---
title: "Generating High-Quality and Informative Conversation"
description: "Second, they replace standard decoding with a stochastic beam search that reranks partial outputs… 15 references, cited by 63 papers here."
authors: ["Yuanlong Shao", "Stephan Gouws", "Denny Britz", "Anna Goldie", "Brian Strope", "Ray Kurzweil"]
year: 2017
venue: "Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing"
source: "https://aclanthology.org/D17-1235/"
doi: "10.18653/v1/d17-1235"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/shao-2017-generating-high-quality-informative-conversation/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/shao-2017-generating-high-quality-informative-conversation/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2017](https://md.research.elamur.ai/free/en/papers/2017/page/2/) / Generating High-Quality and Informative Conversation Responses with Sequence-to-Sequence Models

# Generating High-Quality and Informative Conversation Responses with Sequence-to-Sequence Models

**Authors:** Yuanlong Shao, [Stephan Gouws](https://md.research.elamur.ai/free/en/authors/stephan-gouws/), [Denny Britz](https://md.research.elamur.ai/free/en/authors/denny-britz/), [Anna Goldie](https://md.research.elamur.ai/free/en/authors/anna-goldie/), [Brian Strope](https://md.research.elamur.ai/free/en/authors/brian-strope/) and [Ray Kurzweil](https://md.research.elamur.ai/free/en/authors/ray-kurzweil/) · **Year:** 2017 · **Venue:** [Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing](https://md.research.elamur.ai/free/en/venues/emnlp/) ([EMNLP 2017](https://md.research.elamur.ai/free/en/venues/emnlp/2017/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/D17-1235/) · [PDF](https://aclanthology.org/D17-1235.pdf) · [DOI 10.18653/v1/d17-1235](https://doi.org/10.18653/v1/d17-1235) · [OpenAlex](https://openalex.org/W2734443755)

**Citations:** 179 citations (OpenAlex, 2026-10-14) · [63](https://md.research.elamur.ai/free/en/papers/shao-2017-generating-high-quality-informative-conversation/cited-by/) in our library, of the 141,840 papers whose bibliographies we read

## Summary
Neural seq2seq conversation models tend to produce short, bland replies. When pushed to be longer, their output often becomes repetitive, incoherent or self-contradictory. Working in a single-turn setting, Shao, Gouws et al. (Google) make small changes to the standard model to get long, varied and sensible responses. First, they add attention over the decoder's own past output. Full target-side attention was too memory-hungry at scale, so they propose the "glimpse model." The decoder learns only a fixed-length slice of the reply, and the source plus the earlier part of the reply are placed on the encoder side. Second, they replace standard decoding with a stochastic beam search that reranks partial outputs segment by segment, which adds variety early in generation. Training used over 2.3B web conversation messages. Human raters judged the method's long replies to hold or improve in quality as length grew, while the length-normalized baseline degraded. A hybrid back-off system beat the baseline in pairwise preference. The work shows that cheap modeling and decoding tweaks plus massive data can noticeably improve open-domain response generation.

## Key points
- Target-side attention gave small perplexity improvements on OpenSubtitles 2009. The glimpse model was used at scale because it needs only data preprocessing on top of an unmodified seq2seq implementation.
- On the full data, the glimpse model with K=10 trained faster than the source-attention-only baseline, even though the average response length was 38. The K=1 variant learned much more slowly and was stopped early.
- About one third of the proposed method's responses exceeded 100 characters, compared with a negligible share for the baseline. The baseline used beam search with length normalization.
- Each response received three human ratings. "Major agreement" (at least two of three ratings identical) was 80% for the baseline and 70% for the new method.
- The back-off system uses the plain beam-search reply (no length normalization) when that reply is under 40 characters and the new method's reply otherwise. On 200 prompts it took 133 replies from beam search and 67 from the new model. In pairwise comparison it was preferred 103 times versus 77, with 20 ties.
- The authors release their set of context-free prompts as an evaluation benchmark.

## Interesting ideas
- **Attention via data reshaping:** To approximate attention over already-generated tokens, the earlier output is moved onto the encoder input and the model is trained on short target windows. This turns an architectural change into a preprocessing step, which helps anywhere memory limits long-target training.
- **Rerank early, not at the end:** Rescoring whole N-best lists comes too late, because beam search has already narrowed down to near-identical candidates. Reranking after each short segment steers diversity while it can still matter, which could also benefit summarization or captioning.
- **Sampling inside beam search:** Drawing candidates stochastically within the beam, instead of greedy sampling or pure MAP search, balances variety against the error accumulation that plain sampling suffers from underfit models.
- **Debiasing against generic outputs:** Model scores lean toward the unconditional likelihood of a response. Normalizing them, for example by the marginal or by scores under random prompts, helps surface replies that are specific to the input. It is a cheap trick for any conditional generator prone to "safe" outputs.
- **Length-aware back-off and evaluation:** Plotting human ratings against response length exposed that each method wins in a different length range. A simple length-threshold fallback then combined their strengths, a pattern worth reusing when comparing generation systems.

*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

- [Google](https://md.research.elamur.ai/free/en/companies/google/) — Company, Mountain View, United States · 353 works · 364,806 citations
- [Google (United Kingdom)](https://md.research.elamur.ai/free/en/companies/google-united-kingdom/) — Company, London, United Kingdom · 6 works · 916 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 9,085 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 8,888 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 9,296 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 9,206 works

## Cite

Cite as: Yuanlong Shao, Stephan Gouws, Denny Britz, Anna Goldie, Brian Strope and Ray Kurzweil. 2017. Generating High-Quality and Informative Conversation Responses with Sequence-to-Sequence Models. Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing.

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

```bibtex
@inproceedings{shao2017generating,
  title = {{Generating High-Quality and Informative Conversation Responses with Sequence-to-Sequence Models}},
  author = {Yuanlong Shao, Stephan Gouws, Denny Britz, Anna Goldie, Brian Strope and Ray Kurzweil},
  booktitle = {Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing},
  year = {2017},
  doi = {10.18653/v1/d17-1235},
  url = {https://doi.org/10.18653/v1/d17-1235}
}
```

- [← previous Argumentation Mining in User-Generated Web Discourse](https://md.research.elamur.ai/free/en/papers/habernal-2017-argumentation-mining-user-generated-web/)
- [next → Semantic Specialization of Distributional Word Vector Spaces using Monolingual and Cross-Lingual Constraints](https://md.research.elamur.ai/free/en/papers/mrksic-2017-semantic-specialization-distributional-word-vector/)
- [↑ back to papers of 2017](https://md.research.elamur.ai/free/en/papers/2017/page/2/)

## References

15 references, 9 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Martı́n Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenbe… [find](https://scholar.google.com/scholar?q=Mart%C4%B1%CC%81n+Abadi%2C+Ashish+Agarwal%2C+Paul+Barham%2C+Eugene+Brevdo%2C+Zhifeng+Chen%2C+Craig+Citro%2C+Greg+S%2E+Corrado%2C+Andy+Davis%2C+Jeffrey+Dean%2C+Matthieu+Devin%2C+Sanjay+Ghemawat%2C+Ian+Goodfellow%2C+Andrew+Harp%2C+Geoffrey+)
2. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. [Neural machine translation by jointly learning to align and translate](https://md.research.elamur.ai/free/en/papers/bahdanau-2014-neural-machine-translation-jointly-learning/). arXiv preprint arXiv:1409.0473 .
3. Kyunghyun Cho, Bart Van Merriënboer, Çalar Gülçehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using rnn encoder–decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Lan… [find](https://scholar.google.com/scholar?q=Kyunghyun+Cho%2C+Bart+Van+Merrie%CC%88nboer%2C+C%CC%A7alar+Gu%CC%88lc%CC%A7ehre%2C+Dzmitry+Bahdanau%2C+Fethi+Bougares%2C+Holger+Schwenk%2C+and+Yoshua+Bengio%2E+2014%2E+Learning+phrase+representations+using+rnn+encoder%E2%80%93decoder+for+statis)
4. Christopher Cieri David, David Miller, and Kevin Walker. 2004. [The fisher corpus: a resource for the next generations of speech-to-text](https://md.research.elamur.ai/free/en/papers/cieri-2004-fisher-corpus-resource-next-generations/). In in Proceedings 4th International Conference on Language Resources and Evaluation. pages 69–71.
5. Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2015. [A diversity-promoting objective function for neural conversation models](https://md.research.elamur.ai/free/en/papers/li-2016-diversity-promoting-objective-function-neural/). arXiv preprint arXiv:1510.03055 .
6. Jiwei Li, Will Monroe, Alan Ritter, Michel Galley, Jianfeng Gao, and Dan Jurafsky. 2016. [Deep reinforcement learning for dialogue generation](https://md.research.elamur.ai/free/en/papers/li-2016-deep-reinforcement-learning-dialogue-generation/). CoRR abs/1606.01541.
7. Chia-Wei Liu, Ryan Lowe, Iulian Vlad 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](https://md.research.elamur.ai/free/en/papers/liu-2016-how-not-evaluate-your-dialogue/). In EMNLP. ACL, Austin, Texas, page 21222132. [https://aclweb.org…](https://aclweb.org…) [↗](https://aclweb.org…)
8. Lifeng Shang, Zhengdong Lu, and Hang Li. 2015. [Neural responding machine for short-text conversation](https://md.research.elamur.ai/free/en/papers/shang-2015-neural-responding-machine-short-text/). arXiv preprint arXiv:1503.02364 .
9. Alessandro Sordoni, Michel Galley, Michael Auli, Chris Brockett, Yangfeng Ji, Margaret Mitchell, Jian-Yun Nie, Jianfeng Gao, and Bill Dolan. 2015. [A neural network approach to context-sensitive generation of conversational responses](https://md.research.elamur.ai/free/en/papers/sordoni-2015-neural-network-approach-context-sensitive/). arXiv preprint arXiv:1506.06714 .
10. Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. In Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 27, Curran Associates, Inc., pages 3104–3112. [find](https://scholar.google.com/scholar?q=Ilya+Sutskever%2C+Oriol+Vinyals%2C+and+Quoc+V+Le%2E+2014%2E+Sequence+to+sequence+learning+with+neural+networks%2E+In+Z%2E+Ghahramani%2C+M%2E+Welling%2C+C%2E+Cortes%2C+N%2E+D%2E+Lawrence%2C+and+K%2E+Q%2E+Weinberger%2C+editors%2C+Advances)
11. Jörg Tiedemann. 2009. News from OPUS - A Collection of Multilingual Parallel Corpora with Tools and Interfaces. In N. Nicolov, K. Bontcheva, G. Angelova, and R. Mitkov, editors, Recent Advances in Natural Language Processing (vol V), John Benjamins, Amsterdam/Philadelphia, pages 237–248. [find](https://scholar.google.com/scholar?q=Jo%CC%88rg+Tiedemann%2E+2009%2E+News+from+OPUS+%2D+A+Collection+of+Multilingual+Parallel+Corpora+with+Tools+and+Interfaces%2E+In+N%2E+Nicolov%2C+K%2E+Bontcheva%2C+G%2E+Angelova%2C+and+R%2E+Mitkov%2C+editors%2C+Recent+Advances+in+Na)
12. Oriol Vinyals and Quoc Le. 2015. [A neural conversational model](https://md.research.elamur.ai/free/en/papers/vinyals-2015-neural-conversational-model/). arXiv preprint arXiv:1506.05869 .
13. Tsung-Hsien Wen, Milica Gasic, Dongho Kim, Nikola Mrksic, Pei-hao Su, David Vandyke, and Steve J. Young. 2015. [Stochastic language generation in dialogue using recurrent neural networks with convolutional sentence reranking](https://md.research.elamur.ai/free/en/papers/wen-2015-stochastic-language-generation-dialogue-recurrent/). CoRR abs/1508.01755.
14. Tsung-Hsien Wen, Milica Gasic, Nikola Mrkšić, Lina M. Rojas Barahona, Pei-Hao Su, Stefan Ultes, David Vandyke, and Steve Young. 2016. Conditional generation and snapshot learning in neural dialogue systems. In EMNLP. ACL, Austin, Texas, pages 2153–2162. [https://aclweb.org/anthology/D16-1233](https://aclweb.org/anthology/D16-1233). [↗](https://aclweb.org/anthology/D16-1233)
15. Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V. Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, Jeff Klingner, Apurva Shah, Melvin Johnson, Xiaobing Liu, Łukasz Kaiser, Stephan Gouws, Yoshikiyo Kato, Taku Kudo, Hideto Kazawa, Keith Stevens, George Kurian, N… [find](https://scholar.google.com/scholar?q=Yonghui+Wu%2C+Mike+Schuster%2C+Zhifeng+Chen%2C+Quoc+V%2E+Le%2C+Mohammad+Norouzi%2C+Wolfgang+Macherey%2C+Maxim+Krikun%2C+Yuan+Cao%2C+Qin+Gao%2C+Klaus+Macherey%2C+Jeff+Klingner%2C+Apurva+Shah%2C+Melvin+Johnson%2C+Xiaobing+Liu%2C+%C5%81uk)

*Updated 24 October 2026*