proposed will not assume that format for reasons indicated above (§ 3.3., first objection to (10)).
Grammars as theories: the case for axiomatic grammar (part I)
83
THE CONFLATION OF T AND Tt = the ordered quadruple such that: (1) ^ = the conflation of T and the core of È. (2) T2 = the conflation of T and the applied core of È. (3) L = the amalgamation of the axiomatic language of T and the total language of È. (4) B = the semantic basis of È. The confotion of è and è i^ the quadruple such that: (1) Ô = the conflation of the core of è and the core of ÈÉ . (2) T! = the conflation of the applied core of è and the applied core of èé· (3) L = the amalgamation of the total language of è and the total language of 6j . (4) Â = the 'addition' of the semantic basis of è and the semantic basis of 0j, in the following sense: Â is a triple whose second and third components are the union of, respectively, the second and third components of the semantic bases of è and è÷';·· the first member is the (n + m)-tupel obtained from the domain sequence of è (ç = the number of its components) by adding the components of the domain sequence of È ÷ that are not in the domain sequence of è (keeping sequential order). So far I have not established the precise conditions under which the conflation of Ô (è) and 0j is again an APT but it seems quite likely (on the basis of informal considerations) that the Conflation Theorem can be extended in an appropriate way. 87
For "amalgamation", cf. § 2.3, p. 68. The same constant could be an axiomatic primitive in Ô and defined in Tj, or conversely. 89 This avoids circularity when axiomatic primitives of Ô are defined by means of axiomatic primitives of Tj. 88
84
Hans-Heinrich Lieb
Note that even in the case of AT1 s the conflated theories may not be parts of the conflation as "part" was defined in §§ 2.4 and 3.1: Some axiomatic primitives and axioms of the conflated theories may not be among the axiomatic primitives and axioms of the conflation. Distinguishing between "part" and "subtheory", we may define the latter term as follows: If T and Tj are AT^s, T is a SUBTHEORY of Tj if there is an AT1 T2 such that T! = the conflation of T and T2. This concept is generalized to APTs in an obvious way. The (homogeneous) subtheory relations are obviously reflexive. We now explicate the concepts connected with theory use. T is USED in T! (Ti PRESUPPOSES T) iff: (1) The axiomatic language of T is logically compatible with the basic language of T^90 (2) Some valid non-logical sentences of the basic language of Tj are valid (non-logical) sentences of the axiomatic language of T. (3) A valid sentence of the axiomatic language of T is a valid sentence of the axiomatic language of Tj if and only if it is a valid sentence of the basic language of Tj. T is used completely in Tj (Tj presupposes all of T) if T is used in Tt and all valid nonlogical sentences of the axiomatic language of T are valid (non-logical) sentences of the basic language of Tj 91 T is used m è if T is used in the applied core of è ; è is used in T if the applied core of è is used in T; and è is used in 0j if the applied core of è is used in the applied core of È1. "Used completely" is generalized in the same way. Formulating a theory in terms of another is explicated as follows. T IS FORMULATED IN TERMS OF Tx iff: (1) The axiomatic language of Tt is contained in the basic language of T.92 (2) The axiomatic language of Tj contains a Pre-IFS of predicate logic. (3) No undefined term of the axiomatic language of Tj is a defined term of the axiomatic language of T. (4) For each axiomatic primitive e of T there is an ej, e2, e3 such that: (4a) et is a closed non-logical term of Tj. (4b) ej is a predicate expression and ej e2 is an atomic formula of the axiomatic language of T, or ei is a functor expression and Cj.-.e^. is a term of the axiomatic language of T. (4c) e occurs in e2. (4d) ^ is an axiomatic term of Tx or there is an axiomatic term of Tj that occurs in e2. (4e) e3 is an axiom or definition of T. (4f) e^ occurs in e3. If "each" in (4) is replaced by "some" we say that T is partly formulated in terms of Tj. If T is formulated in terms of Tj and the predicate logic contained in Tj (hence, in T) is sufficiently strong to contain an abstraction operator (like the ë-operator), then for any axiomatic primitive e of T there is an axiomatic term ej of T obtained as follows. Let e2 and e3 be expressions such as ej and e2 in (4) of 90
For "logically compatible", see § 2.3, p. 67. Note that theory use does not require use of an axiom, theorem, or definition of T; a valid non-logical sentence of the axiomatic language of T must be involved but this may be a valid sentence of the basic language of T. (The Conflation Theorem allows only for theory use of this limited type: cf. (3) in the definition of "compatible".) It is only complete theory use that necessarily involves the sentences of T. 92 Because of (1), formulation-in-terms-of implies complete theory use. 91
Grammars as theories: the case for axiomatic grammar (part I)
85
the above definition. Let €4 be an expression obtained from e3 by replacing all constants other than e by appropriate variables. We now define ej as follows: If e4 = €,€! = e. Otherwise, ej is the closed term obtained from e2 e4 by prefixing an abstraction operator that binds all the free variables in e4. Depending on whether 6264 is a formula or a term, ej is a predicate expression or a functor expression.93 If, in addition, there are no defined axiomatic constants of T, reliance on the nonlogical terms of Tj is particularly strong: T is strongly formulated in terms of Tj if T is formulated in terms of ¾± and there are no definitions of T. The concept of formulation-in-terms-of is extended in analogy to the extension of use: T is formulated in terms of è if T is formulated in terms of the applied core of è and the axiomatic primitives of T are non-logical, non-mathematical constants of Ô; è is formulated in terms of T if the applied core of è is formulated in terms of T and the axiomatic primitives of T are non-logical, nonmathematical constants of T; and È is formulated in terms of È1 if the applied core of È is formulated in terms of the applied core of È! ,94 If è is formulated in terms of T, the abstract theory T can be 'completed' into an APT ÈÉ by making appropriate use of è. This result is of great importance for the problem of relating grammars to a theory of language and will therefore be established in a more precise way. We first introduce the following concept, for any AT1 T and APT è. THE è-COMPLETION OF T = the quadruple such that: (1) Tj = the quadruple ^Li> C, S, Sj) such that: (a) Lj = the core language ïßè. (b) C = the axiomatic primitives of T. (c) S = the C-axioms of Lj. (d) Sj = the C-definitions of Lj. (2) T 2 =the quadruple
To give a simple example: If Tj is a theory of language and T a theory of English, e could be "English"; e^ = "Word" ("is a word in", a two-place predicate); e3 = e4 = "(x,English)", with "x" a variable; ej — "(ë÷) (Word(÷,English))" ("the class (or property) of all ÷ such that x is a word in English"). 94 The condition on T is to exclude the awkward case of a logical or mathematical theory being formulated in terms of a non-logical or non-mathematical one, and conversely. 95 For "application", cf. §3.1, p. 75.
86
Hans-Heinrich Lieb
"formulated in terms of" and "axiom system'*) the proof is complex and has to be omitted for lack of space. As one important feature of the proof we mention the following idea: Assume that Tj and T2 in the completion are indeed AT^s. Consider the set consisting of the axiomatic primitives of Tt and the undefined basic non-logical, non-mathematical constants of T2. Then consider the set consisting of the axiomatic primitives of the core of È and the undefined basic non-logical, non-mathematical constants of the applied core of è. Under the conditions of formulation-in-terms-of these two sets are identical, only the role of a given constant may change (the axiomatic primitives of T2 are undefined basic non-logical, non-mathematical constants of the applied core of è). Hence, the semantic basis of è should serve as a semantic basis for the completion. The Completion Theorem enables us to extend the notions of "theory of" and "subject matter of" (§3.3, pp. 80f) even to abstract axiomatic theories, relativized to applied theories that are formulated in terms of them. Let è be an APT and Ô an AT1. Ô is a theory of ÷ relative to è if è is formulated in terms of Ô and the è-completion of Ô is an abstract or applied theory of ÷. ÷ is a subject matter of Ô relative to S if T is a theory of ÷ relative to è. Thus, there is a relative notion of subject matter that would apply to a theory of language even when taken as an abstract axiomatic theory.
4.
Grammars as axiomatic theories
4.1.
The program and its linguistic background We shall make a weaker and a stronger proposal for grammar writing:
(2)
Thesis A. A complete or partial linguistic grammar can and should be formulated as an applied axiomatic theory in the language of predicate logic or set theory.
(3)
Thesis B. A complete or partial linguistic grammar can and should be formulated as an applied axiomatic theory in the language of predicate logic or set theory and in terms of a part or subtheory of a theory of language that presupposes a part or subtheory of a theory of communication.
Both proposals are meant for scientific grammars, not for pedagogic ones. They are to be understood within the framework of axiomatic theories developed in §§ 2 and 3: For "applied axiomatic theory" and "formulated in the language of predicate logic or set theory", cf. §3.1, pp. 73 and 76; for "subtheory", §3.4, p. S4.96 Requiring an applied rather than an interpreted theory is motivated by 96
Part-of and Subtheory-of are reflexive; hence, the two theses allow for formulation in terms of the complete theory. In (3) it is to be understood that a complete grammar would be formulated in terms of a theory of language whereas a partial grammar might require only a proper part or a subtheory.
Grammars as theories: the case for axiomatic grammar (part I)
87
considerations in § 3.3 (cf. discussion of first objection to (10)). Requiring formulation in predicate logic or set theory is anything but self-evident and will have to be justified (below, § 4.3). We are left with the task of explaining the expressions "theory of communication", "theory of language", and "complete or partial linguistic grammar". This amounts to characterizing the linguistic background of our program now that the general framework for theories has been developed. In the present subsection I shall informally develop a general outline; the necessary detail will be filled in later in Part II. I draw a basic distinction, frequently obscured in modern linguistics, between two disciplines: the theory of language and the theory of linguistics; the latter has the theory of linguistic description as one of its branches.97 The theory of language develops individual theories on the properties of natural languages and their systems (see below); the theory of linguistic description develops, among other things, theories of grammars of natural languages and their varieties, such as dialects; any such theory is relative to a theory of language.98 At this point we may ask whether a theory of grammars is relative to a theory of language in the sense of presupposing such a theory. I would indeed take the position that "theory of grammars" is most fruitfully understood in a sense where any axiomatic theory of grammars presupposes a theory of language in the defined sense of "presupposes" (§ 3.4). The present paper may go quite a long way to developing an axiomatic theory of axiomatic grammars but it certainly will not arrive at an axiomatic formulation; hence, relativity to a theory of language can only be taken as an informal analogon to presupposition in the technical sense. Let us say that the proposed theory of grammars assumes a theory of language, to stress the analogy. The same relation may hold between the proposed theory of grammars and the framework for axiomatic theories (which is not yet an axiomatic theory). I indeed believe that the proposed theory of grammars if axiomatized would presuppose an axiomatic theory of language and an axiomatic version of the framework for axiomatic theories, but I will not here pursue this idea any further. Turning to theories of language I shall assume that any theory of language contains exactly one axiomatic constant, to be called the distinguished term of the theory, that is 'meant to refer to languages'.99 Likewise, for any grammar I shall assume that it contains exactly one constant or ç-tuple of constants, to be called the distinguished term of the grammar, that is 'meant to refer to the object of the grammar'. I further assume that an axiomatic theory of language is of limited fruitfulness if it neither is a part or subtheory of a theory of communication nor 97
There is at least one other branch, vi%. the methodology of linguistics, dealing with the methods applied or applicable in any branch of linguistics. 98 For a more detailed account of this position, see Lieb (1970) § 1.7, or Lieb (1968a) § 1. 99 For the problem of explicating this phrase, cf. § 3.3, third objection to (10).
88
Hans-Heinrich Lieb
presupposes such a theory.100 Presupposition seems preferable to a subtheory approach insofar as one might wish to use other theories in a theory of language without granting special status to a theory of communication. Differently from Lieb (1970), I will here adopt the presupposition approach, to which the theory in Lieb (1970) is easily converted. The theory presented in Lieb (1970) is the most comprehensive axiomatic theory that presently exists in linguistics. At the time of working it out I had no clear picture of theory integration, which made me take my theory of language as a part of a theory of communication; also, I had not developed the conception of realized theories as outlined above, § 3.1. I would now say that the theory was intended as an applied axiomatic theory, with the application characterized in general terms only. I shall here assume the theory as part of the linguistic background, restructured as follows: The theory is part of a theory of language, and may be taken as an abstract axiomatic theory if Thesis B is accepted. It presupposes a (part of a) theory of communication, which in turn presupposes a theory of physical time. Most of the theory in Lieb (1970) consists of the presupposed (part of a) theory of communication and can be characterized informally as follows. Consider a linguistic complex > i. e. either a complete natural language through time (an historical language) or a complete variety of an historical language, such as an historical dialect; or a stage or a period of a language or language variety. Any linguistic complex is a communication complex in the sense of Lieb (1970): a non-empty finite class of means of communication. A variety is a subclass of the language, and a stage or a period of a language or variety is a subclass of the latter. The elements of a linguistic complex are called linguistic means of communication. A linguistic means of communication (corresponding, roughly, to a homogeneous idiolect in a synchronic sense) is a class of abstract texts. Each text is a pair of a 'form' and a 'meaning', both abstract if compared to actual utterances and their meanings. More specifically, an appropriate class of texts is a means of communication for somebody during a certain time. It is actually 'used' by the speaker, i.e. certain texts are 'realized' by meaningful utterances of the speaker. Given the class, the speaker, and the interval, there is a system of the means of communication that uniquely determines the means and is, during the interval, at the speaker's disposal. Given a stage of a language or language variety, there is a system for the stage that is an entity of the same kind as the systems of the elements of the stage and is related to them by a relation of abstraction. More specifically, there is a finite sequence of such systems, each more abstract than the preceding one. When a certain degree of abstraction is reached, a system for a stage may also be a system for other stages, and eventually a system for the complete language or variety. It is assumed that there is such a system for any complete natural language or variety.
1
Cf. Lieb (1970) § 1.3, for arguments.
Grammars as theories: the case for axiomatic grammar (part I)
89
Allowing for certain qualifications to be mentioned below (§ 4.2) the theory of Lieb (1970) will be assumed in the present essay in an extended form that incorporates additional assumptions along the following lines (these assumptions would not be expressed in the presupposed theory of communication but by sentences—axioms, theorems, or definitions—of the theory of language). Any system of a linguistic means of communication contains the phonological, the morpbosyntactic, the semantic, and (possibly) the pragmatic subsystem; the phonological system may be replaced by an analogous one, say, a graphemic system. More specifically, the system is the triple or quadruple of its three or four subsystems (ordered as above), or rather, the corresponding unit class.101 Within these subsystems, minor systems may be distinguished, to be called parts of the subsystems. The phonological and morpho-syntactic subsystems are the syntactical subsystems of the complete system. A number of additional assumptions will be formulated in Part II of the present paper. A system for a linguistic complex is structured in the same way as a system of a linguistic means of communication. On the basis of this sketch, at least the following subtheories of the theory of language may be assumed: A theory of systems (to be further subdivided into a phonological, morpho-syntactic, semantic, and (possibly) pragmatic theory) dealing with the structure of systems (for linguistic complexes or of their elements), and a realisation theory dealing with the relations between systems and spatio-temporal entities. This subtheory may perhaps be further subdivided into at least a phonetic theory and a speech act theory, depending on which spatio-temporal entities are primarily involved. This division will be tentatively accepted for the rest of the present essay. The theory of communication used in the theory of language provides most of the basic non-logical constants of the theory and its subtheories: ".. .is a means of communication (MC) for ... during ...", ".. .is a system of..'.'" etc. I will now assume a theory of grammar distinguishing complete and partial linguistic-grammars as follows: The intended subject matter102 of a complete linguistic grammar is either a linguistic complex and a system for the complex, or a linguistic means of communication and a system of the means. The intended subject matter of a partial linguistic grammar is either a linguistic complex and a (part of a) subsystem of a system for the complex, or is a linguistic means of communication and a (part of a) subsystem of a system of the means. If the subject matter involves a means of communication, we shall speak of a (complete or partial) idiolect grammar. 101
The unit class is chosen for logical reasons explained in the next subsection. In the extended theory, systems lose their status as individuals which they had in Lieb (1970); this changes the logical type of certain variables and constants in my earlier theory. 102 "Intended" shows that we are tacitly assuming a pragmatic relation. I will not go into the relevant problems. They are analogous to the ones discussed for theories in § 3.3, third objection to (10).
90
Hans-Heinrich Lieb
The expression "is a grammar of will be construed as follows: A complete grammar of a linguistic complex and a system is any linguistic grammar whose intended subject matter is the complex and the system; a partial grammar of a complex and a system is a linguistic grammar whose intended subject matter is the complex and a (part of a) subsystem of the system; analogously, for linguistic means of communication.103 Accepting either Thesis A or Thesis B substantially enriches the theory of grammar as should be obvious from the framework for axiomatic theories developed in §§ 2 and 3. In particular, we may introduce the following assumptions, again stated informally: A grammar of a linguistic complex and a system is an abstract theory o/104 a couple consisting of the complex and the system; correspondingly, for a linguistic means of communication. The two assumptions may have to be weakened by choosing an ç-tuple that includes other entities besides the ones just mentioned. As the grammar is an applied axiomatic theory, it is an applied theory of an n-tuple that includes the complex (the means of communication), the system, and entities denoted by the application terms of the theory. In the case of an idiolect grammar a speaker and a time interval may be assumed among those entities. If Thesis B is acceptable we further make the following assumptions, for any grammar formulated in terms of the presupposed theory of language: A grammar of a language contains a theorem of the form c 6 C,105 where C is the distinguished term of the theory of language and c the distinguished term of the grammar; a grammar of a language variety (that is not itself a language) contains a theorem of the form: For some cls ct e C and ccq; a grammar of a linguistic means of communication contains a theorem of the form: For some q, G! 6 C and c € ct. Each sentence will be called the distinguished sentence of the corresponding grammar.106 Our main task will consist in demonstrating that the proposals (2) and (3) are feasible, given a theory of language as outlined above. The whole of Part II will be devoted to that demonstration. Throughout, we concentrate on the stronger Thesis B although most of our arguments could be reformulated so as to support at least the weaker proposal. The linguistic background of our program has been characterized in fairly general terms only. In the next subsection I shall be more specific about the logical properties of the theory of language that will be assumed. (The next two subsections are of a more technical nature; the reader interested in the advantages and the history of our approach may turn to § 4.4 immediately.) 103
Further explication would, of course, require an explication of "linguistic grammar". For "abstract theory of and "applied theory of below, cf. § 3.3, p. 80. 105 Or an analogous form, depending on the system of logic involved. 106 The concept of distinguished sentence could be generalized by relativizing it to arbitrary theories of language. This is important in connection with language universals (Lieb to appear, §§6.2ff) but may be neglected. 104
Grammars as theories: the case for axiomatic grammar (part I) 4.2.
91
Questions of logic
Using the theory in Lieb (1970) for our present purposes is more problematic than our previous account might suggest, due to the language in which the theory is formulated. The applied-core language (and, in principle, the total language) of the theory of language in Lieb (1970) is a couple Á÷) of higher-order predicate logic of the following kind: The underlying logic of (Fi> Ä^ is a four-sorted version of Carnap's Language C (Carnap (1958)), i.e. a version with four types of individual expressions ; it is a system assuming the 'simple theory of types' with a finite number of levels.* <^Fl5 Ä^ contains the axiomatic language of an abstract axiomatic theory of the real numbers. Ä! is the set of all D such that: The domain of D = the set of constants, closed terms, and closed formulas of F1-minus-C*, where C* = the set of undefined non-logical, non-mathematical constants of F t ; and D satisfies the rules of value assignment and evaluation (appropriately modified) for Carnap's Language B ((1958) § 25).107 F is a German reading of F^108 For some rendering r of F! as F, Ä is the set of r-forms of elements of Ë ÷. Thus, by the theorem on r-forms (§2.3, p. 67),
(F!,^).
Actually, the preceding account is not quite correct in the following respects: Fj and a certain rendering of Fj as F were characterized in general terms only, and the same holds for the application of the theory of language (its application axioms, semantic basis, and application language). If (3) is acceptable, the theory of language does not have to be an applied theory at all. According to §4.1 we take an extension of the theory in Lieb (1970), characterized informally, as the theory of language to be assumed for our account of grammars and for possible use in the grammars themselves. To avoid confusion due to different natural languages (the present essay is in English), we modify this decision as follows: First, we consider not the theory in Lieb (1970) but a symbolic counterpart of the theory obtained by replacing the theory's language by a version of Carnap's Language C (as just indicated) and replacing the axioms, definitions etc. accordingly. We might then take the theory of language as an extension of the counterpart; i.e. take the latter as a part of the former. This would be problematic on a number of accounts. First, individuals of certain types (systems, texts) must be treated as non-individuals in the new theory of language, which therefore would be an extension not of the counterpart but only of a modification of the counterpart. Second, whereas 0^, ËÉ) is an extensional Pre-IFS of logic (cf. § 2.3, p. 67), this may not be tme of the axiomatic (the total) 107
It has to be assumed that <Ñ1,Ä1> is a Pre-IFS of logic; no proof is possible because the term "minus-C interpreted formalized system of logic" was not defined. (Cf. the definition of "Pre-IFS" in î 2.3). 108 Possibly with the qualifications noted in § 2.2, fn. 49.
92
Hans-Heinrich Lieb
language of the new theory (for reasons to be discussed in the next subsection). Thirdly, we may not be able to restrict ourselves to a finite number of levels: In the semantic part of the theory of language, axiomatic constants of transfinite levels may be required.109 Finally, there are minor problems with versions of Carnap's Language C, in particular: They do not contain constants or variables whose values are ordered ç-tuples; such constants and variables may be needed. For the theory in Lieb (1970) we therefore consider a symbolic counterpart as above, with qualifications as follows: The (total or axiomatic) language of the counterpart is either an extensional Pre-IFS obtained from an appropriate version of Carnap's Language C by making the following changes: (a) modifications as required by restructuring the domains of individuals; (b) minor changes of symbolism; or it is a non-extensional Pre-IFS that is a modification of an appropriate version of Carnap's Language C and also incorporates the previous changes (a) and (b). The theory of language to be assumed for the present essay will then be an extension of this counterpart, and may be so even with respect to logic by containing variables and constants of transfinite levels. Ordered ç-tuples could be assumed as introduced by a variant of the Wiener-Kuratowsky method. We shall, however, do without them and instead take the unit classes of ç-tuples, which are available in Carnap's Language C. The 'minor changes of symbolism' will be pointed out at their first occurrence. They include use of "e" as an auxiliary symbol (not as a constant); thus, "R<x, y>" and "(x, y^ e R" are treated as notational variants of each other. The assumed theory of language will not be developed formally. In informal discussion I may continue to use expressions of an English language reading (obtainable on the basis of Lieb (1968^) or by translating from Lieb (1970)), speaking of classes, elements etc. (cf. Carnap (1958) §28c). I do not wish to maintain that every theory of language should have the logical properties just indicated. On the contrary, if extensional systems are sufficient for formulating a theory of knguage we may try to avoid complexities such as transfinite levels by formulating theories of language in the language of set theory. This would also solve a problem in our informal version of a theory of grammars: The framework for axiomatic theories in §§ 2 and 3 was formulated by using set-theoretical means of expression. Assuming that a formal version of the theory of grammars presupposes both a theory based on our framework for axiomatic theories and the theory of language, an awkward situation develops if the language of set theory is used in one but not the other theory. It will soon become apparent, though, that a general restriction to the language of set theory may be problematic.
1
For the reasons, cf. already Carnap (1942) § 12.
Grammars as theories: the case for axiomatic grammar (part I)
4.3.
93
Problems of feasability
In this subsection we consider various objections that question the general feasability of our program. Both Thesis A and B contain the requirement that a grammar should be formulated in the language of predicate logic or set theory (in the sense of § 3.1, p. 76). Although we did not specify the conditions that would make a PreIFS of logic a Pre-IFS of predicate logic, the requirement may seem too strong if we take "predicate logic" to imply extensionality of the system, as we certainly would in the case of "set theory". Thus, we arrive at the following Objection 1. The program is doomed because the total language of a grammar cannot be extensional. We may argue as follows. According to §4.1, a complete linguistic grammar will contain not only a semantic part but may even contain a pragmatic one, and its application will establish a relation to actual language use. In the semantic part the grammar will have to deal with meanings of (abstract) expressions and in its application, possibly also in its pragmatic part, with attitudes of speakers. For either task a non-extensional knguage is required. If Thesis B is adopted we may even strengthen this argument as follows: The axiomatic language of the theory of language in terms of which a complete grammar is formulated is contained in the grammar's total language. In §4.1 we assumed two subtheories of a theory of language, a systems theory containing a semantic and (possibly) a pragmatic part, and a realization theory containing a speech act part. Developing these three parts requires a non-extensional language for the same reasons as given before in the case of grammars. Hence, the total language of the grammar cannot be extensional. I would immediately subscribe to these arguments except for the crucial assumption: Non-extensional languages are required for dealing with meanings of abstract expressions (i.e. not of concrete utterances or speech acts) and for attitudes of speakers.110 First, consider meanings. The first explicit discussion with which I am acquainted is Carnap (1956) § 38. His result, though inconclusive, supports the hypothesis that an extensional language may be sufficient for the semantic description of interpreted formalized systems of logic.111 Recent developments in logical semantics ('possible-worlds semantics') also seem to support that hypothesis : Intensional entities such as properties are all conceived as set-theoretical constructs, and the semantic metalanguage contains only sentences that could be 110
Note that such attitudes are also involved in the meanings of utterances, which are intended meanings (or understood meanings). 111 "On the basis of these considerations, I am inclined to believe that it is possible to give a complete semantical description even of an intensional language system like S2 in an extensional meta-language like Me. However, this problem requires further investigation." (1956,172).
94
Hans-Heinrich Lieb
sentences of the axiomatic language of an abstract set theory, including constants for specific sets. However, I have been unable to ascertain whether this is sufficient for extensionality in a Carnapian sense.112 Next, consider attitudes, in particular, so-called prepositional attitudes like knowing, believing, meaning, intending etc. It is apparent from the work of Irena Bellert that these concepts may occupy an important place in grammars, and fairly obvious from the work of Bellert and other authors that they should occupy such a place in a theory of language.113 We need non-logical constants such as "believes", "intends" in a grammar and a theory of language; it may be left open whether these are axiomatic constants of a theory of language or of a psychological theory used in the theory of language.114 The question, then, is whether an axiomatic theory of the relevant prepositional attitudes (that might be part of a theory of language or of actual grammars) can be formulated in an extensional language. Relevant discussion is found mainly in logical semantics and the philosophy of language, where an extensive literature on prepositional attitudes has accrued, centering to a large extent around the concept of belief. Mainly since Carnap (1956) (i.e. 1947) it has been widely debated whether belief should be taken as a relation between a person and a sentence or a person and a proposition, or still something else.115 A decision in this matter is immediately relevant for our problem: According to Carnap, an extensional language may suffice if sentences are chosen (1954: (1956) 232); Rescher, who advocates propositions, reaches the conclusion that "an adequate logical theory of belief statements is not, it would seem, to be had unless modal concepts be presupposed as an available tool for its development" ((1968) 53). Again, I am not certain about the extensionality of the language if a possible-worlds semantic is chosen for treating propositional attitudes (cf., e.g., Hintikka'(1969)). 112
Carnap (1956) 168f, considers one of his metalanguages as non-extensional simply because it contains sentences which say what the intension of a certain expression is; in his view, any such sentence is non-extensional. This would seem to answer our question also for the more recent developments. But Carnap's supporting example seems to be wrong; he apparently confuses equivalence in the metalanguage with, equivalence in the object language. He also presupposes that an expression such as "the property Human" would be a predicate expression of the metalanguage because "Human" is; in a semantics taking properties as particular sets and sets as individuals, such an expression would be an individual expression. 113 Cf. Bellert (1972*) (1972£) (to be discussed in § 4.5). For the role of propositional attitudes for a speech act theory (as a subtheory of a theory of language), cf. Searle (1969) as a classic reference. 114 It may be argued that such constants should be taken as logical operators; this would affect the form not the substance of our argument. In any case, we are not concerned with actual Verbs of believing* etc. in natural languages. 115 The question is taken up from a linguist's point of view in Partee (1973). Cf. also Moravcsik (1973) for a survey of part of the relevant literature.
Grammars as theories: the case for axiomatic grammar (part I)
95
Given this situation, I am taking the course already hinted at in the preceding paragraph: I explicitly allow as Pre-IFSs of predicate logic systems that are nonextensional. If necessary, I would be prepared to give up the relevant parts of Theses A and B: It is not essential to the program that the underlying logic should have a particularly simple structure. There is a second objection of a more fundamental nature: Objection 2. The program is problematic because it naively assumes axiomatizability. This objection loses part of its force because we did not require that the (axiomatic or total) language of an axiomatic theory should be axiomatizable (in the sense of §2.2, p. 61). Given a formal theory, abstract, interpreted, or applied, axiomatizability of the theory involves only a subset of the set of valid sentences of the theory's language (cf. § 2.4, p. 70, 3.1, p. 75). To make sense of the objection within our framework we have to assume that an Ordinary' grammar can be represented by a formalized theory of some sort. The question then is: Are those theories axiomatizable in one of the previously introduced senses? For representation as a formalized theory we have to decide on which set of sentences of the theory's language should be the third member of the theory. Objection 2 may already break down because no such set of sentences can be delimited unless the resulting theory is axiomatizable. In the foundations of mathematics it was possible to find non-axiomatizable formalized systems because in some cases the set of valid sentences could be defined semantically as true under such and such an interpretation; it then turned out that the system with those valid sentences was non-axiomatizable. In empirical theories we have the problem of uninterpreted constants that complicates the concept of truth. I have been unable to extend to formalized theories even the relativized concept of truth introduced above for realized axiomatic theories (§ 3.3, Third Objection to (10)). One might think of characterizing a set of valid sentences 'pragmatically' but this should always lead to finite sets unless a semantic or syntactical characterization is included. But suppose we have a concept of truth that applies to interpreted and applied formalized theories, and È is an applied formalized theory and is a theory of a linguistic means of communication.116 It may be natural to require that every true sentence of the form: s is a text of C, is an element of the third component of È (s and C are constants and C is 'meant to refer to the intended subject matter of the theory'). È will be axiomatizable only if the set of all those sentences is recursively enumerable. Assuming that it is corresponds to the one most basic assumption of generative grammar (at least, if "text of C" is understood as "sentence of C"). To my knowledge, only Henry Hiz has argued against that assumption in favor of a weaker one (Hiz (1968) 248f). Suppose the set of those sentences is not recursively enumerable. Then we may either give up the require' For "theory of, cf. above, § 3.3, fn. 82.
96
Hans-Heinrich Lieb
meftt that it should be a subset of the third component of È, which removes the obstacle to axiomatizability; or we may attempt to find a significant subset (of the set of sentences) that is recursively enumerable; that is, we weaken our claim of axiomatizability to a part of È. An axiomatization of that part would then be an incomplete axiomatization' of È, in an obvious sense. I conclude that the assumption of axiomatizability is at present defensible; in the worst case our program would have to be relativized to 'reduced grammars', /. e. to theories that are 'incomplete axiomatizations' of corresponding formalized theories. One might also consider re-formulating it for formalized instead of axiomatic theories. It would then considerably lose in interest: There are additional semantic problems with formalized theories; the results on theory integration of § 3.4 do not seem to generalize to formalized theories in a non-trivial way;117 and the systematizing effect of axiomatic theories would be lost. Objection 3. The program is misguided: It is based on a conception of theories that derives from and aims at theory construction in the natural sciences; this conception is [may be] inadequate, or partially inadequate, for the humanities to which linguistics belongs. I would subscribe to the premises of the objection in their non-dogmatic version ("may be" instead of "is"), taking as the conception of theories the ordinary one in the philosophy of science. I would not, however, accept the conclusion, for the following reasons: (a) The very attempt to apply a certain conception may be the best means to judge its adequacy in a certain field; the framework for axiomatic theories in §§2 and 3 already contains a considerable number of modifications and extensions of the traditional conception that are motivated by the needs of theory construction in linguistics; (b) I would indeed submit that the conception of theories as developed in the philosophy of science, whatever its historic origin, is adequate to a significant degree for any empirical discipline, (hence, for linguistics) although there may be no single discipline yet for which it is completely adequate.118 Objection 4. The program is lop-sided: It does not take into account the non-deductive aspects of grammars and grammar writing. This objection is connected with an important complex of questions on which I can comment only briefly: If the objection is understood as establishing a contrast between deductive theories and theories allowing for probability statements it is clearly mistaken since theories of probability can also be formalized or even axiomatized. If, however, the objection refers to the role of inductive reasoning in theory construction, then it has to be admitted that the program makes no reference to it. But this is only natural: Induction plays an important part in the 117
Note that the decisive definitions all contain reference to axioms. There are good reasons for this hypothesis, and so far I have not seen any argument in favour of the singulary status of the humanities that could not be refuted or reduced to dogma. 118
Grammars as theories: the case for axiomatic grammar (part I)
97
heuristics of theory construction and the confirmation of theories, not in their actual formulation; the program is concerned only with the format of grammars. Indeed, it is one of the advantages of the proposed format that questions of heuristics and confirmation can be formukted in a well-established framework.
4.4.
Advantages of axiomatic grammar writing
Fraenkel et al. (1973, 323) make the following remark on formalization of theories (representation by formalized systems): There are very many mathematicians, and even more so other scientists, who doubt it very much whether mathematical (and other) theories should be formalized even if they can be so in principle, suspecting that the fruits of formalization are not worth the effort.
This applies directly to our program of axiomatization in linguistics, which needs to be justified. There are at least four reasons which recommend an axiomatic format for grammars as stipulated in § 4.1. Although those reasons carry weight independently of any disadvantages that may be connected with generative grammars, I will include some comparison with generative grammars in the following discussion. First advantage: The language of a grammar is a formalized system of logic. This is true for the axiomatic language (or total language) of an axiomatic theory even if a restriction to the language of predicate logic or set theory is not accepted. The system may be a natural language reading of a systern of logic, which then provides the possibility of symbolization. Thus, the linguist does not have to bother with formally specifying the language he is going to use: This task has been solved for him by the logician. All that is needed is appropriate introduction of the non-logical components: the non-logical constants and the axioms and definitions that involve such constants. Second advantage: The metalinguistic approach to the theory of language is avoided. One of the basic ideas of generative grammar can be summarized as follows: Language should be studied by specifying the form of generative grammars that characterize natural languages; more specifically, the generative grammars used by the linguist should eventually be formally constrained in such a way that they characterize precisely the natural languages. Insofar as the symbolism used in a generative grammar can be compared to a constructed language, this approach may be called metalinguistic because it directly studies the formal properties of grammars in order to indirectly arrive at properties of natural languages ("Every natural language has the property of requiring generative grammars with the following formal properties: ..."). This is at best a very clumsy procedure for developing a theory of language; at worst it is a permanent source of pseudo7 TLI1/2
98
Hans-Heinrich Lieb
problems and ad /w-decisions on matters of principle.119 The axioms and theorems of a grammar as stipulated in § 4.1 are sentences of a language which is interpreted to a degree where the axioms and theorems can be understood as assertions on the intended subject matter of the grammar. A theory of language does not deal with the linguistic properties of grammars but with arbitrary natural languages. Grammars and theories of language are rekted by the concepts of theory integration as developed in § 3.4. Third advantage: The discussion of theories in the philosophy of science can be made to bear directly on grammars™
This claim is obviously correct because of our paragraphs 2 and 3 which incorporate and develop many important features of the conception of theories found in the philosophy of science. Because our framework for axiomatic theories is sufficiently similar to that conception, the latter is relevant in still other respects. Generally, the following points can be made. (a) Heuristics and explication. It is well-known that a theory must be distinguished from the ways of arriving at a theory, which in turn can be studied systematically. Developing the heuristics of grammar writing is simplified when it can be seen as a special case of characterizing theory construction in the empirical sciences. In particular, the relation between an informal scientific grammar and a corresponding grammar that has been formulated as an applied axiomatic theory can be understood in terms of a set of simultaneous 'explications'. (b) Interpretation. Transformational generative grammars, or algorithmic grammars in general, run afoul of the problem of 'making sense of them*. The two existing attempts to solve this problem directly or indirectly make use of axiomatic theories. Thus, solving the interpretation problem for generative grammars seems to require a solution of the same problem for axiomatic theories. In § 3, such a solution was attempted, based on the relevant discussion in the philosophy of science. 119 As an example of a pseudo-problem or set of pseudo-problems, take the question of how to develop an 'evaluation measure' for grammars in terms of'formal properties of grammars. As a general requirement, evaluation is to correspond to how closely a grammar is in agreement with certain psycholinguistic assumptions, especially on language learning. Normally, this would be seen as a problem of relating sentences of the grammar to sentences of a psycholinguistic theory, regardless of the form of the sentences involved. From this point of view, the problem of finding corresponding formal properties of the grammar is spurious.—For an ad ^-decision on a matter of principle, see the way linguistics is claimed for psychology by Chomsky (cf. Lieb (1970), Ch. 10). 120 In his somewhat mistitled book, Botha (1968) made an ambitious attempt to apply to transformational generative grammars the concepts which have been developed for theories in the philosophy of science. Unfortunately, he failed to demonstrate that such grammars are theories in the required sense. Even if some of his arguments can be saved in the light of Wang's work or my own (cf. above, §§ 1.5f), this does not hold for his elaborate discussion of 'mentalism', 'competence' and related matters, which therefore is largely empty.
Grammars as theories: the case for axiomatic grammar (part I)
99
(c) Confirmation. The involved discussion of what constitutes the 'data' for a grammar and how the grammar is related to the 'data' could greatly profit from placing it into the general methodological framework for theories: Assuming that grammars are applied axiomatic theories, the discussion could be reformulated as argumentation concerning (a) the heuristics of grammar writing; (b) the interpretation of the total language and the relation between core, applied core, and application of the grammar; (c) the confirmation of grammars. (d) Explanation. Wang's recent attempts to apply the Hempel-Oppenheim schema for nomological-deductive explanation of facts to generative grammars presuppose their conversion into axiomatic theories (Wang 1972^). The value of Wang's proposals is doubtful.121 Still, if a grammar is written as an applied axiomatic theory it should be possible to give a more adequate account of what Chomsky tried to cover by his distinction between observational, descriptive, and explanatory adequacy. The first kind of adequacy concerns questions of heuristics and confirmation. The second concerns the problem of explaining facts of language use by relating them to what is stated (a) in the grammar and (b) in a psycholinguistic theory that indicates how the system described by the grammar is represented in the speaker. The third kind of adequacy concerns the problem of explaining what is stated in the grammar by relating it to what is stated in a theory of language (or language acquisition). The last two problems may perhaps be reconstructed as explanation of facts vs. explanation of laws. In any case, these and other concepts of explanation now apply directly to grammars. (e) Evaluation. An applied axiomatic theory can be evaluated by such criteria as formal simplicity, organizing power, predictive power, and 'placeability', i.e. readiness with which it is linked to other theories in the same or related fields. At present, such evaluations can hardly be more than informal estimates according to certain principles; as such, they are quite important. They may be sufficient, too, for applied axiomatic theories whose application axioms are sufficiently well understood since any such theory has a semantic basis and a pre-interpreted formalized system of logic as its total language. I suggest that problems of grammar evaluation came to occupy a prominent place in generative tranformational grammar mainly because the grammars remained semantically unspecified. Fourth advantage: Grammars can be combined with grammars and other axiomatic theories by theory use or conflation (§3.4). 121
Wang considers atomic sentences of the axiomatic grammar that assign lexemes to lexical categories as statements of facts, and universal sentences on syntactic categories as laws. Sentences of the two types are then taken as the explanans of an explanation whose explanandum is a theorem derive