Elamur Research⌕
☰
home / papers / 2021 / Measuring the ‘I don't know' Problem through the Lens of Gricean Quantity

Measuring the ‘I don't know' Problem through the Lens of Gricean Quantity

Paper · 2021 · Huda Khayrallah and João Sedoc

Summary

Neural chatbots tend to fall back on bland replies such as "I don't know". Earlier work tried to suppress this habit but never measured it directly, so it was unclear why a given fix helped. The authors treat such replies as breaking Grice's maxim of Quantity, which asks a speaker to be informative. From this they build Relative Utterance Quantity (RUQ), an intrinsic check that asks whether the model scores the reference reply above a generic one. They train Transformer chatbots in FAIRSEQ on DailyDialog and on an entropy-filtered version of it. They compare a default FLORES configuration with the winner of a large regularization sweep, and evaluate with RUQ plots, the RUQ score, multi-reference METEOR and human ratings. The baseline models rank "I don't know." above the reference for most training prompts, which is the very data they learned from. Entropy filtering helps a little. Hyperparameter tuning helps far more, bringing the preference for the generic reply under 5%. The diagnostic needs no extra model and no human judges.

Key points

  • RUQ score: for each training prompt, compare the length-normalized model score of "I don't know." with that of the reference, then report the share of prompts where the reference wins. When a prompt has several references, the authors advise using the least likely one.
  • Training-data RUQ is 28.5% for DD-BASE, 37.9% for EF-BASE, 95.3% for DD-BEST and 89.2% for EF-BEST. This means the base models favour the generic reply 71.5% and 62.1% of the time.
  • Multi-reference METEOR is 12.7 for DD-BASE, 13.2 for EF-BASE, 17.8 for DD-BEST and 17.2 for EF-BEST. Tuning adds about 5.1 points, while filtering adds only about 0.5.
  • Across the whole sweep, RUQ tracks METEOR closely (Spearman ρ 0.9). Among the top 20 systems the correlation falls to 0.6, so the two metrics measure different things.
  • The sweep yielded two practical lessons: a SentencePiece vocabulary of 4–8k helps, and label smoothing helps but depends on the vocabulary size. Human ratings were collected on MTurk, with 4 annotators per prompt on 100 prompts; annotators with Cohen's κ below 0.1 were dropped.

Interesting ideas

  • I1. Score a known degenerate output against the gold output under the model itself, on its own training data; a model that prefers the trivial answer to material it was trained on has a defect that no decoding trick will truly cure.
  • I2. When several gold answers are acceptable, test the weakest of them against the generic baseline, because every valid answer should beat an empty one.
  • I3. Tie each conversational failure type to a specific Gricean maxim and give a concrete bad reply for each, which turns vague "bad response" judgements into targeted diagnostics.
  • I4. Plot the average per-token score of the reference, the system output and the generic reply by position in the sentence, to see whether the generic reply wins from the first token or only gains ground later.
  • I5. Check how well a new metric agrees with an established one only among the strongest systems; high agreement over the full range can hide the fact that the two measure different properties.

For Mova

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

  • I1 (learn): Add a negative control that checks how often the parser scores the gold tree above a trivial tree (flat attachment to root, or a left-to-right chain). A falling rate would flag degenerate training before LAS shows it.
  • I2 (answer): On fable QA and school tests, require that even the least preferred correct answer outranks "unknown" or a stock answer. This catches readers that hedge too often.
  • I3 (frames): The prag crate could label system replies by the maxim they break: Quantity for uninformative, Quality for unsupported, Relation for off-topic, Manner for disfluent. Each error then becomes an explainable category.
  • I4 (back): Apply the same curves to the generator and the text → graph → text compressor. They would show where along the sentence bland boilerplate overtakes the content-bearing realization.
  • I5 (learn): Compute the agreement between witness-parser metrics and downstream answer accuracy on top-ranked configurations only. This avoids being misled when choosing among good models.

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

🎓Johns Hopkins UniversityBaltimore · United States187 works · 92,194 citations🎓New York UniversityNew York · United States136 works · 86,390 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 1Research subjects and languages the work is about.Implicature · 72 works

Cite

Cite as Huda Khayrallah and João Sedoc. 2021. Measuring the ‘I don't know' Problem through the Lens of Gricean Quantity. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2021.naacl-main.450

@inproceedings{khayrallah2021measuring,
  title = {{Measuring the ‘I don't know' Problem through the Lens of Gricean Quantity}},
  author = {Huda Khayrallah and João Sedoc},
  booktitle = {Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies},
  year = {2021},
  doi = {10.18653/v1/2021.naacl-main.450},
  url = {https://doi.org/10.18653/v1/2021.naacl-main.450}
}

References

35 references, 17 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Niels Ole Bernsen, Hans Dybkjær, and Laila Dybkjær. 1996. Cooperativity in human-machine and human-human spoken dialogue. Discourse Processes, 21(2):213–236. ↗
  2. Richárd Csáky, Patrik Purgai, and Gábor Recski. 2019. Improving neural conversational models with entropy-based data filtering. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5650–5669, Florence, Italy. Association for Computational Linguistics. find
  3. Jan Deriu, Don Tuggener, Pius von Däniken, Jon Ander Campos, Alvaro Rodrigo, Thiziri Belkacem, Aitor Soroa, Eneko Agirre, and Mark Cieliebak. 2020. Spot the bot: A robust and efficient framework for the evaluation of conversational dialogue systems. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3971–3984, Online. Association for Computational Linguistics. find
  4. Gabriel Forgues, Joelle Pineau, Jean-Marie Larchevêque, and Réal Tremblay. 2014. Bootstrapping dialog systems with word embeddings. In Modern Machine Learning and Natural Language Processing at NeurIPS. find
  5. H. P. Grice. 1975. Logic and Conversation, pages 41 – 58. Brill, Leiden, The Netherlands. find
  6. Prakhar Gupta, Shikib Mehri, Tiancheng Zhao, Amy Pavel, Maxine Eskenazi, and Jeffrey Bigham. 2019. Investigating evaluation of open-domain dialogue systems with human generated multiple references. In Proceedings of the 20th Annual SIGdial Meeting on Discourse and Dialogue, pages 379–391, Stockholm, Sweden. Association for Computational Linguistics. find
  7. Francisco Guzmán, Peng-Jen Chen, Myle Ott, Juan Pino, Guillaume Lample, Philipp Koehn, Vishrav Chaudhary, and Marc’Aurelio Ranzato. 2019. The FLORES evaluation datasets for low-resource machine translation: Nepali–English and Sinhala– English. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6098–6111, Hong Kong, China. Association for Computational Linguistics.
  8. Michael Alexander Kirkwood Halliday. 1989. Spoken and Written Language. Language education. Oxford University Press. find
  9. Sanda Harabagiu, Dan Moldovan, and Takashi Yukawa. 1996. Testing gricean constraints on a wordnet-based coherence evaluation system. In Working Notes of the AAAI-96 Spring Symposium on Computational Approaches to Interpreting and Generating Conversational Implicature, pages 31– 38. find
  10. Tatsunori Hashimoto, Hugh Zhang, and Percy Liang. 2019. Unifying human and statistical evaluation for natural language generation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1689– 1701, Minneapolis, Minnesota. Association for Computational Linguistics. ↗
  11. Prathyusha Jwalapuram. 2017. Evaluating dialogs based on Grice’s maxims. In Proceedings of the Student Research Workshop Associated with RANLP 2017, pages 17–24, Varna. INCOMA Ltd. ↗
  12. Huda Khayrallah and João Sedoc. 2020. SMRT chatbots: Improving non-task-oriented dialog with Simulated Multiple Reference Training. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4489–4505, Online. Association for Computational Linguistics. find
  13. Ryan Kiros, Yukun Zhu, Russ R Salakhutdinov, Richard Zemel, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Skip-thought vectors. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems 28, pages 3294– 3302. Curran Associates, Inc. ↗
  14. Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66–71, Brussels, Belgium. Association for Computational Linguistics.
  15. Batia Laufer and Paul Nation. 1995. Vocabulary Size and Use: Lexical Richness in L2 Written Production. Applied Linguistics, 16(3):307–322. ↗
  16. Alon Lavie and Abhaya Agarwal. 2007. METEOR: An automatic metric for MT evaluation with high levels of correlation with human judgments. In Proceedings of the Second Workshop on Statistical Machine Translation, pages 228–231, Prague, Czech Republic. Association for Computational Linguistics.
  17. Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2016a. A diversity-promoting objective function for neural conversation models. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 110–119, San Diego, California. Association for Computational Linguistics. ↗
  18. Jiwei Li, Will Monroe, Alan Ritter, Dan Jurafsky, Michel Galley, and Jianfeng Gao. 2016b. Deep reinforcement learning for dialogue generation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1192–1202, Austin, Texas. Association for Computational Linguistics.
  19. Margaret Li, Jason Weston, and Stephen Roller. 2019. Acute-eval: Improved dialogue evaluation with optimized questions and multi-turn comparisons. ↗
  20. Yanran Li, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao, and Shuzi Niu. 2017. DailyDialog: A manually labelled multi-turn dialogue dataset. In Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 986–995, Taipei, Taiwan. Asian Federation of Natural Language Processing. ↗
  21. Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74–81, Barcelona, Spain. Association for Computational Linguistics.
  22. Myle Ott, Michael Auli, David Grangier, and Marc’Aurelio Ranzato. 2018. Analyzing uncertainty in neural machine translation. In Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 3956–3965, Stockholmsmässan, Stockholm Sweden. PMLR. ↗
  23. 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, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics. ↗
  24. Mohammed R. H. Qwaider, Abed Alhakim Freihat, and Fausto Giunchiglia. 2017. TrentoTeam at SemEval-2017 task 3: An application of Grice maxims in ranking community question answers. In Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017), pages 271–274, Vancouver, Canada. Association for Computational Linguistics. ↗
  25. Vasile Rus and Mihai Lintean. 2012. A comparison of greedy and optimal assessment of natural language student input using word-to-word similarity metrics. In Proceedings of the Seventh Workshop on Building Educational Applications Using NLP, pages 157–162, Montréal, Canada. Association for Computational Linguistics. find
  26. João Sedoc, Daphne Ippolito, Arun Kirubarajan, Jai Thirani, Lyle Ungar, and Chris Callison-Burch. 2019. ChatEval: A tool for chatbot evaluation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (Demonstrations), pages 60–65, Minneapolis, Minnesota. Association for Computational Linguistics. find
  27. Shikib Mehri and Maxine Eskenazi. 2020a. Unsupervised evaluation of interactive dialog with DialoGPT. In Proceedings of the 21th Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 225–235, 1st virtual meeting. Association for Computational Linguistics.
  28. Iulian V. Serban, Alessandro Sordoni, Yoshua Bengio, Aaron Courville, and Joelle Pineau. 2016. Building end-to-end dialogue systems using generative hierarchical neural network models. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, AAAI’16, page 3776–3783. AAAI Press. ↗
  29. Shikib Mehri and Maxine Eskenazi. 2020b. USR: An unsupervised and reference free evaluation metric for dialog generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 681–707, Online. Association for Computational Linguistics.
  30. Shikhar Sharma, Layla El Asri, Hannes Schulz, and Jeremie Zumer. 2017. Relevance of unsupervised metrics in task-oriented dialogue for evaluating natural language generation. CoRR, abs/1706.09799. ↗
  31. Alexander Miller, Will Feng, Dhruv Batra, Antoine Bordes, Adam Fisch, Jiasen Lu, Devi Parikh, and Jason Weston. 2017. ParlAI: A dialog research software platform. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 79–84, Copenhagen, Denmark. Association for Computational Linguistics. ↗
  32. Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. 2016. Rethinking the inception architecture for computer vision. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818–2826. ↗
  33. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems 30, pages 5998–6008. Curran Associates, Inc.
  34. Sean Welleck, Ilia Kulikov, Stephen Roller, Emily Dinan, Kyunghyun Cho, and Jason Weston. 2020. Neural text generation with unlikelihood training. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net. ↗
  35. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online. Association for Computational Linguistics. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/khayrallah-2021-measuring-i-don-t-know/