An Orthonormal Basis for Topic Segmentation in Tutorial Dialogue

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 971–978, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics An Orthonormal Basis for Topic Segmentation in Tutorial Dialogue Andrew Olney Zhiqiang Cai Department of Computer Science Institute for Intelligent Systems University of Memphis University of Memphis Memphis, TN 38152 Memphis, TN 38152 aolney@memphis.edu zcai@memphis.edu Abstract This paper explores the segmentation of tutorial dialogue into cohesive topics. A latent semantic space was created using conversations from human to human tu- toring transcripts, allowing cohesion be- tween utterances to be measured using vector similarity. Previous cohesion- based segmentation methods that focus on expository monologue are reapplied to these dialogues to create benchmarks for performance. A novel moving window technique using orthonormal bases of se- mantic vectors significantly outperforms these benchmarks on this dialogue seg- mentation task. 1 Introduction Ever since Morris and Hirst (1991)’s ground- breaking paper, topic segmentation has been a steadily growing research area in computational linguistics, with applications in summarization (Barzilay and Elhadad, 1997), information retrieval (Salton and Allan, 1994), and text understanding (Kozima, 1993). Topic segmentation likewise has multiple educational applications, such as question answering, detecting student initiative, and assess- ing student answers. There have been essentially two approaches to topic segmentation in the past. The first of these, lexical cohesion, may be used for either linear segmentation (Morris and Hirst, 1991; Hearst, 1997) or hierarchical segmentation (Yarri, 1997; Choi, 2000). The essential idea behind the lexical cohesion approaches is that different topics will have different vocabularies. Therefore the lexical cohesion within topics will be higher than the lexi- cal cohesion between topics, and gaps in cohesion may mark topic boundaries. The second major ap- proach to topic segmentation looks for distinctive textual or acoustic markers of topic boundaries, e.g. referential noun phrases or pauses (Passonneau and Litman, 1993; Passonneau and Litman, 1997). By using multiple markers and machine learning methods, topic segmentation algorithms may be developed using this second approach that have a higher accuracy than methods using a single marker alone (Passonneau and Litman, 1997). The primary technique used in previous studies, lexical cohesion, is no stranger to the educational NLP community. Lexical cohesion measured by latent semantic analysis (LSA) (Landauer and Du- mais, 1997; Dumais, 1993; Manning and Schütze, 1999) has been used in automated essay grading (Landauer, Foltz, and Laham, 1998) and in under- standing student input during tutorial dialogue (Graesser et al., 2001). The present paper investi- gates an orthonormal basis of LSA vectors, cur- rently used by the AutoTutor ITS to assess student answers (Hu et al., 2003), and how it may be used to segment tutorial dialogue. The focus on dialogue distinguishes our work from virtually all previous work on topic segmen- tation: prior studies have focused on monologue rather than dialogue. Without dialogue, previous approaches have only limited relevance to interac- tive educational applications such as intelligent tutoring systems (ITS). The only existing work on topic segmentation in dialogue, Galley et al. (2003), …

Content truncated to 500 words