Canonical Representation in NLP System Design: A Critical Evaluation

CANONICAL REPRESENTATION IN NLP SYSTEM DESIGN= A CRITICAL EVALUATION Kent Wittenburg and Jim Barnett MCC 3500 West Balcones Center Drive Austin, TX 78759 ABSTRACT This paper is a critical evaluation of an approach to con- trol in natural language processing systems which makes use of canonical structures as a way of collapsing multiple analyses in individual components. We give an overview here of how the Lucy natural language interface system cur- rently realizes this control model and then evaluate what we take to be the strengths and weaknesses of such an approach. In particular, we conclude that the use of canonical struc- tures can restrain combinatorial explosion in the search, but at the cost of breaking down the barriers between modules and of letting processing concerns infect the declarative representation of information. 1 Introduction The traditional design for natural language processing systems is one in which processing proceeds a sentence at a time, with syntactic analysis feeding subsequent semantic and discourse analysis in a "conduit" fashion, to borrow a characterization used in a somewhat different setting by Ap- pelt (1982). The basic advantages of this design stem from the fact that it is inherently modular: control is simple, modules can be developed and debugged independently. The ma~n disadvantage from the processing point of view is that the search can explode as each module detects ambiguities that cannot be resolved until later. 12 Although well-known alternatives to conduit models exist -- the most obvious being the proposal to interleave syntax, semantics, and discourse processing -- we sense that the simplicity and modularity of some form of the conduit model continue to be the determin- ing design factor in most applied natural language systems to date, at least for those that can be said to have independent modules to start with. In this paper we will discuss the pros and cons of a design paradigm that stays within the basic conduit model. It is a paradigm characterized by the at- tempt to procrastinate the resolution of ambiguity by means of representing families of analyses with canonical represen- tations. The discussion will appear as follows. In Section 2 we at- tempt to define and justify the use of canonical represen- tations, highlighting their appeal from a processing point of view. We then give a brief overview of how this paradigm has been applied in the natural language interface prototype 1For example, Martin, Church, and Patil (1981) mention that their phrase structure grammar produced 958 parses for the naturally occurring sentence In as much a8 allocating costs i8 a tough job [ would like to have the total co~ts related to each product. 2Appelt (1982) mentions other problems with the conduit model having to do with its inability to account for inter- actions, .say, between linguistic choices and gestural ones in language generation models. called Lucy (Rich et al. 1987), Our main points appear in Section 4, where we assess the consequences of these design decisions for three of the Lucy system modules. The …

Content truncated to 500 words