Semantic interpretation and
the resolution of ambiguity
by Graeme Hirst
reviewed by
Peter Norvig
Computer Science Div.
University of California
Berkeley CA 94720
norvig@cogsci.Berkeley.EDU
Hirst's book is based on his 1983 doctoral dissertation. Readers who are familiar with the dissertation will find that the book version follows the same layout, chapter for chapter, sub-section for sub-section. The material is revised, but the book does not present significant new research results beyond the thesis work. However, Hirst does a good job of discussing the related work that has appeared in the intervening four years, and he adds a new chapter of unanswered questions, speculations, and exercises for the reader. Some of the jokes have been removed, and the result is perhaps a little more ``scholarly'' in tone, while still maintaining an interesting and amusing style.
Hirst says his work is ``Montague-inspired,'' and he uses Transformational Grammar, but nothing rests on the details of either the grammar or the semantic representation language. His arguments should be applicable to a wide range of theoretical orientations, and should not be dismissed out of hand by those who count themselves in different camps.
Hirst divides the problem of semantic interpretation into three parts, which are reflected both in the organization of the book, and in the computer implementation he has built. First he addresses the overall problem of interpretation, laying out six desirable qualities. He describes his system, named Absity, which he claims meets five of the six criteria. Absity uses strict composition of well-formed semantic objects. It is driven by syntax, and maintains a strict correspondence between syntactic and semantic types, but it allows for interaction between parser and interpreter.
The second part of the book is devoted to lexical disambiguation. This part, like the other two, consists of two chapters, the first describing the problem, and the second outlining a solution. The review of the psycholinguistic literature in the ``problem'' chapter is particularly good. Hirst's solution is to use ``polaroid words:'' fake semantic objects which eventually develop into real ones. I have always found the concept of Polaroid words to be an apt and compelling metaphor, but Hirst should make clearer the distinction between metaphor and implementation. As a technical term ``Polaroid words'' is equivalent, as far as I can tell, to the less evocative term ``disjunction.'' (``Developing'' a polaroid word is equivalent to eliminating a disjunct, and sharing of pictures is equivalent to the destructive unification commonly used in unification grammars.)
Part three addresses structural disambiguation. Again, the introduction to the problem is good, with many interesting examples, but the solution is less satisfying. Hirst proposes a ``semantic enquiry desk'' which can be asked questions about the existence or plausibility of proposed interpretations. Unfortunately, the implementation only handles two types of ambiguity: prepositional phrase attachment, and gap finding in relative clauses. Furthermore, as the name ``semantic enquiry desk'' suggests, Hirst takes a wholly semantic view of structural disambiguation, ignoring the fact that certain syntactic constructions may be more frequent than others, or that pragmatic factors may influence the interpretation.
For me, the most serious problem is a lack of commensurability, and indeed a lack of interaction between the various interpretation and disambiguation components. For example, the table on page 174 gives a decision algorithm for the attachment of a restrictive PP. The algorithm combines the principles of referential success, plausibility, verb expectations and reference failure avoidance, but it does so in a completely serial, binary fashion. The principles are strictly ordered in importance, and there is no way for them to interact, nor any way to express, for example, degrees of plausibility. Furthermore, there is no provision for interaction between two different ambiguities as posed to the semantic enquiry desk (although there is a mechanism for allowing polaroid words to communicate with each other). Recent work in semantic interpretation such as Charniak (1988) and Hobbs (1988) addresses just these problems; it is not clear how limited Hirst's approach is without these capabilities, nor how easy it would be extend his approach.
Hirst is unusually candid in characterizing what his system can and cannot do, but he does not do a very good job at showing us examples, or providing statistics. It is certainly useful to know that Absity can do X and Y but not Z, but the reader would get a better feel for the strengths and weaknesses of the system if Hirst would add ``... and it successfully interprets N% of a 100-sentence corpus of text, taken from...''
While the level of argumentation is usually high, there were a few logical breakdowns. One that particularly irks me is found on page 135. In discussing the ambiguity of the attachment of the adverb ``quickly'' in (1) below, Hirst uses the ungrammaticality of (2) to argue that the ``type-quickly'' reading of (1) must also be ungrammatical.
(1) A good secretary can type quickly written reports.
(2) * A good secretary can type quickly reports.
The faulty argument goes as follows:
``The error is that in general an adverb may not be placed between a verb and its object NP; any adverb in such a position must in fact be part of the NP, if that is at all possible, and sentences in which it is not, such as (3), are at best marginally well formed:
(3) *?Nadia closed rapidly the stopcock.'' (Hirst p.135-136) |
Of course, the preconception that any NP can be substituted for any other without affecting grammaticality is simply false, as Ross (1973) has convincingly shown. Ross shows many examples where a heavier NP (like ``written reports'') allows or prefers preposed modifiers where a lighter NP (like ``reports'') takes a postposed modifier.
In each of the book's three parts, I find myself preferring the ``problem'' chapter over the ``solution'' chapter. The discussion of the various problems tend to be broad yet reasonably comprehensive, with good examples chosen for illustration. However, Hirst often jumps from the data to a solution without adequately considering other possible solutions. This is appropriate for a Ph.D. thesis, but is less so for a book. As a result, the reader gets the impression that Hirst knows the subject matter very well, but the conclusions can only be accepted as a matter of faith in his experience. For example, on page 43 Hirst justifies his choice of a compositional system by saying ``Compositionality is clearly a desideratum.'' While this may be largely true, there are also cases (idioms, metaphors, certain grammatical constructions) where non-compositionality is a desideratum. Hirst, not unreasonably, chooses a completely compositional approach, but he does not tell us how to handle the difficulties, nor does he tell us what an alternative choice might entail.
The end result is that the book gives the reader an excellent introduction to some problems in semantic interpretation and a good description of one approach at the problems, but no compelling arguments for necessarily accepting that approach.
Reference
Ross, J. R., (1973) ``Nouniness,'' in O. Fujimura, ed., Three Dimensions of Linguistic Theory. TEC, Tokyo.