Book Reviews: Reversible Grammar in Natural Language Processing

This is a collection of 15 revised and expanded papers from the 1991 Berkeley ACL workshop of the same title. Collectively, the contributions cover all the main theoretical and practical motivations for reversibility of grammars: increased accuracy of description gained; economy of development and maintenance effort; consistency of language use in translation or interface applications; and possibility of self-monitoring for ambiguity in generation or paraphrase of parsing input. Some other connected issues are also discussed. "A reversible constraint-based logic grammar," by Palmira Marrafa and Patrick Saint-Dizier, describes a formalism derived based on Ait-Kaci's psi-terms, and shows how the operation of type construction can be used both to generate and to parse types representing sentences and their meanings. Some illustration of the formalism is given by applying it to problems of Portuguese syntactic description. Marc Dymetman, in "Inherently reversible grammars," distinguishes different ways in which "reversibility" can be understood. He focuses on a notion of "inherent finite reversibility" of grammars: a property possessed by a grammar if it is possible to derive sound, complete, and terminating parsing and generation algorithms for it. He discusses some grammars of artificial languages that do not satisfy this condition and tries to define a property of "moderation" that distinguishes the two classes of grammars. The intuitive content of this is something like compositionality. Gunter Neumann and Gertjan van Noord ("Reversibility and self-monitoring in natural language generation") point out that a reversible grammar is necessary for self-monitoring for ambiguity in generation, or for paraphrasing of ambiguity when parsing. They describe an algorithm for carrying out both of these tasks, illustrated mostly with examples of PP-attachment ambiguity. I found this one of the most interesting papers in the volume. However, there is an important practical issue not discussed by the authors that arises when such techniques are used in practical applications, namely that in many cases multiple readings are not distinguishable by a nonexpert. For example, each of the following has at least two parses but most nonlinguists (and some linguists) are incapable of discerning any difference in meaning: