Discourse Segmentation of Multi-Party Conversation

Discourse Segmentation of Multi-Party Conversation Michel Galley Kathleen McKeown Columbia University Computer Science Department 1214 Amsterdam Avenue New York, NY 10027, USA {galley,kathy}@cs.columbia.edu Eric Fosler-Lussier Columbia University Electrical Engineering Department 500 West 120th Street New York, NY 10027, USA fosler@ieee.org Hongyan Jing IBM T.J. Watson Research Center Yorktown Heights, NY 10598, USA hjing@us.ibm.com Abstract We present a domain-independent topic segmentation algorithm for multi-party speech. Our feature-based algorithm com- bines knowledge about content using a text-based algorithm as a feature and about form using linguistic and acous- tic cues about topic shifts extracted from speech. This segmentation algorithm uses automatically induced decision rules to combine the different features. The em- bedded text-based algorithm builds on lex- ical cohesion and has performance compa- rable to state-of-the-art algorithms based on lexical information. A significant er- ror reduction is obtained by combining the two knowledge sources. 1 Introduction Topic segmentation aims to automatically divide text documents, audio recordings, or video segments, into topically related units. While extensive research has targeted the problem of topic segmentation of written texts and spoken monologues, few have stud- ied the problem of segmenting conversations with many participants (e.g., meetings). In this paper, we present an algorithm for segmenting meeting tran- scripts. This study uses recorded meetings of typi- cally six to eight participants, in which the informal style includes ungrammatical sentences and overlap- ping speakers. These meetings generally do not have pre-set agendas, and the topics discussed in the same meeting may or may not related. The meeting segmenter comprises two compo- nents: one that capitalizes on word distribution to identify homogeneous units that are topically cohe- sive, and a second component that analyzes conver- sational features of meeting transcripts that are in- dicative of topic shifts, like silences, overlaps, and speaker changes. We show that integrating features from both components with a probabilistic classifier (induced with c4.5rules) is very effective in improv- ing performance. In Section 2, we review previous approaches to the segmentation problem applied to spoken and written documents. In Section 3, we describe the corpus of recorded meetings intended to be seg- mented, and the annotation of its discourse structure. In Section 4, we present our text-based segmenta- tion component. This component mainly relies on lexical cohesion, particularly term repetition, to de- tect topic boundaries. We evaluated this segmenta- tion against other lexical cohesion segmentation pro- grams and show that the performance is state-of-the- art. In the subsequent section, we describe conver- sational features, such as silences, speaker change, and other features like cue phrases. We present a machine learning approach for integrating these con- versational features with the text-based segmenta- tion module. Experimental results show a marked improvement in meeting segmentation with the in- corporation of both sets of features. We close with discussions and conclusions. 2 Related Work Existing approaches to textual segmentation can be broadly divided into two categories. On the one hand, many algorithms exploit the fact that topic segments tend to be lexically cohesive. Embodi- ments of this idea …

Content truncated to 500 words