Book Reviews: The Formal Complexity of Natural Language

To provide some details, Cleopatra's parsing is "lexically driven" in that "each word in the input sentence invokes procedures that direct the parsing process". Samad observes that "we can invoke any arbitrary function when evaluating a constraint". Similarly, the integration procedures that build meaning structures may also involve "arbitrary Lisp functions". Cleopatra attacks ambiguity by use of "confidence levels" related to such phenomena as "the relative frequencies of occurrence of different senses of a word, the likelihood of particular structure% and the correspondence of conjuncts". Parsing is viewed as a parallel process, but the author clearly and honestly states that the existing program is breadth-first. While the author claims to have achieved a "sharp contrast to the limited domains of previous natural language interfaces", I cannot concur. Although he does handle certain issues in greater detail than have most previous systems (for example,, time references and some types of conjunction), there are also, as he so often and honestly admits, many typical distinctions that have been largely ignored (apparently, without suffering a great loss). For example, the system "is not very sophisticated about auxiliaries or adverbs" and does not yet "know about" person or number. Samad suggests that the "final determinant" of a system with the practical aspirations of his is the evaluation it receives from its intended users. But his claim that "Cleopatra is more than a vehicle to demonstrate the feasibility of our ultimate goal. It is a useful CAD tool" is simply not substantiated. Although "we are confident that experimental studies will confirm the utility of Cleopatra", there's no indication that such investigations have been conducted. In conclusion, the author is to be commended for his interest in building a complete system that can be used for some meaningful purpose. His frank and honest discussion of the details and limitations of his work are also to be praised, and the presence of some inherently interesting example sentences in the domain under study should be mentioned. However, I find little in the book that helps clarify any problems of language processing, nor do I suspect the techniques presented can provide any "value added" over what's available from existing literature. I am also disappointed by the fact that most of the book concerns implementational issues discussed at the level of data structures. Although I cannot recommend the book as a text, it could certainly be found useful as a case study.