Elamur Research⌕
☰
home / papers / 2003 / Discourse Segmentation of Multi-Party Conversation

Discourse Segmentation of Multi-Party Conversation

Summary

The paper addresses topic segmentation of multi-party meetings. This setting has been studied far less than written text or monologue, and it is harder: speech is informal, often ungrammatical, and speakers overlap. The authors use 25 meetings from the ICSI Meeting corpus, each annotated for topic boundaries by at least three people. They propose a two-part segmenter. The first part, LCseg, is a lexical-cohesion algorithm. It builds chains of repeated stemmed terms, weights them, and finds points where cohesion between neighbouring windows drops sharply. The second part adds conversational cues: cue phrases, silences, overlapping speech and changes in speaker activity. A C4.5rules classifier merges these cues with LCseg's output. LCseg alone is competitive with leading lexical segmenters. The combined system clearly reduces error on meetings (Pk 23.00% vs. 31.91% for LCseg alone), and the gain is statistically significant. The work shows that what is said and how the conversation unfolds give complementary evidence about topic shifts in dialogue.

Key points

  • Data: ICSI Meeting corpus, 25 meetings of about 60 minutes on average, with roughly 6.5 participants each.
    • Every speaker change was a candidate boundary, about 770 per meeting.
    • The majority-vote reference has about 7.5 segments per meeting.
    • Under Cochran's Q, inter-annotator agreement was significant at the 0.05 level for 19 of the 25 meetings.
  • LCseg:
    • It uses only term repetition, ignoring synonymy and other semantic relations.
    • A chain is split whenever the term is absent for h consecutive sentences.
    • Each chain is scored as frequency × log(L / chain length).
    • Cohesion at each break is the cosine similarity between two adjacent windows, using the chain scores as weights.
    • On concatenated WSJ stories it reaches an average Pk of 15.31 when the number of segments is unknown.
  • Conversational features: CUE, SIL (gaps), OVR, ACT (speaker activity, compared via divergence between distributions) and LC (both LCseg's posterior and the raw cohesion function). The best analysis window per feature was chosen automatically, for example 30 s on the left side for silence and 5 s on both sides for cue phrases.
  • Evaluation: 25-fold cross-validation, training on 24 meetings and testing on the held-out one.
    • Combined system: Pk 23.00% and WD 25.47%.
    • LCseg alone: Pk 31.91% and WD 35.88%.
    • U00 (Utiyama and Isahara): Pk 37.39% and WD 40.43%.
    • The combined system's improvement is significant (p ≈ 2e-4).
  • Learned rules: lexical cohesion was the most influential feature, but the combination beat either knowledge source used alone.

Interesting ideas

  • I1. Prefer compact chains: Chains that are short but dense get more weight, because a term spread across the whole document says little about where boundaries lie. This is useful in any cohesion-based segmentation, including for Ukrainian texts.
  • I2. Break chains at long gaps: Splitting a repetition chain when the term disappears for a while keeps weak, long-distance links from blurring topic structure. It is a cheap trick for any lexical-chain method.
  • I3. Pick each feature's form by its solo performance: Rather than hand-tuning window size and direction, each candidate version of a feature is tested alone, and the best predictor of boundaries is kept. This simple, principled way to set feature parameters carries over to many sequence-labeling tasks.
  • I4. Feed one model's output into another: A complete unsupervised segmenter becomes a single input to a supervised classifier, contributing both its probability estimate and its raw signal. This gives hybrid systems that keep the strengths of a robust baseline.
  • I5. Set the boundary count adaptively: When the true number of segments is unknown, low-scoring candidates are discarded first. The remaining ones are kept if they exceed the mean minus a multiple of the standard deviation. This is a portable heuristic for choosing peaks or valleys in noisy score curves.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Natural language processing

Topics: Discourse analysis

Cited by

104 papers of our library cite it (of the 140884 whose bibliographies we read); 2 of them have a page here, 102 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.3115/1075096.1075167

@inproceedings{galley2003discourse,
  title = {{Discourse Segmentation of Multi-Party Conversation}},
  author = {Galley, Michel and McKeown, Kathleen R. and Fosler-Lussier, Eric and Jing, Hongyan},
  booktitle = {Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics},
  year = {2003},
  doi = {10.3115/1075096.1075167},
  url = {https://doi.org/10.3115/1075096.1075167}
}

References

27 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. D. Beeferman, A. Berger, and J. Lafferty. 1999. Statistical models for text segmentation. Machine Learning, 34(1–3):177–210. find
  2. F. Choi. 2000. Advances in domain independent linear text segmentation. In Proc. of NAACL’00. find our card
  3. W. Cochran. 1950. The comparison of percentages in matched samples. Biometrika, 37:256–266. find
  4. B. Grosz and J. Hirschberg. 1992. Some intonational characteristics of discourse structure. In Proc. of ICSLP-92, pages 429–432. find
  5. B. Grosz and C. Sidner. 1986. Attention, intentions and the structure of discourse. Computational Linguistics, 12(3). find our card
  6. M. Hajime, H. Takeo, and O. Manabu. 1998. Text segmentation with multiple surface linguistic cues. In COLING-ACL, pages 881–885. find
  7. M. Hearst. 1994. Multi-paragraph segmentation of expository text. In Proc. of the ACL. find
  8. J. Hirschberg and D. Litman. 1994. Empirical studies on the disambiguation of cue phrases. Computational Linguistics, 19(3):501–530. find
  9. J. Hirschberg and C. Nakatani. 1996. A prosodic analysis of discourse segments in direction-giving monologues. In Proc. of the ACL. find
  10. J. Hirschberg and C. Nakatani. 1998. Acoustic indicators of topic segmentation. In Proc. of ICSLP. find
  11. A. Janin, D. Baron, J. Edwards, D. Ellis, D. Gelbart, N. Morgan, B. Peskin, T. Pfau, E. Shriberg, A. Stolcke, and C. Wooters. 2003. The ICSI meeting corpus. In Proc. of ICASSP-03, Hong Kong (to appear). http://www.cs.columbia.edu/˜galley/research.html ↗
  12. M.-Y. Kan, J. Klavans, and K. McKeown. 1998. Linear segmentation and segment significance. In Proc. 6th Workshop on Very Large Corpora (WVLC-98). find
  13. H. Kozima. 1993. Text segmentation based on similarity between words. In Proc. of the ACL. find
  14. S. Levinson. 1983. Pragmatics. Cambridge University Press. find
  15. D. Litman and R. Passonneau. 1995. Combining multiple knowledge sources for discourse segmentation. In Proc. of the ACL. find
  16. J. Morris and G. Hirst. 1991. Lexcial cohesion computed by thesaural relations as an indicator of the structure of text. Computational Linguistics, 17:21–48. find
  17. C. Nakatani, J. Hirschberg, and B. Grosz. 1995. Discourse structure in spoken language: Studies on speech corpora. In AAAI-95 Symposium on Empirical Methods in Discourse Interpretation. find
  18. R. Passonneau and D. Litman. 1993. Intention-based segmentation: Human reliability and correlation with linguistic cues. In Proc. of the ACL. find
  19. R. Passonneau and D. Litman. 1997. Discourse segmentation by human and automated means. Computational Linguistics, 23(1):103–139. find
  20. L. Pevzner and M. Hearst. 2002. A critique and improvement of an evaluation metric for text segmentation. Computational Linguistics, 28 (1):19–36. find our card
  21. R. Quinlan. 1993. C4.5: Programs for Machine Learning. Machine Learning. Morgan Kaufmann. find our card
  22. J. Reynar. 1994. An automatic method of finding topic boundaries. In Proc. of the ACL. find
  23. J. Reynar. 1999. Statistical models for topic segmentation. In Proc. of the ACL. find
  24. G. Salton and C. Buckley. 1988. Term weighting approaches in automatic text retrieval. Information Processing and Management, 24(5):513–523. find
  25. G. Tür, D. Hakkani-Tür, A. Stolcke, and E. Shriberg. 2001. Integrating prosodic and lexical cues for automatic topic segmentation. Computational Linguistics, 27(1):31–57. find
  26. M. Utiyama and H. Isahara. 2001. A statistical model for domain-independent text segmentation. In Proc. of the ACL. find
  27. J. Xu and B. Croft. 1998. Corpus-based stemming using cooccurrence of word variants. ACM Transactions on Information Systems, 16(1):61–81. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/galley-2003-discourse-segmentation-multi-party-conversation/