Discourse as a Knowledge Resource for Sentence Disambiguation

Abstract

A consistent text contains rich resources of information, such as collocation patterns , that can be used to resolve ambiguities within it s sentences. For exa mple, attachment ambiguities in a sentence ca n be resolved by selecting a candidate attachment that matches attachments found in other sentences in the same discourse. Thus, discourse can be r ega rded as a valuable knowledge resource for sentence analysis. In this paper, we examine some feat ur es of discourse as a knowledge res our ce, and propose a framework for natural language processing that provides a simple algorithm for using information extracted from discourse together with information stored in knowledge bases. The ex perimental results of using our fram ework to disambiguat e sentences in technical documents offer good prospects for improving the accuracy of a broad-coverage natural language processing system that handles various texts without constructing knowledge bases for each text in advance. Some noteworthy features of discourse information are also deduced from the res ults of our experiments.