Elamur Research⌕
☰
home / papers / 2018 / Retrieve and Refine: Improved Sequence Generation Models For Dialogue

Retrieve and Refine: Improved Sequence Generation Models For Dialogue

Summary

Neural generators for open-domain chat often give bland, short replies built from common words. Retrieval systems can return livelier responses, but they can only choose from a fixed pool, so their answers may not fit the context. The authors combine the two in RetrieveNRefine (RetNRef). A Key-Value Memory Network retriever picks a candidate utterance, which is appended to the dialogue history. A 2-layer attentive LSTM Seq2Seq model then reads this input and produces the final reply. Two heuristics strengthen the approach. RetNRef+ drops the persona sentences from the generator input, which pushes the generator to rely more on the retrieved candidate. RetNRef++ adds a rule that outputs the retrieved text verbatim whenever the generated reply largely overlaps with it. On the ConvAI2 persona-chat task, perplexity gives little useful signal. Human judges, however, rated RetNRef++ as more engaging than both the Seq2Seq and Memory Network baselines. Its outputs also resembled human utterances more closely in length and rare-word use. The work shows that retrieval can help generative dialogue models produce less generic replies.

Key points

  • Data: ConvAI2, a modified and larger version of PersonaChat, with about 160,000 utterances in about 11,000 dialogues. Personas come from a pool of 1155. Implementation uses ParlAI.
  • Training: For each training turn, the top 100 retrieved candidates are reranked by embedding similarity to the gold reply. The copy rule in RetNRef++ triggers at more than 60% word overlap.
  • Perplexity results:
    • No retrieval: 31.4.
    • Memory Network retrieval: 31.8.
    • Random retrieval: 32.0.
    • Gold reply's nearest neighbour as the retrieval: 25.9.
    • Gold reply itself: 9.2.
    • The history-truncating variant: 48.4, despite its better human scores.
  • Human evaluation (600 dialogues):
    • Engagingness was 3.80 for RetNRef++, 3.66 for the Memory Network and 2.70 for Seq2Seq.
    • In A/B tests, RetNRef++ won about 54% of decisive comparisons against each baseline (ties excluded), which was statistically significant.
    • RetNRef++ was weaker at persona use, scoring 0.65 vs 0.90 for Seq2Seq.
  • Output statistics:
    • Average reply length was 12.7 words for RetNRef++, 11.7 for Seq2Seq and 13.0 for humans.
    • Words with frequency under 1k made up 10.9% of RetNRef++ output, against 11.5% for humans and 5.8% for Seq2Seq.
    • About 53% of RetNRef++ outputs overlap the retrieved text by more than 80%, so it mixes copying with novel generation.

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

Cited by

90 papers of our library cite it (of the 140884 whose bibliographies we read); 9 of them have a page here, 81 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/w18-5713 · arXiv 1808.04776

@inproceedings{weston2018retrieve,
  title = {{Retrieve and Refine: Improved Sequence Generation Models For Dialogue}},
  author = {Weston, Jason and Dinan, Emily and Miller, Alexander},
  booktitle = {Proceedings of the 2018 EMNLP Workshop SCAI: The 2nd International Workshop on Search-Oriented Conversational AI},
  year = {2018},
  doi = {10.18653/v1/w18-5713},
  eprint = {1808.04776},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.18653/v1/w18-5713}
}

References

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

  1. Ziqiang Cao, Wenjie Li, Sujian Li, and Furu Wei. 2018. Retrieve, rerank and rewrite: Soft template based neural summarization. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), volume 1, pages 152–161. find
  2. Iulian Vlad Serban, Ryan Lowe, Laurent Charlin, and Joelle Pineau. 2016. Generative deep neural networks for dialogue: A short review. arXiv preprint arXiv:1611.06216. — [arXiv:1611.06216](https://arxiv.org/abs/1611.06216) find
  3. Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. In Advances in neural information processing systems, pages 3104–3112. find
  4. Angela Fan, David Grangier, and Michael Auli. 2017. Controllable abstractive summarization. arXiv preprint arXiv:1711.05217. find our card
  5. David Grangier and Michael Auli. 2017. Quickedit: Editing text & translations via simple delete actions. arXiv preprint arXiv:1711.04805. — [arXiv:1711.04805](https://arxiv.org/abs/1711.04805) find
  6. Yingce Xia, Fei Tian, Lijun Wu, Jianxin Lin, Tao Qin, Nenghai Yu, and Tie-Yan Liu. 2017. Deliberation networks: Sequence generation beyond one-pass decoding. In Advances in Neural Information Processing Systems, pages 1782–1792. find
  7. Jiatao Gu, Yong Wang, Kyunghyun Cho, and Victor OK Li. 2017. Search engine guided non-parametric neural machine translation. arXiv preprint arXiv:1705.07267. — [arXiv:1705.07267](https://arxiv.org/abs/1705.07267) find
  8. Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing dialogue agents: I have a dog, do you have pets too? arXiv preprint arXiv:1801.07243. find our card
  9. Kelvin Guu, Tatsunori B Hashimoto, Yonatan Oren, and Percy Liang. 2017. Generating sentences by editing prototypes. arXiv preprint arXiv:1709.08878. find our card
  10. Zongcheng Ji, Zhengdong Lu, and Hang Li. 2014. An information retrieval approach to short text conversation. arXiv preprint arXiv:1408.6988. — [arXiv:1408.6988](https://arxiv.org/abs/1408.6988) find
  11. Marcin Junczys-Dowmunt and Roman Grundkiewicz. 2017. An exploration of neural sequence-to-sequence architectures for automatic post-editing. arXiv preprint arXiv:1706.04138. find
  12. Chia-Wei Liu, Ryan Lowe, Iulian V Serban, Michael Noseworthy, Laurent Charlin, and Joelle Pineau. 2016. How not to evaluate your dialogue system: An empirical study of unsupervised evaluation metrics for dialogue response generation. arXiv preprint arXiv:1603.08023. find
  13. Alexander Miller, Adam Fisch, Jesse Dodge, Amir-Hossein Karimi, Antoine Bordes, and Jason Weston. 2016. Key-value memory networks for directly reading documents. arXiv preprint arXiv:1606.03126. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/weston-2018-retrieve-refine-improved-sequence-generation/