Most words have more than one meaning. Yet natural language contains enough redundancy so that, in context, the appropriate meaning of each word is clear. The disambiguation process is so natural that most of the time we are not aware that it is taking place at all until we are confronted with a pathological example (usually either in the form of a comedian’s pun or a linguist’s numbered example). The objective of Hirst’s research (originally a Ph.D. project under Charniak at Brown) is to construct a working computer model of how semantic interpretation takes place. Hirst describes his work as cognitive science, not just AI. Though he is primarily interested in the AI agenda of solving the problem, he devotes a good deal of effort to understanding the underlying psycholinguistic mechanisms and seeking to clarify them by constructing computational mechanisms that share their characteristics.
Hirst’s project builds on the ongoing natural language research at Brown, and in particular incorporates the Frail frame language and the Paragram parser. Upon this foundation (modified as required for the present project), he constructs a semantic interpreter and implements mechanisms for both lexical and structural disambiguation.
The heart of the system is Absity, a semantic interpreter that coordinates semantic and syntactic interpretation in tandem so that each contributes to the other. Absity relies on strongly typed lexical and semantic objects, and this and its tandem character lead Hirst to characterize it as “Montague-inspired.” It is strongly compositional, combining lower-order semantic objects to form higher-order ones. Absity leans heavily on the parallel between the linguistic model of a verb as a nexus of cases and the AI formalism of a frame as a collection of slots. For example, prepositions are construed as slot names for the associated verbal frame, and syntactic functions that are marked in English by position rather than prepositions receive pseudoprepositions to fit them into this structure. Hirst exploits this parallel between cases and slots for practical reasons, but it reflects theoretical needs not otherwise met in most current frame systems [1].
Two mechanisms address lexical disambiguation: marker passing and Polaroid Words. As the frame system encounters each word in the sentence, it marks the corresponding frame, which in turn propagates marks to connected frames. Converging markers define a path between concepts and thus identify an association that can often disambiguate a word. The system includes appropriate restrictions on marker passing to guard against combinatorial explosion. A Polaroid word is a process with the responsibility of disambiguating a single word. Like a Polaroid photograph, such a process “develops” the meaning of its word over time. Originally, it records all possible meanings, and it eliminates impossible ones as it receives information from other words. Even before development is complete, the rest of the system can manipulate the Polaroid Word and consult the partial conclusion that it contains as an aid to resolving other questions in the system.
The provisions for structural disambiguation are rather less developed. The parser interrogates a process known as the Semantic Enquiry Desk (SED) to resolve two kinds of structural ambiguity: prepositional phrase attachment and gap finding in relative clauses. The SED in turn uses the Polaroid Word mechanism to investigate the consequences of tentative decisions.
Potential readers should note at least two issues in Hirst’s work that reach beyond the interests of natural language researchers.
The notion that syntactic and semantic interpretation must be conducted in tandem with one another is not new, but Hirst provides a particularly clean architecture for providing this coordination while at the same time keeping the two roles cleanly modularized. Central to his strategy is the principle that intermediate results are well-formed objects and can be queried even before they are fully resolved (the Polaroid Words metaphor). This principle is likely to have broad application in distributed AI for many problem domains other than language. Sociologically, one of the things that distinguishes interactions in a team from hierarchical or competitive interaction is the availability of information about partial results or unresolved options. Hirst provides a clear example of how this can be done, at least in one domain.
Hirst is particularly forceful and articulate in his insistence that objects in a linguistic knowledge base are more than just symbols; they are semantic objects that correspond to real things “out there” in the real world. He contrasts this position with one that attempts to achieve intelligent behavior simply by manipulating symbols. Once one has raised the question of the relation between computer bits and external reality, it is difficult to avoid going on to engage the view, articulated for the AI community by Winograd and Flores [2], of whether there really are things “out there,” or whether the categories into which people divide the world are rather imposed on a continuum by the observers. It is not reasonable to expect Hirst to resolve this question, but his position puts him squarely in the midst of the debate, and the reader’s position in that discussion will certainly affect the acceptability of Hirst’s model.
The book is written in engaging, entertaining prose, with references through the year of publication and thorough indices by name and by subject, and it is not marred by obvious typographical errors. The typography is clear, with consistent use of varying typefaces. As an exemplar of the “published dissertation” genre, the book is of primary interest to other researchers in the field. While this work is not a textbook, it does give fairly detailed introductions to its major ideas, and it concludes with a set of exercises, which range in complexity from a reasonable weekly assignment in a seminar to a dissertation topic. Thus it would be more useful than many other published dissertations in a graduate seminar on natural language processing. Overall, it is a responsible piece of creative research, clearly presented and attractively packaged.
