Discourse Segmentation by Human and Automated Means

Discourse Segmentation by Human and Automated Means Rebecca J. Passonneau* Bellcore and Columbia University Diane J. Litman t AT&T Labs-Research The need to model the relation between discourse structure and linguistic features of utterances is almost universally acknowledged in the literature on discourse. However, there is only weak consensus on what the units of discourse structure are, or the criteria for recognizing and gener- ating them. We present quantitative results of a two-part study using a corpus of spontaneous, narrative monologues. The first part of our paper presents a method for empirically validating multiutterance units referred to as discourse segments. We report highly significant results of segmentations performed by naive subjects, where a commonsense notion of speaker intention is the segmentation criterion. In the second part of our study, data abstracted from the subjects" segmentations serve as a target for evaluating two sets of algorithms that use utterance features to perform segmentation. On the first algorithm set, we evaluate and compare the correlation of discourse segmentation with three types of linguistic cues (referential noun phrases, cue words, and pauses). We then develop a second set using two methods: error analysis and machine learn- ing. Testing the new algorithms on a new data set shows that when multiple sources of linguistic knowledge are used concurrently, algorithm performance improves. 1. Introduction Each utterance of a discourse contributes to the communicative import of preceding utterances, or constitutes the onset of a new unit of meaning or action that subse- quent utterances may add to. The need to model the relation between the structure of such units (referred to here as discourse segment structure) and linguistic features of utterances I is almost universally acknowledged in the literature on discourse. How- ever, natural language systems rarely exploit the relation between discourse segment structure and linguistic devices because there is very little data about how they con- strain one another. We have been engaged in a two-part study addressing this gap. We report on a method for empirically validating discourse segments, and on our development and evaluation of algorithms to identify these segments from linguistic features of discourse. We show that human subjects can reliably perform discourse segmentation using speaker intention as a criterion. We also show that when multiple sources of linguistic knowledge are used (referential noun phrases, cue words, and pauses), algorithm performance approaches human performance. The excerpt in Figure 1 illustrates the two aspects of discourse that our study addresses. 2 The first pertains to an abstract structure consisting of meaningful dis- course segments and their interrelations. The utterances in segments X and Z of Fig- • 445 South Street, Morristown, NJ 07960 (E-mail: beck@bellcore.com); Department of Computer Science, New York, NY 10027 (E-mail: becky@cs.columbia.edu) 600 Mountain Avenue, Murray Hill, NJ 07974 (E-mail: diane@research.att.com) 1 By the term utterance we mean the spoken or written use of a sentence or other linguistic unit, 2 This excerpt is taken from a corpus of spoken narratives (Chafe 1980) described below. @ 1997 Association for Computational …

Content truncated to 500 words