Elamur Research⌕
☰
home / papers / 1998 / Computing Dialogue Acts from Features with Transformation-Based Learning

Computing Dialogue Acts from Features with Transformation-Based Learning

Paper · 1998 · Ken Samuel; Sandra Carberry; and K. Vijay-Shanker

Summary

The paper tackles automatic recognition of dialogue acts (labels such as SUGGEST, REJECT, ACCEPT) that capture what a speaker intends by an utterance, often beyond its literal meaning. Earlier machine-learning attempts relied mostly on preceding act tags, and a few on words, length and speaker direction, using Semantic Classification Trees or smoothed N-Grams. Their success was modest: no reported system topped 75% accuracy. The authors adapt Brill's Transformation-Based Learning (TBL), which had worked well for part-of-speech tagging, to this task. TBL starts from a crude default labelling and then learns an ordered list of correction rules drawn from human-written templates. To make the huge rule space tractable, they use a Monte Carlo strategy that samples candidate rules rather than enumerating them all. They also add richer utterance features, including automatically generated cue phrases that a human can prune. Preliminary tests on the VerbMobil spoken-dialogue corpora are promising. The work matters because it brings a transparent, rule-producing learner to discourse-level interpretation.

Key points

  • TBL training loop: assign an initial tag, generate template-based rules that fix the current errors, score each by its net accuracy gain over the whole corpus, keep the best one, apply it, and repeat until the gain drops below a threshold.
  • Prior systems surveyed used SCTs or N-Grams with deleted interpolation, on German, English and Japanese dialogues; the authors judge task type and feature choice to matter more than the learning algorithm.
  • Following Ramshaw & Marcus, TBL can exploit tags it has already computed and resists overtraining better than decision trees. Its drawback is that it needs hand-crafted rule templates.
  • Monte Carlo sampling of the rule space substantially sped up training without losing accuracy. Evaluation used the VerbMobil corpora, with preliminary results only.
  • Proposed extensions:
    • confidence scores from agreement among several trained models (Committee-Based Sampling);
    • weakly-supervised learning that seeds unsupervised TBL with high-confidence tags;
    • rules restricted to backward context for live dialogue;
    • incremental training;
    • focus tracking.

Interesting ideas

  • I1. Learning proceeds like painting with progressively finer brushes: a sweeping default label comes first, then ever narrower correction rules, each one readable and applied in a fixed order.
  • I2. Instead of scoring every possible rule instantiation, the learner randomly samples candidates from the template space, which cuts cost sharply without hurting accuracy.
  • I3. Confidence is derived from how strongly several independently trained models agree on a tag, which gives a rule learner the uncertainty estimate it otherwise lacks.
  • I4. High-confidence automatic tags are treated as the "unambiguous" seeds that bootstrap learning on untagged data.
  • I5. Learning two rule sets, one using only past context and one refining once later utterances arrive, lets the same model work both incrementally and offline.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (frames): speech-act labelling in prag could start from a default act per clause type and then apply a learned, ordered list of readable correction rules, keeping every decision explainable.
  • I2 (learn): self-induced repair rule mining over large silver corpora could sample rule candidates rather than enumerate them, making each induction round much cheaper.
  • I3 (learn): agreement among witness parsers or among several repair-rule lists can serve as a per-item confidence score, filtering which silver examples enter training.
  • I4 (learn): silver trees or frames with full witness agreement could seed further rule induction on raw text, while disputed items are held back for teacher tiers.
  • I5 (frames): speech-act and implicature decisions could be made provisionally while reading and then revised by a second rule pass once the following sentences are known.

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

 Institutions

🎓University of DelawareNewark · United States20 works · 1,056 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 2Research subjects and languages the work is about.Dialogue systems · 624 worksSpeech act theory · 229 works

Cite

Cite as Ken Samuel; Sandra Carberry; and K. Vijay-Shanker. 1998. Computing Dialogue Acts from Features with Transformation-Based Learning. Applying Machine Learning to Discourse Processing: Papers from the 1998 AAAI Spring Symposium.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.cmp-lg/9806002 · arXiv cmp-lg/9806002

@inproceedings{carberry1998computing,
  title = {{Computing Dialogue Acts from Features with Transformation-Based Learning}},
  author = {Ken Samuel; Sandra Carberry; and K. Vijay-Shanker},
  booktitle = {Applying Machine Learning to Discourse Processing: Papers from the 1998 AAAI Spring Symposium},
  year = {1998},
  doi = {10.48550/arxiv.cmp-lg/9806002},
  eprint = {cmp-lg/9806002},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.cmp-lg/9806002}
}

References

18 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Alexandersson, J.; Maier, E.; and Reithinger, N. 1995. A robust and efficient three-layered dialogue component for a speech-to-speech translation system. In Proceedings of the European Association for Computational Linguistics. ↗
  2. Andernach, T. 1996. A machine learning approach to the classification of dialogue utterances. In Proceedings of NeMLaP-2. ↗
  3. Brill, E. 1995a. Transformation-based error-driven learning and natural language processing: A case study in part-of-speech tagging. Computational Linguistics 21(4):543–566.
  4. Brill, E. 1995b. Unsupervised learning of disambiguation rules for part of speech tagging. In Proceedings of the Very Large Corpora Workshop. find
  5. Chen, K.-h. 1995. Topic identification in discourse. In Proceedings of the Seventh Meeting of the European Association for Computational Linguistics, 267–271. find
  6. Dagan, I., and Engelson, S. P. 1995. Committee-based sampling for training probabilistic classifiers. In Proceedings of the Twelfth International Conference on Machine Learning, 150–157. ↗
  7. Hinkelman, E. A. 1990. Linguistic and Pragmatic Constraints on Utterance Interpretation. Ph.D. Dissertation, University of Rochester, Rochester, New York. Technical Report UR CS TR #238. find
  8. Hirschberg, J., and Litman, D. 1993. Empirical studies on the disambiguation of cue phrases. Computational Linguistics 19(3):501–530.
  9. Dialogue Act INFORM SUGGEST SUPPORT INFORM SUGGEST ARGUE Lambert, L. 1993. Recognizing Complex Discourse Acts: A Tripartite Plan-Based Model of Dialogue. Ph.D. Dissertation, The University of Delaware, Newark, Delaware. Technical Report #93-19. find
  10. Mast, M.; Niemann, H.; Noeth, E.; and Schukat-Talamazzini, E. G. 1995. Automatic classification of dialog acts with semantic classification trees and polygrams. In Proceedings of the Fourteenth International Joint Conference on Artificial Intelligence, 71–78. New Approaches to Learning for Natural Language Processing Workshop. find
  11. Nagata, M., and Morimoto, T. 1994a. First steps toward statistical modeling of dialogue to predict the speech act type of the next utterance. Speech Communication 15:193–203. find
  12. Nagata, M., and Morimoto, T. 1994b. An information-theoretic model of discourse for next utterance type prediction. Transactions of Information Processing Society of Japan 35(6):1050–1061. find
  13. Ramshaw, L. A., and Marcus, M. P. 1994. Exploring the statistical derivation of transformation rule sequences for part-of-speech tagging. In Proceedings of the 32nd Annual Meeting of the ACL, 86–95. Balancing Act Workshop. find
  14. Reithinger, N., and Klesen, M. 1997. Dialogue act classification using language models. In Proceedings of EuroSpeech-97, 2235–2238. find
  15. Reithinger, N., and Maier, E. 1995. Utilizing statistical dialogue act processing in verbmobil. In Proceedings of the 33rd Annual Meeting of the ACL, 116–121. ↗
  16. Reithinger, N.; Engel, R.; Kipp, M.; and Klesen, M. 1996. Predicting dialogue acts for a speech-to-speech translation system. Technical Report Verbmobil-Report 151, DFKI GmbH Saarbruecken. ↗
  17. Samuel, K. B. 1996. Using statistical learning algorithms to compute discourse information. Technical Report #97-11, The University of Delaware. Dissertation proposal. find
  18. Samuel, K. 1998. Lazy transformation-based learning. In Proceedings of the Eleventh International Florida Artificial Intelligence Research Symposium Conference, 235–239. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/samuel-1998-computing-dialogue-acts-features-transformation/