Elamur Research⌕
☰
home / papers / 2018 / A Context-based Approach for Dialogue Act Recognition using Simple Recurrent Neural Networks

A Context-based Approach for Dialogue Act Recognition using Simple Recurrent Neural Networks

Paper · 2018 · Chandrakant Bothe, Cornelius Weber, Sven Magg et al.

Summary

The paper addresses dialogue act (DA) recognition: labelling each utterance in a conversation as a statement, question, backchannel, answer, and so on. The authors observe that most earlier classifiers look at one utterance in isolation. Many DA classes are short and lexically ambiguous, so this fails. A bare "yeah" can be a Backchannel, a Yes-Answer or an Agree/Accept depending on what came before. After reviewing how DAMSL-style corpora (Switchboard, MRDA) are annotated, they propose a context-based model. Each utterance is encoded by a pre-trained character-level mLSTM language model trained on out-of-domain text. A simple RNN then runs over the current utterance and a few preceding ones to predict the current label. Evaluated on the Switchboard Dialogue Act corpus (SwDA), the model reaches 77.3% accuracy. The prior context-based state of the art was 73.9%, and adding context gives roughly 3 points over their own utterance-only baseline. Because the model uses only past utterances, it fits real-time dialogue systems, unlike models that also see future turns.

Key points

  • Corpus: SwDA (SWBD-DAMSL tag set), with the MRDA annotation scheme also analysed; the reported inter-annotator agreement on SwDA is 84%.
  • Utterance encoder: a single-layer multiplicative LSTM (mLSTM) with 4,096 hidden units, trained as a character language model by Radford et al. on a corpus of roughly 80 million reviews of Amazon products.
  • Two utterance vectors are compared: the hidden state at the last character ("feature") and the mean of the hidden states over all characters ("average").
  • The baseline is an MLP with softmax over single utterances. The context model is an RNN fed the representations of utterances t−2, t−1 and t.
  • Earlier results cited for comparison:
    • HMM with n-grams (Stolcke et al.): 71%
    • Naive Bayes: 66%
    • HMM (Tavafi et al.): 74.32%
    • Lee and Dernoncourt, CNN/RNN: 73.1%
    • Kalchbrenner and Blunsom: 73.9%
    • This work: 77.3%

Interesting ideas

  • I1. Short, ambiguous utterances like "yeah" get their function from the previous turn, so the label should be predicted from a small window of preceding turns rather than the utterance alone.
  • I2. The model restricts context to past turns only, which keeps it usable online; models that see the whole conversation score high partly by peeking at the future.
  • I3. A frozen, generic character-level language model trained on unrelated text served as a reusable utterance encoder, with only a small classifier trained on top.
  • I4. The authors justify the context-based design by inspecting annotated dialogues, showing the same string receiving different DA labels in different contexts.
  • I5. Visualising the learned representations as clusters showed confusable classes (Backchannel, Yes-Answer, Agree/Accept) grouping together, with context separating them inside the cluster.

For Mova

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

  • I1 (frames): the speech-act component in prag could read the act of the previous utterance (e.g. a yes-no question) and use it to decide whether "yeah" is an answer, agreement or mere backchannel.
  • I2 (learn): when evaluating context-sensitive components, Mova could add a causal-only setting as a control, so gains from leaking later text are not mistaken for real understanding.
  • I3 (tokens): a character-level encoder could give robust signals for noisy or punctuation-rich input (fragments, disfluencies) where word-level tagging breaks down.
  • I4 (answer): test sets for dialogue or fable QA could include minimal pairs where an identical reply has different meanings depending on the preceding question, to check that readers use context.
  • I5 (learn): clustering Mova's frame or act decisions by class could reveal which labels are confused and whether added context actually pulls them apart.

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

🎓Universität HamburgHamburg · Germany39 works · 4,669 citations

📚 Shelves 4 Computer Science · 9,086 works Artificial Intelligence · 8,889 works Arts and Humanities · 9,296 works Language and Linguistics · 9,206 works

🧭 Topics 2 Dialogue systems · 401 works Speech act theory · 160 works

Cite

Cite as Chandrakant Bothe, Cornelius Weber, Sven Magg and Stefan Wermter. 2018. A Context-based Approach for Dialogue Act Recognition using Simple Recurrent Neural Networks. Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018).

The paper itself, not this page. BibTeX from the library's record of the work. arXiv 1805.06280

@inproceedings{bothe2018context,
  title = {{A Context-based Approach for Dialogue Act Recognition using Simple Recurrent Neural Networks}},
  author = {Chandrakant Bothe, Cornelius Weber, Sven Magg and Stefan Wermter},
  booktitle = {Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)},
  year = {2018},
  doi = {10.63317/2b2w5ksnfqov},
  eprint = {1805.06280},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.63317/2b2w5ksnfqov}
}

References

30 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Allen, J. and Core, M. (1997). Draft of DAMSL: Dialogue Act Markup in Several Layers. find
  2. Austin, J. L. (1962). How to Do Things with Words. Oxford University Press. find
  3. Elman, J. L. (1990). Finding structure in time. Cognitive science, 14(2):179–211. ↗
  4. Godfrey, J. J., Holliman, E. C., and McDaniel, J. (1992). SWITCHBOARD: Telephone speech corpus for research and development. In Proc. of the International Conference on Acoustics, Speech, and Signal Processing, volume 1, pages 517–520. find
  5. Grau, S., Sanchis, E., Castro, M. J., and Vilar, D. (2004). Dialogue act classification using a Bayesian approach. In 9th Conference Speech and Computer SPECOM 2004. find
  6. Grosz, B. J. (1982). Discourse Analysis. Sublanguage. Studies of Language in Restricted Semantic Domains, pages 138–174. Ji, Y., Haffari, G., and Eisenstein, J. ). A Latent Variable Recurrent Neural Network for Discourse Relation Language Models. In Proc. of NAACL-HLT, pages 332–342. find
  7. Jurafsky, D., Shriberg, E., and Biasca, D. (1997). Switchboard Dialog Act Corpus. Technical report, International Computer Science Inst. Berkeley CA. find
  8. Jurafsky, D., Shribergy, E., Fox, B., and Curl, T. (1998). Lexical, Prosodic, and Syntactic Cues for Dialog Acts. In The ACL/COLING Workshop on Discourse Relations and Discourse Markers. our card
  9. Jurafsky, D. (1997). Switchboard SWBD-DAMSL Shallow-Discourse-Function Annotation Coders Manual, draft 13. Technical Report 97-01, University of Colorado Institute of Cognitive Science, pages 225–233. find
  10. Kalchbrenner, N. and Blunsom, P. (2013). Recurrent Convolutional Neural Networks for Discourse Compositionality. In Workshop on Continuous Vector Space Models and their Compositionality, ACL, pages 119–126. our card
  11. Khanpour, H., Guntakandla, N., and Nielsen, R. (2016). Dialogue Act Classification in Domain-Independent Conversations Using a Deep Recurrent Neural Network. In COLING, pages 2012–2021. our card
  12. Kingma, D. and Ba, J. (2014). Adam: A method for stochastic optimization. arXiv: 1412.6980. ↗
  13. Krause, B., Lu, L., Murray, I., and Renals, S. (2016). Multiplicative LSTM for sequence modelling. Workshop track - ICLR 2016. find
  14. Kumar, H., Agarwal, A., Dasgupta, R., Joshi, S., and Kumar, A. (2017). Dialogue Act Sequence Labeling using Hierarchical encoder with CRF. arXiv:1709.04250v2. ↗ our card
  15. Lakomkin, E., Bothe, C., and Wermter, S. (2017). GradAscent at EmoInt-2017: Character and Word Level Recurrent Neural Network Models for Tweet Emotion Intensity Detection. In Proc. of the 8th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis at EMNLP 2017, pages 169–174. ACL. find
  16. Lee, J. Y. and Dernoncourt, F. (2016). Sequential Short-Text Classification with Recurrent and Convolutional Neural Networks. arXiv preprint arXiv:1603.03827. ↗ our card
  17. Liu, Y., Han, K., Tan, Z., and Lei, Y. (2017). Using Context Information for Dialog Act Classification in DNN Framework. In Proc. of the 2017 Conference on EMNLP, pages 2160–2168. ACL. our card
  18. Meng, Z., Mou, L., and Jin, Z. (2017). Hierarchical RNN with Static Sentence-Level Attention for Text-Based Speaker Change Detection. In Proc. of ACM Conference on Information and Knowledge Management, pages 2203–2206. find
  19. Ortega, D. and Vu, N. T. (2017). Neural-based Context Representation Learning for Dialog Act Classification. Proc. of the SIGDIAL 2017 Conference, pages 247–252. our card
  20. O’Shea, J., Bandar, Z., and Crockett, K. (2012). A Multi-classifier Approach to Dialogue Act Classification Using Function Words. In Transactions on Computational Collective Intelligence VII, pages 119–143. Springer. find
  21. Radford, A., Jozefowicz, R., and Sutskever, I. (2017). Learning to Generate Reviews and Discovering Sentiment. arXiv: 1704.01444. ↗ our card
  22. Searle, J. R. (1979). Expression and Meaning: Studies in the Theory of Speech Acts. Cambridge University Press. find
  23. Shriberg, E., Stolcke, A., Jurafsky, D., Coccaro, N., Meteer, M., Bates, R., Taylor, P., Ries, K., Martin, R., and Van Ess-Dykema, C. (1998). Can Prosody Aid the Automatic Classification of Dialog Acts in Conversational Speech? Language and Speech, 41(3-4):443–492. ↗
  24. Shriberg, E., Dhillon, R., Bhagat, S., Ang, J., and Carvey, H. (2004). The ICSI meeting recorder dialog act (MRDA) corpus. Technical report, International Computer Science Inst. Berkeley CA. ↗ our card
  25. Skantze, G. (2007). Error Handling in Spoken Dialogue Systems-Managing Uncertainty, Grounding and Miscommunication: Chapter 2, Spoken Dialogue Systems. KTH Computer Science and Communication. find
  26. Stolcke, A., Ries, K., Coccaro, N., Shriberg, E., Bates, R., Jurafsky, D., Taylor, P., Martin, R., Van Ess-Dykema, C., and Meteer, M. (2000). Dialogue Act Modeling for Automatic Tagging and Recognition of Conversational Speech. Computational Linguistics, 26(3):339–373. ↗ our card
  27. Surendran, D. and Levow, G.-A. (2006). Dialog act tagging with support vector machines and hidden Markov models. In Interspeech – ICSLP. find
  28. Tavafi, M., Mehdad, Y., Joty, S. R., Carenini, G., and Ng, R. T. (2013). Dialogue Act Recognition in Synchronous and Asynchronous Conversations. In SIGDIAL Conference, pages 117–121. ACL. ↗ our card
  29. Tran, Q. H., Zukerman, I., and Haffari, G. (2017). Preserving Distributional Information in Dialogue Act Classification. In Proc. of Conference on EMNLP, pages 2141– 2146. ACL. our card
  30. Yang, X., Liu, J., Chen, Z., and Wu, W. (2014). Semi-supervised Learning of Dialogue Acts Using Sentence Similarity Based on Word Embeddings. In Proc. of International Conference on Audio, Language and Image Processing, pages 882–886. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/bothe-2018-context-based-approach-dialogue-act/