File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/abstr/69/c69-1801_abstr.xml
Size: 25,129 bytes
Last Modified: 2025-10-06 13:45:44
<?xml version="1.0" standalone="yes"?> <Paper uid="C69-1801"> <Title>III-i EXTENSION AND MODIFICATION O~ TH~ THEORY</Title> <Section position="1" start_page="0" end_page="12" type="abstr"> <SectionTitle> I-I FEASIBILITY </SectionTitle> <Paragraph position="0"> This paper discusses the feasibility of applying a model of language use based on a modification and extension (to be discussed below) of the generative semantic (transformational) theory of language competence recently developed by Paul Postal, George Lakoff, John Robert Ross, ~ames D. McCawley, and others, to problems of computational linguistics.</Paragraph> <Paragraph position="1"> The theory of generative semantics, to be discussed in section II, is an outgrowth of, and reaction to, Chomsky's 1965 theory of transformational linguistics.</Paragraph> <Paragraph position="2"> It is a radical theory which deals with a very great range of problems with very abstract methods. Trose working in this paradigm hold that there is a linguistic level reflecting conceptual or semantic structure which is directly convertible into surface syntax by a single set of garden-variety transformations, with no O significant intermediary level, that is, no deep structure&quot;. These of us working in generative semantics believe that methods substantially those long familiar in linguistics can achieve very absract , very general results which treat semantics in a more serious and enlightening way than ever before. I do not, I think, support this very strong claim very well in section II, but I provide summaries of several studies and a lengthy bibliogrpahy of works which when consulted will hopefully give some feeling for what is being attempted, I think not without results.</Paragraph> <Paragraph position="3"> But generative semantics is a model, or rather, a theory, of competence, like most serious theories of language now held to by American linguists. ~ven if, as might be claimed, our semantic structures are to be merely variants of the structures long familiar from formal logic, so that if our assumptions are correct, we will ultimately be able to directly transform surface structures into underlying semantic structures, the majority of actual sentences, as well as all hypersentential structures, the treatment of which has been swept under the rug of &quot;performance&quot;, will remain unhandleable. null Accordingly, I propose initially cert@in extensions and modifications of the theory to make it in some sense a model of performance. But if we are to apply it to the computer, a major component must still be added. The impetus to this application is ~he possibility of creating an understanding machine, dewcribed in section IV below.</Paragraph> <Paragraph position="4"> Since the actual human interpretation of language depends on past knowledge (consider which of these sentences is good and why: As for Albuquerque, the ~iffel Tower is pretty.</Paragraph> <Paragraph position="5"> As for Paris, the Eiffel ~ower is pretty.</Paragraph> <Paragraph position="6"> And the se : ShirLey is a blonde and Susan is Nordic-looking too.</Paragraph> <Paragraph position="7"> Shirley is a linguist and Susan is Nordic-lloking too.) the old split between semantics, syntax, and pragmatics must be revised, and our model closely linked with a memory and possibly a logic component as well.</Paragraph> <Paragraph position="8"> Obviously this defines a very difficult task, but insofar as such goals as HT, artificial intelligence, and machine reading of handwritten material or writing of spoken material involve comprehension on the part of the machinej o~ which there seems to be no doubt, these important goals will continue to ~lude us until such time as we can devise such an understanding machine as I have ~escribed below.</Paragraph> <Paragraph position="9"> I believe that generative semantics lays the foundation for studies relevant to such a development, and it is in this context that my proposals are made.</Paragraph> <Paragraph position="10"> In section II I will d~scuss generative semantics.</Paragraph> <Paragraph position="11"> In section III I will discuss the body of my proposals here.</Paragraph> <Paragraph position="12"> In section IV I will discuss what should be required of a generalized &quot;understanding&quot; machine.</Paragraph> <Paragraph position="13"> .... tl I II \] \] ...... I \]\] Part II. The theory of Generative Semantics.</Paragraph> <Paragraph position="14"> ~ae theory of ~enerative semantics is an outgrowth and reaction to the theory of transformational grammar as represented in Chomsky's 1965 book, As s~cts of the Theor~ of S~tax (MIT Press). To a very ~-I-~ extent, this theory has been the development of a small group of former students of Chomsky,s or their close colleagues. John (HaJ) Ross has said that the theory is really Just an attempt to explicate Pa~l Postal's work of five years ago to date. If Postal was the founder of this school, if you can call it that, its main workers have been HaJ Ross and George Lakoff, who between 1965 and 1968 swept aside most of transformational linguistics as it then was. But perhaps best known of the group is J~mes McCawley, who graduated from MIT in 1965 with a Ph.D. based on work in ubono- logy, not syntax or semantics. He promptly amazed Lakoff and 2oss by some very substantive work in the latter areas as well as phonology. s~udent of McCawley's I will be emphasizing his contributions here, and those of my collea~s at Chicago, Jerry L. Morgan and Georgia M. Green, but it should be kept in mind that people like Ross, Lakoff, Postal, Arnold Zwicky, David Perlmutter, Emmon Bach, Robin Lakoff, and several others, have made the current theory possible, and that many others, such as Robert Wall, Lauri Kartunnen, Ronald Langacker, and others, have contributed as well. It should also be kept in mind that the Case Grammar of Fillmore and the work done by Gruber, while differing from generative semantics, have contributed a great deal to it.</Paragraph> <Paragraph position="15"> ~ae basic theory of generative semantics is built upon an attempt to relate the underlying semantic structure of language to the surface, phonetic manifestation of that underlying structure. That is, a phonetic reality is recognized, and a semantic reality is recognized. But unlike other versions of transformational grammar, this theory assigns no special status to syntax; syntax is subsumed in the semantics.</Paragraph> <Paragraph position="16"> McCawley has Jokingly referred to his theor~ as being one of either &quot;semantax&quot; or &quot;synantics&quot;. ~11e name generative semantics is not a particularly good one, since it implies that the ~oal of the theory is, as with the work of Chomsky, to &quot;separate the grsumuatical sequences&quot; of a language&quot;from the ~E~__ammatical other words, to generate all and only grammatical sentences of a language. ~his is not at all the goal of generative semantics. Rather, what we want to do is in some rigorous way specify the correlations of underlying semantic entities and surface phonetic entities: to specify for any underlying semantic structure what its possible phonetic realizations in some language are, and for some phonetic structure what underlying semantic structures it can represent. Naturally, so~e descriptive ability is predicated as well, that is, we want to be able to define ambiguity in some algorithmic fashion, we want to be able to define levels or classes of ill-correlation between structures on different levels, etc. Chomsky would say that a sentence~like &quot;Golf plays John&quot; is eminently deserving of a star; we would say (I) if it's supposed to mean'John plays golf', it doesn't succeed in conveying the message; (2) if it's suppesed to mean 'John loves Marsha', then it's really bad; and (3) if Golf is a man.s name and Gohn the name of a game or role, it's a good sentence --- indeed, one can very well imagine arcane circumstances under which one might utter that sentence with the intent of saying that the game plays John, that the tail wags the dog~ as it were. Suppose, for example, that John's wife were tired of him spending all his free time playing golf and she grumbled to a heighbor about it, and the neighbor rather unfeelingly replied, &quot;Oh well, John plays golf.&quot; I can ~ery well imagine John's wife complaining bitterly, &quot;Oh no, golf plays John.&quot; In any case, it is for hus unimaginative approach to language that Chomsky has been Jokingly called a &quot;bourgeois formalist&quot;o Even when we use stars, we try to keep in mind that Just about any valid phonological string of a language conveys one or more meanings in some context, and that it is artificial to take a string out of context and declare it good or bad. So &quot;generative semantics&quot; is a bad ns~e.</Paragraph> <Paragraph position="17"> The following diagram of the components of the theory is based on McCawley's paper in the proceedings of the 4th Regional Feeting of the Chicago Linguistic Society (1968). A theory very similar is discussed in Ronald Langacker's book LAnguage and its Structure (Harbrace, 1968), pp. 114-34.</Paragraph> <Paragraph position="18"> The above diagram comes from a report prepared by myself, Jerry Morgan, and Georgia Green, called the Uamelot ~o' which attempted to describe the curren-~te of rmational re, search in the Sum~uer cf 1968, particularly in reference to the LSA Summer Linguistic Institute at the University of lllinois, where HaJ Ross, George Lakcff, and Jim McCawley had lectured to large groups on a huge number of very '~airy&quot; (i.e., difficult and tickleishly novel) topics.</Paragraph> <Paragraph position="19"> In that report (which was prepared for Victor Yngve ), we raised several questions concerning the above representation. We asked: i. ~hat will an adequate semantic representation have to include? What form will it have? 2. ~hat can a transformation do? What does one lock like ? stage 3. At what A and in what manner are semantic repres@~tations converted into words of real languages? ~ese were by no means all of the questions asked.</Paragraph> <Paragraph position="20"> Needless to say, the answering cf these questions has hardly begun and will undoubtedly guarantee linguists a few gocd centuries of work at least. It is only in the last decade that syntax has been the subject of serious work, and we are still only discovering how ignorant we are. Semantics is even newer, less than a decade old. If anyone doubts that this is true, consider a) what the above 3 questions would have meant to a linguist in (say) 1955, and b)why he would have been wrong in his (lack of) comprehension of them. One of the great contributions of Postal and Ross has been II-4 their constant critical look at transformational grammar. One of the things they saw was that our transformations were (and are) extremely powerful devices, with practically no constraints placed on their formulation. ~at I will do here is summarize some of the attempts at partial answers to the three above questions. In this way I can delimit and explicate generative semantics best.</Paragraph> <Paragraph position="21"> i will start by abstracting parts of two papers by McCawley that deal with the nature of semantic representation. In a paper in the Japanese Journal ~otoba no Uchu (World of Language) in 1967, McCawley argued tha~ semantic representation would be similar to syntactic representation as familiar from ~-type grammar, but that it would also be quite similar to symbolic logic as familiar from the tons of work that have followed Principia and such studies. That semantic representation should resemble syntactic representation makes sense if only because we are arguing for a single set of rules that transforms (i.e., reEates) the underlying structure into (to) the surface structures. There will be more about that later.</Paragraph> <Paragraph position="22"> McCawley argues as follows: the following devices have all had a role in symbolic logic: I. propositional connectives&quot; 'and', 'or', 'not'. 2. constants denoting individuals.</Paragraph> <Paragraph position="23"> 3. predicates, denoting properties and relationships.</Paragraph> <Paragraph position="24"> 4. set symbols and the quantifiers 'all' and 'there exists '.</Paragraph> <Paragraph position="25"> 5. descriptions of sets and individuals.</Paragraph> <Paragraph position="26"> * x The following devices play a role in natural languages : I. all igs. have words for 'and', 'or', and 'not'.</Paragraph> <Paragraph position="27"> (he notes however that these words in natural igs.</Paragraph> <Paragraph position="28"> may connect more than sentences ) 2. &quot;indices&quot; denoting individuals; John loves John might be represented as x I loves x2, but John loves II-5 himself is x I loves x I.</Paragraph> <Paragraph position="29"> 3. predicates are expressed in natural Igs. (by verbs, adjectives, nouns, etc.) expressioms such as 4. &quot;Words such as all and&at least one are two members of a rather larg--e--clasB of expressions which are used to indicate not only the existence of an individual or a set but the absolute or relative number of members in that set.&quot; 5. sets and individuals can be expressed as descriptions using modified noun phrases.</Paragraph> <Paragraph position="30"> McCawley then gives further reasons for supposing symbolic logic representation to be proper for semantic representation. (See the bibliography to this section where t~is and other papers that can be consulted for these arguments in detail are listed. ) In a paper prepared for the symposium on &quot;Cognitive Studies and Artificial Intelligence Research&quot; held by the Wenner-~ren Foundation at the University of Chicago in March of this year, McCaWley discussed semantic representation at length. Some of what he had to say there should be noted. He claimed, &quot;semantic representation must indicate the immediate constituent structure of the elements invol~ed in it {i.e. examples showing that different meanings can comsist of the same semantic elements combined in different ways\[ are easy to come b~)&quot; {p.l) He gave the example of John doesn't beat his wife because he loves her.</Paragraph> <Paragraph position="31"> If the negation applies to John beats his wife, the se'~tence means 'the reason ~at John doesn't beat his wi~e is that he loves her', whereas if it applies to the John heats his wife because he loves her., the mg. is 'the reason that John beats his wi~e -~not- ~Hat he loves her. ' Notice that here a surface form represents at least two different underlying structures which nonetheless contain precisely the same semantic elements-- grouped differently, however.</Paragraph> <Paragraph position="32"> Another point made is that &quot;semantic representations must include .., some indication of presupposed coreference.&quot; (p.2) That is, the followingsentence in neutral (i.e. null) context is ambiguous three-ways: John told Harry that his wife was pretty.</Paragraph> <Paragraph position="33"> Whose wife? John's? Harry's? or a third's? It could be any. However, if we know who his refers to, there is no such ambiguity. This may--seem trivial, but it is a point often ignored.</Paragraph> <Paragraph position="34"> McCawley then gives an argument for referential indices being different from expressions used to de~ribe. The sentence Max d~bied that he kissed the girl Be kissed. / is not contradictory if &quot;the girl he kissed&quot; is the speaker's description.</Paragraph> <Paragraph position="35"> Another notion is that of presupposed set membero ship. lh Max is more intelligent than most Americans.</Paragraph> <Paragraph position="36"> said with primary stress ~n most, the sentence is good if and only if Max is presupp--~a to be American, that is, the sentence implies Max is American. With primary stress on Americans, however, Max is presupposed not to be American. Presupposition is in general a very hairy topic which was recently the subject of an entire conference (at the Ohio State University). We know very little about the nuamces of implication and are only beginning even to identify the problems. But if a machine is ever to rea~ Ga_tcher i~ the Rye catching all the nuances of the italicized words, we had better find out how stress is used to alter the presuppositlonal set of a sentence. I need not be so unsubtle as to suggest the extreme value of such researches to psychology. Perhaps they already know about all this, for all I know. In any case I cannot restrain myself from inclucing McCawley's beautiful example CIA Agents are more stupid than most Americans.</Paragraph> <Paragraph position="37"> He had primary streos on the ~ but I prefer to think of it as going on the Americans.</Paragraph> <Paragraph position="38"> Z would like ~o interject at this point a minor apology. I have been rather fan-clubish here and have waved my hand a lot. Frankly I see no value in rehearsing here all the arguments available elsewhere. But I would like the rea~r to bear in mind my skimpy resume in no way reflects~the quality of the original. Let me also note, lest I seem u~uduly credulous towards tjheDthoughts of. C~irman Quang 4mild-maunered linguist * . mc~awley Is In reality Q. p. Dong, Chairman of Unamerican Studies at an unknown universityJ, that most of us working within the paradigm of generative semantics would be the first to admit that our theories haven't a pra~er of being right, that is, they~~ approach even a partially realistic and naturalistic &quot;J theory of language. If we like it better than other paradigms it is because we believe that no other cureent theory is any better and that this one at least has a good chance of self-improvement. (End of apologia. ) If semantic representation looks much like logical representation, it also differs from it.</Paragraph> <Paragraph position="39"> In the Kotoba no Uchu paper McCawley noted the following differences : I. &quot;It is necessary to admit predicates which assert properties not only of individuals but also of sets and propositions.&quot; 2. &quot;In mathematics one enumerates certain objects which ~one~will talk about, defines other obJecSs in te~ms o? these objects, and co~Ifines\[onesel~ to a discussion of objects which\[oneS has either postulated or defined .... However, one does not begin a conversation by giving a list of postulates and definitions..... *..people often talk about things which either do not exist or which they have identified incorrectly* indices exist in the minds of the speaker rather than in the real world; they are conceptual entities which the individual speaker creates in interpreting his experience.&quot; In the Wenner-Gren symposium, McCawley had more to say about the difference between logic and language.</Paragraph> <Paragraph position="40"> 1. Immediate constituent structure (trees)rather than parentheses are basic. First, &quot;semantic representations are to form the input to a system of ~ransformations that relate meaning to superficial form; to the extent that these transformations have been formulated and Justified, they appear to be stateable only in terms of constituent structure and constituent type, rather than in terms of configurations of parentheses and terminal symbols.&quot; Secondly, &quot;it may be necessary to operate in terms of semantic representations in which symbols have no left-to-right ordering .... &quot; 2. There will have to be more 'logi~al operators', such as most, almost all, and m~.</Paragraph> <Paragraph position="41"> 3. &quot;And and ... or ... cannot be regarded as Just binary operators but-~ust be allowed to take an arbitrary number of operands.&quot; 4. The quantifiers must be restricted rather than unrestricted as in most logical systems. Some quantiflers imply existence: All dogs like to bite postmen, involves the presupposition that dogs exist, whereas the unrestricted quantifiers logicians use have no such presupposition.</Paragraph> <Paragraph position="42"> 5. &quot;Adequate semantic representation of sentences involving'shifters' (Jakobson, 1957) such as I, YOu~_ ~ now, ..., gestures and deictic ~6rds like this--and that, and tenses, will have to include re~rence-K~-the speech act. The most promising approach to this aspect of semantic representation ... is Rose's (1969)elaboration of Austin's (1962) notion of 'performative verb'.&quot; (See now too Searle's book, Speech A~ts, CUP, 1969 ---RIB) 6. &quot;The range of indices will ~ave to be enormous. In particular, it will have to include not only indices that purport to refer to physical objects, but also indices corresponding to mythical or literary objects, so that one can represent the meaning of sentences such as The Trobriand Islanders believe in Santa 61aus, but they call him Ubu Ubu.&quot; 7. McC. rejects &quot;the traditional distinction between 'predicate' and 'logical operator' and trea~s~ such 'logical operators' as quantifiers, conjunctions, and negation as predicates....&quot; II-9 To clarify the relationship of semantic to syntactic representations let me quote here from McCawley's Kotoba no Uchu paper: Since the rules for combining items into larger units in symbolic logic formulas must be stated in terms of categories such as 'preposition', Ipredicatel, and 'index'~ these categories can be regarded as labels on the nodes of these trees. And since ... these categories all appear to correspond to syntactic categories, the same symbols (S, V, NP, etc. ) may be used as node labels in semantic representations as are used in syntactic representations. Accordingly, semantic representations appear to be extremely close in formal nature to syntactic representations, so close in fact that it becomes possible to catalogue the conceivable formal differences and determine whether those differences are real or apparentdeg Among such differences he lists: I. &quot;The items in a s#ntactic representation must be assigned a linear order, whereas it is not obvious that linear ordering of items in a semantic representation makes shy sense.&quot; 2. &quot;Syntactic representations inwolve lexical items from the language as their terminal nodes, whereas the terminal nodes in a semantic representation are semantic units rather than lexical units.&quot; &quot;There are many syntactic categories which appear to play no role in semantic representation, for ex., verb-phrase, preposition, and prepositional phrase.&quot; (At the 5th Regional Meeting of the CLS, April of this year, A. L. Becket of the University of Michigan presented a paper in which he argued prepositions are underlying predicates; prepositional phrases are accordingly verb-phrases. ) McCawEey concluded nonetheless that these differences do not provide an argument that semantic ~epresentations are different in formal nature from syntactic representations. Again, I will omit his reasons for that conclusion. I might summarize all this by saying: i. Semantic representatio~ is a modification of the representations long familiar from ~ormal logic.</Paragraph> <Paragraph position="43"> B.</Paragraph> <Paragraph position="44"> II-lO 2. Such representations do not radically differ from the surface syntactic representations of Aspectstype grammar.</Paragraph> <Paragraph position="45"> Let me close by posing more problems. McCawley asks the following questions at the end of his Eotoba no Uchu paper. While they do not specifically reEate to semantic structure, I include them to give some idea of what we believe to be the sort of questions that a serious theory of language should prowide Justifiable answers for : I. How do the mgs. of words change as a language evolves? 2. How does a child learn rags. in learning to speak his native language? 3. ~at mechanisms are involved in phenomena such as metaphor .... ? (Dorothy Lambert has written a Ph.D. thesis at Michigan on the subject of metaphor within the paradigm of Case Grammar.</Paragraph> <Paragraph position="46"> This 500 page dissertation is probably one of the best studies of the subject to date from a linguistic point of view.S--RIB) 4. To what extent are the units of semantic representatiomq univers el ? 5. To what extent does the lexicon of a language have a structure? 6. Can all languages express the same ideas? 7. To what extent doe's one's language affect his thinking? 8. To what extent is one's ability to learn lexical items conditioned by his knowledge of the world? I will now turn to the second question raised above on p. II-3o This question has as yet received little study. It is a very difficult topic, but a very important one. I will confine myself here to a few brief comments and a few references.</Paragraph> </Section> class="xml-element"></Paper>