Browse Books

Reviewer: Fred J. Damerau

This book describes work in natural language processing, now more than five years old, which investigated the feasibility of doing semantic analysis in parallel with syntactic analysis, a process called early semantic analysis. This had, of course, been attempted before, and the book comments briefly on earlier work. The study was restricted to the interpretation of noun phrases in the context of understanding mechanics problems. In this restricted domain, each noun phrase can be assumed to have a concrete referent in a model of the world, a situation which cannot be obtained in general. Even in such a restricted domain, it is often not possible to uniquely identify the referent of a noun phrase simply on the basis of local context; information from later in the same sentence or other sentences is necessary. The essential idea in this work is to separate reference identification from semantic interpretation by creating dummy referents for noun phrases, and then using an inference procedure to deduce correspondence between the concrete reference and the dummy once sufficient information has been accumulated. The representation system stores with each dummy reference a candidate set of world model referents which satisfy the stored set of constraints accumulated so far for this reference, and which satisfy the stored minimum number condition. This allows semantic information stored with the dummy reference to influence syntactic analysis before the exact world model referent is known. A number of problems arise with this approach, such as quantification of the referent sets, dependencies between partially interpreted referents, and others. These are discussed at some length in the book. Not dealt with are quantification over infinite sets, negation, and nonreferential noun phrases. The author makes no strong claims for the theoretical significance of this work. He sees it as an approach to efficiency in early semantic analysis: “All that we have managed to show is that a more considered approach to early semantic analysis, making use of partial information and avoiding arbitrary decisions, can overcome these particular search problems. . . . Unfortunately, the feasibility of early semantic analysis can only be estimated by informal and intuitive arguments about the relative efficiency of different programs.” The quotation above appears to me to characterize very well the essential point regarding mixing semantic and syntactic analysis versus a two-stage process of syntactic analysis followed by semantic analysis on the resulting parse tree. From the point of view of computational linguistics, there appears to be no theoretical import whatever regarding the method chosen; it is simply a disagreement as to which approach results in a more efficient analysis. Until we reach a stage where there are a number of similarly capable natural language processing systems, all implementing the same algorithms with about the same level of skill, it is hard to see how one could arrive at anything like a definitive resolution of this efficiency question. In the meantime, it is useful to have reasonably complete descriptions of the approaches which have been tried, so that one can form an opinion as what might be worth doing next. From that point of view, this is a useful book and can be recommended to practitioners.