Bidirectional grammars and the design of natural language generation systems

Abstract

Intuit ively considered, a grammar is bidirectional if it can be used by processes of approximately equal computat ional complexity to parse and generate sentences of a language. Because we, as computat ional linguists, are concerned with the meaning of the sentences we process, a bidirectional grammar must specify a correspondence between sentences and meaning representations, and this correspondence must be represented in a manner tha t allows one to be computed from the other. Most research in computat ional linguistics has focused on one or the other of the two sides of the problem, with the result tha t relatively little a t tent ion has been given to the issues raised by the incorporation of a single grammar into a system for tasks of both comprehension and generation. Clearly, if it were possible to have t ruly bidirectional grammars in which both parsing and generation processes were efficient, there would be some compelling reasons for adopting them. First , Occam's razor suggests tha t , if language behavior can be explained by hypothesizing only one linguistic representation, such an explanation is clearly preferable to two tha t are applicable in complementary circumstances. Also, from the practical s tandpoint of designing systems tha t will carry on sophisticated dialogues with their users, a single unified formalism for specifying the syntax and semantics of the language is likely to result in a simpler, more robust implementation. The problems of mainta ining consistency between comprehension and generation components when one of them changes have been eliminated. The lexicon is also simpler because its entries need be made but once, and there is no problem of mainta ining consistency between different lexical entries for unders tanding and generation. It is obvious tha t not all grammars are bidirectional. The most fundamental requirement of any bidirectional grammar is tha t it be represented declaratively. If any information is represented procedurally, it must of necessity be represented differently for parsing and generation processes, resulting in an asymmetry between the two. Any change in the grammar would have to be made in two places to mainta in the equivalence between the syntactic and semantic analyses given to sentences by each process. A grammar like DIAGRAM [8] is an example of a grammar for which the encoding of linguistic information