Generating Language with a Phrasal Lexicon

Abstract

In this paper, we ask: How should language be represented in a generator program? In particular, how do the concepts the generator must express, the grammar it is to use, and the words and phrases with which it must express them, relate? The answer presented here is that all linguistic knowledge — all language — should be contained in the lexicon. The argument is the following: A generator performs three types of task to produce text (deciding what material to include; ordering the parts within paragraphs and sentences; and expressing the parts as appropriate phrases and parts of speech). It gets the information it requires to do these tasks from three sources: from the grammar, from partially frozen phrases (including multi-predicate phrasal patterns), and from certain words. In a functionally organized system, there is no reason why an a priori distinction should be made between the contents of the lexicon and the contents of the grammar. From the generator’s perspective, the difference between these sources is not important. Rules of grammar, multi-predicate phrases, and phrasal and verb predicate patterns can all be viewed as phrases, frozen to a greater or lesser degree, and should all be part of the lexicon. Some such “phrases” can be quite complex, prescribing a series of actions and tests to perform the three tasks: these can be thought of as specialist procedures. Others can be very simple: templates. This paper also describes the elements that constitute the lexicon of a phrasal generator program and the way the elements are used.

This paper was written while the author was at Yale University Computer Science Department, 2158 Yale Station, New Haven, CT 06520-2158, U.S.A.

This work was supported in part by the Advanced Research Projects Agency monitored by the Office of Naval Research under contract N00014-82-K-0149. It was also supported by AFOSR contract F49620-87-C-0005.