Elamur Research⌕
☰
home / papers / 2020 / DCR-Net: A Deep Co-Interactive Relation Network for Joint Dialog Act Recognition and Sentiment Classification

DCR-Net: A Deep Co-Interactive Relation Network for Joint Dialog Act Recognition and Sentiment Classification

Paper · 2020 · Libo Qin, Wanxiang Che, Yangming Li et al.

Summary

This work addresses two utterance-level labelling problems in dialog: identifying each turn's dialog act and classifying its sentiment. The authors argue that earlier joint systems only shared parameters between the tasks. That approach transfers knowledge poorly and hides how one task affects the other. DCR-Net first encodes the dialog with a shared hierarchical encoder: a word-level BiLSTM turns each utterance into a vector, and self-attention across utterances adds dialog context. Two copies of this representation, one for acts and one for sentiment, then pass through a stack of co-interactive relation layers, where each stream is updated using the other. Three layer types are compared: plain concatenation, an MLP over the concatenation, and utterance-level co-attention with a residual connection. Separate softmax decoders output the labels. On Mastodon and Dailydialog, the model beats the previous best joint system by 4.3 and 3.4 F1 points on dialog acts and by 5.7 and 12.4 on sentiment. Plugging in BERT raises scores further. The results support explicit cross-task interaction over implicit sharing.

Key points

  • Datasets: two corpora are used, Dailydialog, introduced by Li et al. (2017), and the Mastodon corpus, released by Cerisara et al. (2018). Every utterance carries both a dialog act label and a sentiment label.
  • Over the prior state-of-the-art joint model, F1 gains are +4.3 / +3.4 for dialog act recognition and +5.7 / +12.4 for sentiment (Mastodon / Dailydialog).
  • Three relation layers are studied: Concatenation, MLP, and Co-Attention (an extension of Xiong et al.'s mechanism to the utterance level).
  • Stacking relation layers helps up to about three. Deeper stacks perform worse, which the authors attribute to vanishing gradients or overfitting.
  • Replacing the encoder with BERT sets a new best result on both tasks.

Interesting ideas

  • I1. Two label streams that explicitly exchange information in both directions work better than a shared hidden representation, for example when an "Agreement" act implies the speaker shares the previous turn's polarity.
  • I2. The cross-task exchange is repeated over several rounds, but with diminishing returns that turn negative after about three steps.
  • I3. Co-attention weights showed which earlier utterance drove a label: the turn being agreed with got the highest weight in both directions.
  • I4. The paper compares fusion methods from trivial (concatenation) to richer (MLP, co-attention) within one architecture, so each added complexity has to justify itself.

For Mova

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

  • I1 (frames): The speech-act and sentiment/irony components of prag could pass explicit evidence to each other. An agreement act would then carry the polarity of the turn it agrees with, producing a traceable inference.
  • I2 (frames): Mutual refinement between speech-act and implicature decisions could run as a bounded number of passes. The cap should be tuned on held-out data rather than iterating until convergence.
  • I3 (world): A speech act such as agreement or rebuttal could be stored as an explicit link to its target utterance in the world of the text, similar to a coreference antecedent. Questions about who agreed with what would then have a structural answer.
  • I4 (learn): Every new module coupling could be evaluated against the simplest possible baseline coupling. This shows whether the gain comes from the interaction design itself.

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

🎓Harbin Institute of TechnologyHarbin · China76 works · 27,918 citations

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 9,086 worksArtificial Intelligence · 8,889 worksArts and Humanities · 9,296 worksLanguage and Linguistics · 9,206 works

🧭 Topics 3Research subjects and languages the work is about.Sentiment & emotion · 608 worksDialogue systems · 401 worksSpeech act theory · 160 works

Cite

Cite as Libo Qin, Wanxiang Che, Yangming Li, Minheng Ni and Ting Liu. 2020. DCR-Net: A Deep Co-Interactive Relation Network for Joint Dialog Act Recognition and Sentiment Classification. arXiv:2008.06914.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2008.06914 · arXiv 2008.06914

@misc{qin2020dcr,
  title = {{DCR-Net: A Deep Co-Interactive Relation Network for Joint Dialog Act Recognition and Sentiment Classification}},
  author = {Libo Qin, Wanxiang Che, Yangming Li, Minheng Ni and Ting Liu},
  year = {2020},
  doi = {10.48550/arxiv.2008.06914},
  eprint = {2008.06914},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2008.06914}
}

References

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

  1. [2018] Cerisara, C.; Jafaritazehjani, S.; Oluokun, A.; and Le, H. T. 2018. Multi-task dialog act and sentiment recognition on mastodon. In Proc. of COLING. ↗
  2. [2018] Chen, Z.; Yang, R.; Zhao, Z.; Cai, D.; and He, X. 2018. Dialogue act recognition via crf-attentive structured network. In Proc. of SIGIR.
  3. [2019] Chen, Z.; Wang, X.; Xie, X.; Wu, T.; Bu, G.; Wang, Y.; and Chen, E. 2019. Co-attentive multi-task learning for explainable recommendation. In Proc. of IJCAI. find
  4. [2017] Conneau, A.; Schwenk, H.; Barrault, L.; and Lecun, Y. 2017. Very deep convolutional networks for text classification. In Proc. of ACL.
  5. [2019] Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proc. of NAACL.
  6. [1997] Hochreiter, S., and Schmidhuber, J. 1997. Long short-term memory. Neural computation. ↗
  7. [2016] Ji, Y.; Haffari, G.; and Eisenstein, J. 2016. A latent variable recurrent neural network for discourse relation language models. arXiv preprint arXiv:1603.01913. ↗
  8. [2017] Johnson, R., and Zhang, T. 2017. Deep pyramid convolutional neural networks for text categorization. In Proc. of ACL. find
  9. [2013] Kalchbrenner, N., and Blunsom, P. 2013. Recurrent convolutional neural networks for discourse compositionality. arXiv preprint arXiv:1306.3584. ↗
  10. [2018] Kim, M., and Kim, H. 2018. Integrated neural network model for identifying speech acts, predicators, and sentiments of dialogue utterances. Pattern Recognition Letters. ↗
  11. [2011] Kim, H.-S.; Seon, C.-N.; and Seo, J.-Y. 2011. Review of korean speech act classification: machine learning methods. Journal of Computing Science and Engineering. ↗
  12. [2014] Kim, Y. 2014. Convolutional neural networks for sentence classification. In Proc. of EMNLP.
  13. [2014] Kingma, D. P., and Ba, J. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980. ↗
  14. [2018] Kumar, H.; Agarwal, A.; Dasgupta, R.; and Joshi, S. 2018. Dialogue act sequence labeling using hierarchical encoder with crf. In Proc. of AAAI. ↗
  15. [2016] Lee, J. Y., and Dernoncourt, F. 2016. Sequential short-text classification with recurrent and convolutional neural networks. arXiv preprint arXiv:1603.03827. ↗
  16. [2017] Li, Y.; Su, H.; Shen, X.; Li, W.; Cao, Z.; and Niu, S. 2017. Dailydialog: A manually labelled multi-turn dialogue dataset. In Proc. of IJCNLP. ↗
  17. [2017] Liu, Y.; Han, K.; Tan, Z.; and Lei, Y. 2017. Using context information for dialog act classification in dnn framework. In Proc. of EMNLP.
  18. [2019] Majumder, N.; Poria, S.; Hazarika, D.; Mihalcea, R.; Gelbukh, A.; and Cambria, E. 2019. Dialoguernn: An attentive rnn for emotion detection in conversations. In Proc. of AAAI. ↗
  19. [2018] Nguyen, D.-K., and Okatani, T. 2018. Improved fusion of visual and language representations by dense symmetric co-attention for visual question answering. In Proc. of CVPR. find
  20. [2018] Qiao, C.; Huang, B.; Niu, G.; Li, D.; Dong, D.; He, W.; Yu, D.; and Wu, H. 2018. A new method of region embedding for text classification. In Proc. of ICLR. find
  21. [2019] Qin, L.; Che, W.; Li, Y.; Wen, H.; and Liu, T. 2019. A stack-propagation framework with token-level intent detection for spoken language understanding. In Proc. of EMNLP. ↗
  22. [2019] Raheja, V., and Tetreault, J. 2019. Dialogue Act Classification with Context-Aware Self-Attention. In Proc. of NAACL. ↗
  23. [2016] Shi, Y.; Yao, K.; Tian, L.; and Jiang, D. 2016. Deep LSTM based feature mapping for query classification. In Proc. of NAACL. find
  24. [2018] Tan, Z.; Wang, M.; Xie, J.; Chen, Y.; and Shi, X. 2018. Deep semantic role labeling with self-attention. In Proc. of AAAI.
  25. [2015] Tang, D.; Qin, B.; and Liu, T. 2015. Document modeling with gated recurrent neural network for sentiment classification. In Proc. of ACL.
  26. [2019] Tao, C.; Wu, W.; Xu, C.; Hu, W.; Zhao, D.; and Yan, R. 2019. One time of interaction may not be enough: Go deep with an interaction-over-interaction network for response selection in dialogues. In Proc. of ACL. ↗
  27. [2017] Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L. u.; and Polosukhin, I. 2017. Attention is all you need. In Proc. of NIPS. Curran Associates, Inc.
  28. [2018] Wang, B. 2018. Disconnected recurrent neural networks for text categorization. In Proc. of ACL. find
  29. [2018] Wu, Z.; Dai, X.-Y.; Yin, C.; Huang, S.; and Chen, J. 2018. Improving review representations with user attention and product attention for sentiment classification. In Proc. of AAAI. find
  30. [2016] Xiao, Y., and Cho, K. 2016. Efficient character-level document classification by combining convolution and recurrent layers. arXiv preprint arXiv:1602.00367. ↗
  31. [2016] Xiong, C.; Merity, S.; and Socher, R. 2016. Dynamic memory networks for visual and textual question answering. In Proc. of ICML. ↗
  32. [2016] Xu, J.; Chen, D.; Qiu, X.; and Huang, X. 2016. Cached long short-term memory neural networks for document-level sentiment classification. In Proc. of EMNLP. find
  33. [2016] Yang, Z.; Yang, D.; Dyer, C.; He, X.; Smola, A.; and Hovy, E. 2016. Hierarchical attention networks for document classification. In Proc. of NAACL.
  34. [2017] Yin, Q.; Zhang, Y.; Zhang, W.; and Liu, T. 2017. Chinese zero pronoun resolution with deep memory network. In Proc. of EMNLP, 1309–1318.
  35. [2015] Zhang, X.; Zhao, J.; and LeCun, Y. 2015. Character-level convolutional networks for text classification. In Proc. of NIPS, 649–657. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/qin-2020-dcr-net-deep-co-interactive/