The Syntax, Semantics and Inference Mechanism in Natural Language

The Syntax, Semantics, and Inference Mechanism of Natural Language

Francis Y. Lin Computational Linguistics Unit School of English and Philosophy University of ~¥ales Cardiff

Abstract depends on its syntax. Some KR formalisms also have a semantic component, which determines what a well- It is both desirable and plausible to treat natu- formed formula means and how meaning is ascribed to ral languageitself as a "knowledgerepresentation it. If natural language is a formalism for representing (KR) formalism. Every KRformalism has syntax and support certain inferences. The syntax of a knowledge, then its syntax provides specifications of KRformalism specifies the form in which knowl- well-formed formulae. In addition, reasoning must be edge must be encoded, and its inference mecha- performed directly on the basis of syntax, as in any nism depends on its syntax. If natural language artificial KRscheme. This is a point that has been ne- is a formalism for representing knowledge,then glected so far. If we apply this requirement to existing its syntax provides specifications of well-formed formulae. In addition, reasoning must be done accounts for syntax, we will easily find they are inade- directly on the basis of syntax, as in any arti- quate, that is, existing syntactic rules are not designed ficial KRscheme. In this respect, existing syn- to support inference and it is doubtful that they can tactic theories are inadequate, because syntax in be used directly to perform reasoning. What we need these theories does not support reasoning. In this then is a new conception of syntax of natural language, paper I present a new conception of the syntax, semantics, and inference mechanismof natural and a corresponding new conception of semantics and language. The central ideas are (1) that worduse reasoning. determines the forms of sentences, and (2) that these forms are used to express meaningand per- Syntax and Word Use form inference. Syntax, semantics, and inference are seen as an integrated whole, rather than sep- Ordinary speakers can potentially recognise and pro- arate and autonomousfields as in manyexisting duce an infinitely large number of sentences. This studies of natural language. phenomenon is called the competence of the language speaker (Chomsky, 1957). As with Chomsky, this phenomenon can only be accounted for by assuming that Introduction the language speaker utilises a finite set of grammar Existing knowledge representation (KR) systems rules which are capable of generating an infinite set of not use natural language directly or fully. The com- sentences. One major goal of generative grammar is mon practice is to translate natural language into an to account for this phenomenon. A generative gram- artificial formalism, on which subsequent operations mar typically consists of a set. of abstract rules and arc based. There have been attempts to make KR principles, from which all sentences in a language can formalisms close to natural language to gain wider lin- be derived. These rules and principles are so abstract guistic coverage or to obtain more efficient reasoning that they can hardly be said to be knownto the ordi- (Barwise L: Cooper, 1981; Kayser, 1987; Alshawi, 1992; nary speaker in any real sense. So they are assumed Iwanska, 1992). But, since people use natural language to be ’tacitly ~ knownto the speakers, or even ’innate" to exchange information, to gain new knowledge, to ex- in their genes (Chomsky, 1988). press thought, to perform reasoning, and so on, it is not Chomsky’s conception of grammar rules is in di- only desirable but also plausible to treat natural lan- rect contrast to that of Wittgenstein. According to guage itself as a KR formalism. Every KR formalism the later Wittgenstein, language consists of language has syntax and supports certain inferences. The s3/ntax games which are played according to rules (Wittgen- of a KR formalism specifies the form in which knowl- stein, 1972). Rules must be able to be expressed and edge must be encoded, and its inference mechanism they must be transparent to participants in a rule-

i01 governed practice (Baker &: Hacker, 1985, pp. 62-3). (15) It’s kind of somebody to do something. In this section I outline an account of grammarwhose (16) *It’s hungry of somebodyto do something. rules are actually knownto the speakers. I shall start It is easy to see that usaxes, as those illustrated, are with the notion of sentence structure and try to expli- general w~vs of how words are used and that they are cate the notion of grammar via some examples. finite in number. Usaxes are grammarrules, according Consider the following sentence: to which we recognise sentences and determine their (1) John kissed Mary in front of Peter oll purpose, grammaticality. We have learned them, and they are because he wanted to annoy him. knownto us. In the rest of this section I shall show The structure of the sentence, in my analysis, is: how this finite number of explicit grammar rules can (2) somethingx happened because vsomething generate an infinite number of sentences. happened1 Usaxes are concrete, in the sense that they are ex- where somethingx happened refers to: pressible in natural language. On the other hand they (3)2 John kissed Mary in front of Peter on purpose. are also general. The more general a usax is the more This sentence in turn has a structure, which is: sentences it can cover, that is, it can generate. For (4) somebody did something on purpose. example, there exists an increasing order of generality where somebody is John, somebody did something is: amongthe following sentences: (5) John kissed Mary in front of Peter. (5) < (6) < (17) < (18) The structure of sentence (5) is: where (17-19) are: v.(6) solnebodyx did something in front of somebody (17) Somebody did something. where somebody~ did something refers to: (18) Something happened. (7) John kissed Mary. (19) Somethingis the case. whose structure is: (19) covers events, states, properties, and whatever you v.(8) somebody~ kissed somebody have. In fact (19) is the most general form of sentences It can be seen froln the above example that what I (or usaxes). call structures are expressions of word use. For exam- The process of generalisation starts, as illustrated pie, (2, 4, 6, and 8) show how the words (and phrases) above, from any sentence and ends in the most gen- ’because’, ’on purpose’, ’in front of’, and ’kiss’ are used; eral form, which is (19). What is the converse pro- and one only needs a little reflection to confirm that cess? It is an easy matter to see that it is generation, this is the case. I shall therefore call a structure like (2) 3 in the Chomskyan sense. For example, one can be- a usax, and say that a word is used grammatically if it gin with (19), through (18), (17), and (6), conforms to its usax (or one of its usaxes if it has more with (5). 4 Notice that each general usax can generate than one). We can now have a better idea about the a large number of usaxes which are less general: and notions of sentence structure and can appreciate the the latter are capable of generating a great manyeven equivalence relationship between the following propo- less general usaxes. Note also that each least general sitions: usax can still generate a large numberof concrete sen- (9) A string of words is a sentence. tences. Through this process a grammar can generate (10) The string of words has a sentence structure. very many sentences. To ensure that a grammar can (11) The sentence is gralnmatical. generate an infinite numberof sentences, recursion will (12) All the words have been used grammatically. have to be build into it. But I shall say no more about A few other examples may help to clarify the no- this here. tions of structure, usax, and grammaticality. (13) To sum up, usaxes are grammarrules, they are finite a grammatical sentence, it conforms to the structure and are knownto the speakers. Through the process of (15), which is a usax for ’kind’. (14) is ungralnmatical, generalisation, one gets more and more general usaxes because there is no such usax as (16): and through the converse process, which is generation, (13) It’s kind of you to come to see me. one produces (or can produce) all the sentences in (14) *It’s hungry of you to come to see me. language. 11 use somebodyxto meanthat there is an entity x which is referred to by the word ’somebody’.The referents will 4But one does not need to always start from (19) be left out whenthere are no ambiguities present. generate a sentence. One can start from anywhere in the 2For simplicity the analysis of the second half of (1), chain of generation. Note also that one does not ahvays which is ’he wanted to annoyhim’, will be omitted here. need to reach the end of the generation process either. One 3This notion is employedto stress the relationship be- can stop in the midway,to arrive at certain sentences: e.g. tween the use of words and the syntax of sentences. (17).

102 Semantics and Inference ing roles to objects mentioned in a sentence. Another Semantics is the study of meaning. My aims in this problem is that roles are usually not capable of cap- turing enough, let alone complete, meaning. Consider section are to present my view on meaning, and to provide justification for it. I shall show that the the- ’John dried the shirt’ for instance, the roles can not ory developed here subsumes several major semantic capture the meaning such as ’the shirt was wet and theories. it is now dry’. Viewed from the vantage point of the current theory, roles are merely short-form for some (a Wittgenstein rejects the idea that meaning is some- subset) of the meaning-bearing sentences, e.g. ’John thing abstract., something like a ’shadow’ of words and did the kissing’, ’Mary was kissed’. Thus, in a certain sentences (Wittgenstein, 1972, PP. 117&120). He says sense role-based theories can be seen as a simplified that to know tlle meaning of a word/sentence we must version of the present theory, in which only certain look at how it is used and in what situation it is types of inference are in operation (see next section). used (Wittgenstein, 1972, PP. 43&592). Tile meaning of a word/sentence is the explanation we give (in A third major semantic theory utilises the notion each of the concrete situation in which it is used). In of truth conditions, the most important version being Wittgenstein’s words, ’meaning is what an explanation Davidson’s. Davidson explicates the meaning of a sen- of meaning explains’ (Wittgenstein, 1974, p. 69). Now, tence in terms of its truth conditions, conditions un- how is explanation done? As one cannot, simply mem- der which the sentence is true (Davidson, 1967, 1970). orise all sentences (in order to recognise or produce As Dmnmett (1976) points out, the theory does not them), one cannot simply memorise the meaning of all explain how the speaker can know those truth condi- sentences (in order to explain the meaning}. To know tions. It also implies that the ordinary speaker masters the meaning of a sentence, which is one of a potentially the predicate-calculus and the translation rules which infinite set, we inust perform some inferences. paraphrases the ordinary sentences into their David- I submit that tile meaning of a word/sentence is sonian replacements, which is both ’unrealistic’ and the set of sentences which can be inferred from the ’unnecessary’ (Strawson, 1992, p. 103). It is not diffi- cult to see that, like the roles-based theories, the truth word/sentence, and I shall also call the latter the meaning-bearing sentences of the former. By inference conditions of a sentence are a subset of its meaning- (or reasoning) I mean not only formal inference (e.g. bearing sentences. deduction), but also inforlnal or practical reasoning Moravcsik (1990) contends that ’the meaning of (e.g. speech acts), and inference of other sorts. Thus word "w" is that in virtue of which an element of re- semantics in the present framework is compatible with ality counts as a w’. For example, the meaning of ’wa- Lyons’ linguistic semantics (Lyons, 1995). According ter’ is that it is a clear, odourless liquid, and is used to this conception, there is no meaning in isolation: for drinking, and irrigation, etc. The present account meaning arises from interconnections (Strawson, 1992; largely agrees with Moravcsik’s view of meaning. What Quine & Ulian, 1978). he regards as tile meaning of a word is effectively the One influential semantic theory is based on semantic set of sentences we can infer from the word. These primitives (I

103 ference, though they may be unaware of this. In other u(25) There is something, in something words, the theories I have just studied only employ (26) Somethingv has something, in it. certain types of inference. I shall call equivalent sentences such as (23-26) meaning equivalences. This type of inference will be called Inference in Natural Language equivalisation. Having established that the meaning of a sentence is The third type of commonlyused inference is dele- the set of sentences which can be inferred from it, the tion, which basically means that we can infer from task of semantics then becomes the task of explaining some sentences some of their parts. The following sen- how inference is done. In parallel to the conception tence is a deletion rule: of grammar presented in Section 2 I shall provide an (27) If you did something at a certain time, in account for semantics based on a finite number of in- certain place, in a certain way, or for a certain reason. ference rules which are knownto the speakers etc., then you did it. Inference of meaning can take manyforms (or types), The fourth type of inference is what I shall call ex- four of which will be discussed here. I shall first exposition. It is used to expose the semantic roles (some- plicate what the four types of inference are by giving times called case roles or participant roles) in sen- examples, and then briefly compare lny treatment of tences. Below are a few examples of exposition rules: them with solne other approaches. The most common (28) If x did y, then what x did is to do type of meaning inference is performed through the (29) If x did y, then what (or the person who) meaning of words (and phrases), which speakers are yisx. aware of and are able to explain. Manynew words are (30) Ifx did y to z, then what (or/the person whom) learned also in this way, i.e. through their explanation x did y to is z. in terms of other words (Johnson-Laird. 1983, oh. 10). It is easy to see that all these four types of inference For example, we know: are based on usaxes: Given any sentence, we can ab- (20) If you do something on purpose, you do it de- stract its usax. This usax will match one (or more) liberately. semantic rules, and corresponding inferences will be (21) If you do somethixlg in front of somebody, you obtained. It is here that grammar elucidated in Sec- do it when he is present. tion 2 and semantics discussed in this section come to (22) If you kiss somebody, you touch him with your contact with each other, and the interaction is very lips. smooth. I shall call sentences like (20-22) meaningelucidations, It is also here that we see more clearly the relation- following (Higginbotham, 1989). 5 It is readily seen ships between my account of meaning and other seman- that meaning elucidations also have a structure (e.g. tic theories. By using elucidation and equivalisation we If x then y), which consists of structures of the con- can replace theories based on semantic primitives. By stituent sentences (compare (20-22) with (4, 6, 8)). using exposition we can replace semantic roles. And we apply the meaning elucidations (20-22) to sentences by using deletion we can replace at least the part of (4, 6, 8) respectively, then we should be able to infer Davidson’s theory which accounts for the derivation of the corresponding sentences. I shall use the term elu- (32) from (31): cidation to refer to this type of inference. (31) John kissed Mary in front of Peter. Another commonway of explaining the meaning of (32) John kissed Mary. a sentence is to use its equivalent or near-equivalent In doing so, we also avoid many fundamental problems sentences. For example, (23) is equivalent to (24), inherent in those theories, which were mentioned in the (25) meanstile saine as (26): previous section. (23) You gave something to somebody. (24) You gave somebody something. Conclusions and Discussions It is bot.h desirable and plausible to use natural lan- 5This is not to advocate a ’dictionary’ view of meaning, which implies that meaningis fixed. Accordingto Wittgen- guage as a means of representing knowledge and per- stein (Wittgenstein, 1972, P. 456), if somebodyjust said forming reasoning. This requires that the syntax of kissed Mary"and I nowsimply mimichis voice and also say natural language should specify the format in which ’I kissed Mary’,then I didn’t really kiss Mary(i.e. touch her knowledge is encoded, and that reasoning should op- with my lips). But this is due to the meaningof the word erate directly on the basis of syntax as in any artificial ’inimic’. It. is true that the meaningof words/sentencesde- pends on the situations in whichthey are used, but we must KR scheme. Existing accounts of syntax are inade- knowthe meaning of the words/sentences which describe quate, because the syntax developed so far is separate such situations and perform corresponding inference. from and cannot support reasoning. A new conception

104 of syntax, semantics, and reasoning is needed. Dummett, M. (1976). What is a theory of meaning? In this paper I presented a new way of looking at the (ii). Reprinted in his The Seas of Language, syntax, semantics and inference mechanism of natural Clarendon Press, Oxford, 1993. language. The basis of investigation is word use. In Fillmore, C. (1968). The case for case. In Bach, E., learning a language we learn not only words, but also & Harms, R. T. (Eds.), Universals in Linguistic how they are used and what they mean. The rules of Theory. Rinehart and Winston, New York. grammar, meaning, and inference are closely related to word use and they are known to ordinary speakers. In Haegeman, L. (1994). Introduction to Government and order to understand these notions I have taken neces- Binding Theory (2rid edition). Blackwell, Ox- sary excursions into linguistics and philosophy. What ford. was achieved was a unified account for these notions. Halliday, M. A. K. (1994). Introduction of Functional As I argued, usaxes are grammar rules, they deter- Grammar (revised edition). Edward Arnold, mine sentence recognition and generation. They are London. the forms in which meaning is expressed, and reasoning is performed. Thus, syntax, semantics, and rea- Higginbotham, J. (1989). Elucidations of meaning. soning form an integrated whole, rather than separate Linguistics and Philosophy, 12, 465-517. and autonomous components of natural language. Iwanska, L. (1992). A general semantic model of nega- Further work is still needed to develop the present tion in naturM languages: representation and in- account of language into a satisfying theory. One direc- ference. In Proceedings of the Third Conference tion is to explain within this framework more syntactic on the Principles of Knowledge Representation phenomena,e.g. passivisation, reflexives, etc, as dealt and Reasoning. with in Government and Binding Theory (Haegeman, 1994). Another direction is to explore further the in- Jackendoff, R. (1976). Towards an explanatory seman- ference process based on usaxes (see Lin (1996)). tic representation. Linguistic Inquiry, 7, 89-150. Johnson-Laird, P. N. (1983). Mental ,l,lodels. Cam- Acknowledgements bridge University Press, Cambridge. I am greatful to Sir Peter Strawson for discussions, Katz, J. (1972). Semantic Theory. Harper & Row, eon]ments, and encouragement. All errors are mine. New York. Reference Kayser, D. e. a. (1987). A strategy for reasoning Alshawi, H. (Ed.). (1992). The Core English Engine. natural language. Applied Artificial Intelligence. MIT Press. 1, 205-231. Liu, F. Y. (1996). On the notion of logical form. Com- Baker, G. P., & Itacker, P. M. S. (1985). Wittgenstein: putational Linguistics Unit, School of English Rules. Grammarand Necessity. Blackwell, Ox- and Philosophy, University of Wales Cardiff. ford. Manuscript. Barwise, J., & Cooper, R. (1981). Generalized quanti- Lyons, J. (1995). Linguistic Semantics. Cambrige Uni- tiers and natural language. Linguistics and Phi- versity Press, Cambridge. Iosophy, 4, 159-219. Moravcsik, J. M. E. (1990). Thought and Language. Chomsky, N. (1957). Syntactic Structures. Mouton, Routledge, London. The Hague. Quine, W. V. O., & Ulian, J. (1978). The Web of Belief (rev. edition). RandomHouse, New York. Chomsky. N. (1988). Language and Problems of IO~owledge. The MIT Press, London. Strawson, P. F. (1992). Analysis and Metaphysics: An Introduction to Philosophy. Oxford University Davidson, D. (1967). Truth and meaning. Reprinted Press, Oxford. in his hlquiries into Truth and Interpretation, Clarendon Press, Oxford, 1984. Wittgenstein, L. (1972). Philosophical Investigations (3rd edition). Blackwell, Oxfford. Davidson. D. (1970). Semantics for natural languages. Wittgenstein, L. (1974). Philosophical Grammar. Reprinted in his Inquiries into Truth and Inter- Blaekwell, Oxford. Rhees, R. (ed.). pretatiou. Clarendon Press. Oxford. 1984.

105