Foreword

Yorick Wilks,     New Mexico State University

This is an important book, one that focuses on problems in language understanding by computer that many, in traditions other than that of the present editors, would prefer to forget.

They are close to a tradition (though they may not agree with this characterization!) I shall call “computational semantics” in a sense fairly well understood in artificial intelligence and computational linguistics but one not beyond dispute by other users of the term “semantics.” I shall just state the sense intended, because my purpose here is not to defend the use, which has been done often enough, but to use it to introduce the chapters in this book.

What is “Computational Semantics?”

As is often the case, a position can best be described by contrasting it with others, and in this case I can contrast Computational Semantics usefully with three positions: syntactic analysis, logical semantics, and expert systems. There are refinements and subtleties in all those positions but, speaking broadly, Computational Semantics is opposed to any claims as to the necessity or sufficiency, for computational language understanding, of logical semantics or syntactic analysis. To say that is not to deny, in certain cases, the usefulness of a syntactic analysis, nor the aesthetic pleasures of a subsequent formal axiomatization. Computational Semantics holds, however, that, for the solution of practical and enormous problems like lexical ambiguity resolution, the real theories of language understanding, which is to say, the substantive algorithms and the right level of analysis, lie elsewhere.

The comparison with natural language interfaces to expert systems is rather different: There exists a position, for which it is difficult to give textual sources (in a way that it is all too easy for the above two positions), which claims that, if the appropriate “expert” knowledge structure is adequate, then practical problems of language semantics (lexical ambiguity, etc.) do not arise since the knowledge base contains only appropriate word senses. This claim is simply false, unless (a) the domain chosen is trivial, or (b) the domain chosen includes language itself, as in Small’s Word Expert parsing system [Chapter 1, this volume]. However, to say that is not to mark computational semantics off from knowledge realms and their formal expression: on the contrary, it is to suggest that knowledge of language and “the world” are not separable, just as they are not separable into databases called, respectively, dictionaries and encyclopedias.

Computational Semantics and Connectionism

No introduction to anything these days would be complete without some reference to connectionism: the current cluster of artificial intelligence theories around the notion of very simple computing units, connected in very large numbers, and “learning from experience” by means of shifting aggregated weights in a network. This development may offer a way forward in many areas of artificial intelligence, including computational semantics. Two of the chapters in this book follow this line; Kawamoto’s excellent model of the acquisition of lexemes and Cottrell’s model of lexical access. Connectionism shares many of the differences that computational semantics has with the approaches noted above: an emphasis on the integration of semantics and syntax, continuity between linguistic and other forms of world knowledge, and a type of inference that is not reconcilable with the kind offered by logic-based approaches. Moreover, connectionism has stressed notions such as that of active competition between representational structures, with the tendency of the more strongly connected representations to “win,” a notion to be found explicitly in computational semantics systems such as Preference Semantics, and Small’s Word Expert approach [Chapter 1, this volume].

An important difference, as regards lexical ambiguity resolution in particular, arises here between so-called sub-symbolic approaches within connectionism [Smolensky 1988] and those usually called localist [Cottrell, 1985; Waltz and Pollack, 1985]. This difference, which is yet in no way settled, bears very much on the subject matter of this book: in a sub-symbolic approach to computational semantics one would not necessarily expect to distinguish representations for particular word-senses, they would be simply different patterns of activation over a set of units representing sub-symbolic features, where similar senses would lead to similar patterns. On the other hand, localist approaches [Waltz and Pollack, 1985; Cottrell, 1985] to computational semantics have assumed real distinguished word senses in their symbolic representations at the outset and have then given weighting criteria for selecting between them.

The difference between these two notions of connectionism, as they apply to issues of word sense in computational semantics, captures precisely the issue at the heart of lexical ambiguity resolution—namely, are there, in any real sense, discrete word senses, or is it all just a matter of fuzzy, numerical-boundary classifications of individual examples of use? That this issue is a serious one can be seen from the consequence that, if there is no real basis, psychological or computational, to word-sense classes, then there is no topic to be called lexical ambiguity resolution, and this book should be on the same library shelf as alchemy, phrenology, and the Greek myths. Before going on to reassure the reader on this issue, let me complete the historical introduction by setting the contrasts above in an earlier context.

Some Brief History of Computational Semantics

What is the historical origin of computational semantics, as defined above? In some broad sense nearly all the work in the early 1970s on artificial intelligence approaches to natural language understanding can be said to fall under the description computational semantics: the two approaches that fit most straight forwardly were Schank’s Conceptual Dependency [1975] and Wilks’s Preference Semantics [1975], but in a yet broader sense Winograd’s SHRDLU [1972] was also within the description given above, since his own emphasis at the time was not on reference (to the limited number of blocks on his simulated table) but on “procedural semantics.” Very much the same can be said of Simmons’s networks [1973] and Charniak’s inferential system [1972].

Versions of these systems all still appear in systems under development today: the emphasis in the Conceptual Dependency group has shifted from describing their work as natural language processing to describing it as human memory research, but the change is largely cosmetic, since at Yale the approach is still applied to machine translation [see Lytinen, Chapter 4, this volume] and database front-ends, among other things. A recent paper by Lehnert [1986], although its title suggests a continuing concern with memory, is very much a return to early Conceptual Dependency concerns and is certainly computational semantics: in fact a fusion of Conceptual Dependency and Preference Semantics approaches.

More recent work in the computational semantics tradition has been concerned with the central issue, isolated above and discussed below, of the discreteness of word-senses (as opposed to a model of continuity between them, as connectionism prefers) and the question of how, if distinctions are to be made between senses, this can be done at the appropriate time, as information enters the system, rather than on the basis of early all-or-nothing commitment, in the way earlier computational semantics systems all did. Key work here has been by Small [Adriaens and Small, Chapter 1, this volume], Hirst [Chapter 3, this volume] and Mellish [1985].

The emphasis there is on the boundaries between word-senses, continuity and vagueness [Lytinen, Chapter 4, this volume] and the ways in which sense decisions can be made appropriately if the phenomena are less discrete than was once supposed. These are the real concerns of human and machine understanders, and ones about which logic-based semantics has little or nothing to say.

Are Word-Senses Real?

Let us now, in conclusion, review the issue itself rather than its research history, and do so by asking the basic question this way: is it right to assume a notion of “word sense,” unexamined, and direct from traditional lexicography? What evidence have we that a serious computational semantics can be based on such a notion?

To put the matter another way: Many have claimed that the inability of programs to cope with lexical ambiguity was a major reason for the failure of early computational linguistics tasks like machine translation. Yet, does it follow from that that the lexical ambiguity distinguished by conventional dictionaries, or even LDOCE (Longman’s Dictionary of Contemporary English, [Proctor 1979]) has any real significance for that task? LDOCE is worth a special mention here. It is a real dictionary, on sale as a book, but which has also, in electronic form, been used by computational linguists (e.g., [Wilks et al., 1987]) as a knowledge base for English. It is peculiarly suitable for this because it is written complete with fine...