Elamur Research
☰
home / papers / 2015 / A Neural Conversational Model

A Neural Conversational Model

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "arXiv — non-exclusive licence to distribute" does not allow redistributing the text itself. Original — arxiv.org/pdf/1506.05869

Summary

Before this work, conversational agents were usually limited to narrow domains and depended on hand-built rules and multi-stage pipelines. Vinyals and Le test whether a generic sequence-to-sequence (seq2seq) LSTM, trained end-to-end without rules, can learn to hold a dialogue. They treat conversation as predicting the next utterance. The model encodes the preceding context and then generates the reply one token at a time. It is trained with cross-entropy, and replies are produced with greedy decoding or beam search. They use two corpora: a closed-domain IT helpdesk troubleshooting chat log and the large, noisy, open-domain OpenSubtitles collection of movie dialogue. On both, the network reaches lower perplexity than an n-gram baseline. On helpdesk data it sometimes guides a user to a working fix. On subtitles it answers factual, philosophical and common-sense questions, many of which never appear in training. Human judges preferred it to the rule-based CleverBot more often than the reverse. Its main weaknesses are short, generic answers and no consistent persona. The paper became a foundational baseline for neural dialogue generation.

Key points

  • IT helpdesk: about 30M training tokens and 3M validation tokens. The model is a single-layer LSTM with 1024 cells and a 20K-word vocabulary. It reaches perplexity 8, compared with 18 for the n-gram model.
  • OpenSubtitles (Tiedemann, 2009): 62M training sentences (923M tokens) and 26M validation sentences (395M tokens). Because the subtitles don't mark speakers, each pair of consecutive sentences is treated as a turn between two people.
  • Human evaluation: 200 questions, each judged by four raters, counted only when at least three of them agreed. NCM won 97, CleverBot won 60, 20 were ties and 23 had no agreement.
  • Known failure: paraphrased questions can get contradictory answers. For example, the model says it is a lawyer in one reply and a doctor in the next.
  • Open problem: the authors say there is still no good automatic metric for dialogue quality. Perplexity and human judgments both have clear limitations.

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: Other

Cited by

414 papers of our library cite it (of the 140673 whose bibliographies we read); 26 of them have a page here, 388 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.1506.05869 · arXiv 1506.05869

@misc{vinyals2015neural,
  title = {{A Neural Conversational Model}},
  author = {Vinyals, Oriol and Le, Quoc},
  year = {2015},
  doi = {10.48550/arxiv.1506.05869},
  eprint = {1506.05869},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.1506.05869}
}

References

18 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Bahdanau, D., Cho, K., and Bengio, Y. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. find our card
  2. Bengio, Y., Ducharme, R., Vincent, P., and Janvin, C. A neural probabilistic language model. The Journal of Machine Learning Research, 3:1137–1155, 2003. find
  3. Hochreiter, S. and Schmidhuber, J. Long short-term memory. Neural Computation, 1997. find
  4. Jean, S., Cho, K., Memisevic, R., and Bengio, Y. On using very large target vocabulary for neural machine translation. CoRR, abs/1412.2007, 2014. find our card
  5. Jurafsky, D. and Martin, J. Speech and language processing. Pearson International, 2009. find
  6. Kalchbrenner, N. and Blunsom, P. Recurrent continuous translation models. In EMNLP, 2013. find
  7. Lester, J., Branting, K., and Mott, B. Conversational agents. In Handbook of Internet Computing. Chapman & Hall, 2004. find
  8. Luong, T., Sutskever, I., Le, Q. V., Vinyals, O., and Zaremba, W. Addressing the rare word problem in neural machine translation. arXiv preprint arXiv:1410.8206, 2014. find our card
  9. Mikolov, T. Statistical Language Models based on Neural Networks. PhD thesis, Brno University of Technology, 2012. find
  10. Mikolov, T., Karafiát, M., Burget, L., Cernockỳ, J., and Khudanpur, S. Recurrent neural network based language model. In INTERSPEECH, pp. 1045–1048, 2010. find
  11. Shang, L., Lu, Z., and Li, H. Neural responding machine for short-text conversation. In Proceedings of ACL, 2015. find our card
  12. Sordoni, A., Galley, M., Auli, M., Brockett, C., Ji, Y., Mitchell, M., Gao, J., Dolan, B., and Nie, J.-Y. A neural network approach to context-sensitive generation of conversational responses. In Proceedings of NAACL, 2015. find our card
  13. Sutskever, I., Vinyals, O., and Le, Q. V. Sequence to sequence learning with neural networks. In NIPS, 2014. find
  14. Tiedemann, J. News from OPUS - A collection of multilingual parallel corpora with tools and interfaces. In Nicolov, N., Bontcheva, K., Angelova, G., and Mitkov, R. (eds.), Recent Advances in Natural Language Processing, volume V, pp. 237–248. John Benjamins, Amsterdam/Philadelphia, Borovets, Bulgari… find
  15. Turing, A. M. Computing machinery and intelligence. Mind, pp. 433–460, 1950. find
  16. Vinyals, O., Kaiser, L., Koo, T., Petrov, S., Sutskever, I., and Hinton, G. Grammar as a foreign language. arXiv preprint arXiv:1412.7449, 2014a. find
  17. Vinyals, O., Toshev, A., Bengio, S., and Erhan, D. Show and tell: A neural image caption generator. arXiv preprint arXiv:1411.4555, 2014b. find our card
  18. Will, T. Creating a Dynamic Speech Dialogue. VDM Verlag Dr, 2007. find