Elamur Research⌕
☰
home / papers / 2023 / Contextual Dynamic Prompting for Response Generation in Task-oriented Dialog Systems

Contextual Dynamic Prompting for Response Generation in Task-oriented Dialog Systems

Paper · 2023 · Sandesh Swamy, Narges Tabari, Chacha Chen et al.

Summary

Adapting large language models to task-oriented dialog response generation usually means full fine-tuning, which is slow and costly in compute. Prefix-tuning is cheaper, but it uses one static prefix per task, so the prompt cannot reflect what happened earlier in a given conversation. The authors make the prefix depend on the input instead. A frozen T5 encoder reads the earlier dialog turns, and a small trainable feedforward network turns that representation into prefix vectors. Those vectors are then joined with only the current user utterance before generation. In a second variant, the latest dialog state (filled and requested slots) is fed into the encoder along with the context. Because the encoder stays frozen, the number of trainable parameters matches vanilla prefix-tuning. On MultiWOZ 2.2, with delexicalized responses, the combined score rises by about 3 points from context alone and by about 20 points when dialog state is added. Human judges also preferred the context-aware agents. The work shows that parameter-efficient adaptation can be made sensitive to each turn without extra tunable weights.

Key points

  • Setup: the base model is T5. Only the prefix-generating MLP is trained, while the T5 language model and the T5 encoder that reads the context stay frozen.

  • Data: MultiWOZ 2.2 has 7 domains plus a "general" domain, with 8437 training dialogs and 1000 each for dev and test.

  • Metrics: BLEU, Inform and Success are computed per dialog, and combined score = BLEU + 0.5·(Inform + Success).

  • Combined score results:

    • prefix-tuning: 70.54
    • prefix-tuning with dialog state: 68.76
    • contextual dynamic prompt: 73.46
    • contextual dynamic prompt with dialog state: 90.94 (Inform 77.2, Success 68.8, though BLEU drops slightly to 17.94)

    The last variant also produced the most distinct 3-grams (2390, against 1671 for the baseline).

  • Human evaluation: annotators covered a 10% subset of 100 conversations with 728 turns. The best context-aware agent won 38 conversations, while the vanilla prefix-tuning agents won only 3 and 7.

Interesting ideas

  • I1. Instead of placing the whole dialog history in the input, the history is compressed into a few soft prefix vectors that steer generation, so context acts as a control signal rather than as text to copy.
  • I2. Simply adding dialog state to vanilla prefix-tuning slightly hurt it, yet routing that state through the context encoder gave a large gain, so how structured state is injected matters as much as whether it is.
  • I3. The paper keeps the trainable budget fixed by freezing the encoder that produces conditioning features, which makes the comparison with the static baseline fair.
  • I4. Evaluation pairs a surface metric (BLEU) with task metrics (Inform and Success, checked per dialog), and the best system won on the task metrics despite a lower BLEU.
  • I5. Responses are generated in delexicalized form with slot placeholders like [name] and [ref], so the model learns dialog moves while the concrete values are filled in separately.

For Mova

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

  • I1 (back): Generation from the meaning graph could be conditioned on a compact summary of the discourse so far, such as active entities and open questions, kept separate from the material of the current sentence to be verbalized.
  • I2 (world): When the world-of-text state feeds a downstream step, it is worth testing several injection routes against each other rather than assuming that more state always helps.
  • I3 (learn): Ablations of added signals, such as frames or coreference features for a parser or reader, should hold trainable capacity constant so that any gains can be credited to the signal itself.
  • I4 (answer): Readers and generators should be scored on whether the required facts and slots are actually delivered, not only on n-gram overlap, which can mislead.
  • I5 (back): Text generation could first produce a template with typed holes linked to graph entities and fill values from the world in a second step, which keeps each step explainable.

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

🏢AmazonSeattle · United States128 works · 41,697 citations🎓University of ChicagoChicago · United States52 works · 50,893 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,894 worksDialogue systems · 654 works

Cite

Cite as Sandesh Swamy, Narges Tabari, Chacha Chen and Rashmi Gangadharaiah. 2023. Contextual Dynamic Prompting for Response Generation in Task-oriented Dialog Systems. Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2023.eacl-main.226

@inproceedings{swamy2023contextual,
  title = {{Contextual Dynamic Prompting for Response Generation in Task-oriented Dialog Systems}},
  author = {Sandesh Swamy, Narges Tabari, Chacha Chen and Rashmi Gangadharaiah},
  booktitle = {Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics},
  year = {2023},
  doi = {10.18653/v1/2023.eacl-main.226},
  url = {https://doi.org/10.18653/v1/2023.eacl-main.226}
}

References

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

  1. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901.
  2. Paweł Budzianowski, Tsung-Hsien Wen, Bo-Hsiang Tseng, Iñigo Casanueva, Ultes Stefan, Ramadan Osman, and Milica Gašić. 2018. Multiwoz - a large-scale multi-domain wizard-of-oz dataset for task-oriented dialogue modelling. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP).
  3. Kyunghyun Cho, Bart van Merriënboer, Dzmitry Bahdanau, and Yoshua Bengio. 2014. On the properties of neural machine translation: Encoder–decoder approaches. In Proceedings of SSST-8, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation, pages 103–111, Doha, Qatar. Association for Computational Linguistics.
  4. Tianyu Gao, Adam Fisch, and Danqi Chen. 2020. Making pre-trained language models better few-shot learners. arXiv preprint arXiv:2012.15723. ↗
  5. Donghoon Ham, Jeong-Gwan Lee, Youngsoo Jang, and Kee-Eung Kim. 2020. End-to-end neural pipeline for goal-oriented dialogue systems using gpt-2. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 583–592.
  6. Wanwei He, Yinpei Dai, Yinhe Zheng, Yuchuan Wu, Zheng Cao, Dermot Liu, Peng Jiang, Min Yang, Fei Huang, Luo Si, Jian Sun, and Yongbin Li. 2022. Galaxy: A generative pre-trained model for task-oriented dialog with semi-supervised learning and explicit policy injection. Proceedings of the AAAI Conference on Artificial Intelligence, 36(10):10749– 10757. find
  7. Ehsan Hosseini-Asl, Bryan McCann, Chien-Sheng Wu, Semih Yavuz, and Richard Socher. 2020. A simple language model for task-oriented dialogue. Advances in Neural Information Processing Systems, 33:20179– 20191. ↗
  8. Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for NLP. CoRR, abs/1902.00751. ↗
  9. Chia-Hsuan Lee, Hao Cheng, and Mari Ostendorf. 2021. Dialogue state tracking with a language model using schema-driven prompting. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4937–4949, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. find
  10. Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691. ↗
  11. Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190. ↗
  12. Zhaojiang Lin, Andrea Madotto, and Pascale Fung. 2020a. Exploring versatile generative language model via parameter-efficient transfer learning. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 441–459, Online. Association for Computational Linguistics.
  13. Zhaojiang Lin, Andrea Madotto, Genta Indra Winata, and Pascale Fung. 2020b. Mintl: Minimalist transfer learning for task-oriented dialogue systems. arXiv preprint arXiv:2009.12005. ↗
  14. Qi Liu, Lei Yu, Laura Rimell, and Phil Blunsom. 2021. Pretraining the noisy channel model for task-oriented dialogue. Transactions of the Association for Computational Linguistics, 9:657–674. ↗
  15. Shikib Mehri, Tejas Srinivasan, and Maxine Eskenazi. 2019a. Structured fusion networks for dialog. In Proceedings of the 20th Annual SIGdial Meeting on Discourse and Dialogue, pages 165–177, Stockholm, Sweden. Association for Computational Linguistics. find
  16. Shikib Mehri, Tejas Srinivasan, and Maxine Eskenazi. 2019b. Structured fusion networks for dialog. arXiv preprint arXiv:1907.10016. ↗
  17. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318. ↗
  18. Baolin Peng, Chunyuan Li, Jinchao Li, Shahin Shayandeh, Lars Liden, and Jianfeng Gao. 2021. Soloist: Buildingtask bots at scale with transfer learning and machine teaching. Transactions of the Association for Computational Linguistics, 9:807–824. find
  19. Fabio Petroni, Tim Rocktäschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and Sebastian Riedel. 2019. Language models as knowledge bases? arXiv preprint arXiv:1909.01066. ↗
  20. Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. 2018. Improving language understanding by generative pre-training. https://openai.com/blog/language-unsupervised/. ↗
  21. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9. find
  22. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1–67. ↗
  23. Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. 2017. Learning multiple visual domains with residual adapters. CoRR, abs/1705.08045. ↗
  24. Timo Schick and Hinrich Schütze. 2020. Exploiting cloze questions for few shot text classification and natural language inference. arXiv preprint arXiv:2001.07676. ↗
  25. Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. 2020. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980. ↗
  26. Yixuan Su, Lei Shu, Elman Mansimov, Arshit Gupta, Deng Cai, Yi-An Lai, and Yi Zhang. 2021. Multi-task pre-training for plug-and-play task-oriented dialogue system. arXiv preprint arXiv:2109.14739. ↗
  27. Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014. Sequence to sequence learning with neural networks. In Proceedings of the 27th International Conference on Neural Information Processing Systems - Volume 2, NIPS’14, page 3104–3112, Cambridge, MA, USA. MIT Press. ↗
  28. Tsung-Hsien Wen, Milica Gasic, Nikola Mrksic, Pei-Hao Su, David Vandyke, and Steve Young. 2015. Semantically conditioned lstm-based natural language generation for spoken dialogue systems. arXiv preprint arXiv:1508.01745. ↗
  29. Yunyi Yang, Yunhao Li, and Xiaojun Quan. 2021. Ubar: Towards fully end-to-end task-oriented dialog system with gpt-2. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 14230– 14238. find
  30. Xiaoxue Zang, Abhinav Rastogi, Srinivas Sunkara, Raghav Gupta, Jianguo Zhang, and Jindong Chen. 2020. Multiwoz 2.2: A dialogue dataset with additional annotation corrections and state tracking baselines. In Proceedings of the 2nd Workshop on Natural Language Processing for Conversational AI, ACL 2020, pages 109–117.
  31. Qi Zhu, Bing Li, Fei Mi, Xiaoyan Zhu, and Minlie Huang. 2022. Continual prompt tuning for dialog state tracking. arXiv preprint arXiv:2203.06654. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/swamy-2023-contextual-dynamic-prompting-response-generation/