Text Generation: The Problem of Text Structure

摘要

One of the key problems in Al text organization. A poorly organized text can be unreadable or even misleading. Two Al approaches to text organization are compared: McKeown's TEXT system and Rhetorical Structure Theory (RST). They share many assumptions about the nature of text yet they are found to be in strong contrast. TEXT identifies text organization with whole-text nonrecursive structures, while RST uses small recursive ones. RST has an elaborate apparatus of relations between parts of texts, and of the 'nuclearity' of particular parts; TEXT has no correlates for these. RST works with a wide range of relation types, TEXT with just one of them. TEXT is an implemented system, whereas RST is developmental. Most important, TEXT develops text organizations so that they resemble patterns extracted from previous text, while RST strives for an organization which is justifiable as meeting the goals of the text being generated. This contracts raises many of key issues discussed in this paper about the nature of text organization and how it can be created by programs. (Author)