Probabilistic Grammars for Natural Languages

Abstract

Although a fully adequate grammar for a substantial portion of any natural language does not exist, a vigorous and controversial discussion of how to choose among several competing grammars has already developed. On occasion, criteria of simplicity have been suggested as systematic scientific criteria for selection. The absence of such systematic criteria of simplicity in other domains of science inevitably raises doubts about the feasibility of such criteria for the selection of a grammar. Although some informal and intuitive discussion of simplicity is often included in the selection of theories or models in physics or in other branches of science, there is no serious systematic literature on problems of measuring simplicity. Nor is there any systematic literature in which criteria of simplicity are used in a substantive fashion to select from among several theories. There are many reasons for this, but perhaps the most pressing one is that the use of more obviously objective criteria leaves little room for the addition of further criteria of simplicity. The central thesis of this paper is that objective probabilistic criteria of a standard scientific sort may be used to select a grammar.

This research has been supported in part by the National Science Foundation under Grant NSFGJ-443X.