Elamur Research
☰
home / papers / 2021 / Surface Form Competition: Why the Highest Probability Answer Isn't Always Right

Surface Form Competition: Why the Highest Probability Answer Isn't Always Right

Written for the Research library (claude-opus-5-5) from the paper text. Original — aclanthology.org/2021.emnlp-main.564.pdf

Summary

Zero-shot multiple-choice evaluation often scores each option by the probability a language model gives it. The authors argue this method underrates models because of what they call surface form competition. A model spreads probability over every string that could express the right answer, such as "Bathtub", "A bathtub" and "Whirlpool bath". As a result, a correct option that is rare or has many paraphrases can lose to a wrong option that is simply a more common string. To fix this, they propose Domain Conditional PMI (PMI_DC). It divides the conditional probability of each option by that option's probability after a short task-specific "domain premise", for example "? the answer is:" for CommonsenseQA. This scores how much the question raises the likelihood of the answer. The method needs no training or fine-tuning. They compare it with raw probability (LM), length-normalised log-likelihood (AVG) and contextual calibration (CC) across all GPT-2 and GPT-3 sizes on more than a dozen datasets. PMI_DC wins consistently, which matters for anyone using generative LMs to pick among options.

Key points

  • The paper compares four scoring functions: LM (highest probability), AVG (log-likelihood averaged per token), CC (contextual calibration from Zhao et al., 2021) and the proposed PMI_DC. It tests them on every GPT-2 and GPT-3 model and on more than a dozen multiple-choice datasets, including COPA, CommonsenseQA, ARC, OpenBookQA, RACE, HellaSwag and BoolQ.
  • PMI_DC estimates an option's prior probability with a task-specific domain premise rather than a generic one. In contrast, Brown et al. (2020) used only "A:" and reported a similar score on just three datasets.
  • In a reformulated "COPA Flipped" setup, the model scores the premise instead of the answer, which removes surface form competition. In that setup all scoring methods do well, which supports the competition explanation. One example: two correct answers ("it was 3 AM" vs "it was 3:30AM") differ by about 4 nats in log probability under standard COPA scoring but score almost the same when flipped.
  • PMI_DC does not reliably win on three datasets. HellaSwag favours AVG, ARC Easy favours LM, and on BoolQ an unconditional baseline does best; there, only PMI_DC with the largest GPT-3 model beats the majority baseline.
  • The code is available at github.com/peterwestuw/surface-form-competition.

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

151 papers of our library cite it (of the 140673 whose bibliographies we read); 16 of them have a page here, 135 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2021.emnlp-main.564 · arXiv 2104.08315

@inproceedings{holtzman2021surface,
  title = {{Surface Form Competition: Why the Highest Probability Answer Isn't Always Right}},
  author = {Holtzman, Ari and West, Peter and Shwartz, Vered and Choi, Yejin and Zettlemoyer, Luke},
  booktitle = {Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing},
  year = {2021},
  doi = {10.18653/v1/2021.emnlp-main.564},
  eprint = {2104.08315},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.18653/v1/2021.emnlp-main.564}
}

References

58 references, 16 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Rami Al-Rfou, Dokook Choe, Noah Constant, Mandy Guo, and Llion Jones. 2019. Character-level language modeling with deeper self-attention. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 3159–3166. find
  2. Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165. find
  3. Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. 2019. Boolq: Exploring the surprising difficulty of natural yes/no questions. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics:… find our card
  4. Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457. find
  5. Ido Dagan, Oren Glickman, and Bernardo Magnini. 2005. The pascal recognising textual entailment challenge. In Machine Learning Challenges Workshop, pages 177–190. Springer. find
  6. Joe Davison, Joshua Feldman, and Alexander Rush. 2019. Commonsense knowledge mining from pre-trained models. 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 1173–… find
  7. Marie-Catherine De Marneffe, Mandy Simons, and Judith Tonhauser. 2019. The commitmentbank: Investigating projection in naturally occurring discourse. In proceedings of Sinn und Bedeutung, volume 23, pages 107–124. find our card
  8. Katrin Erk, Sebastian Padó, and Ulrike Padó. 2010. A flexible, corpus-driven model of regular and inverse selectional preferences. Computational Linguistics, 36(4):723–763. find
  9. Tianyu Gao, Adam Fisch, and Danqi Chen. 2020. Making pre-trained language models better few-shot learners. arXiv preprint arXiv:2012.15723. find our card
  10. Jonathan Gordon and Benjamin Van Durme. 2013. Reporting bias and knowledge acquisition. In Proceedings of the 2013 workshop on Automated knowledge base construction, pages 25–30. find
  11. Sergio Guadarrama, Niveda Krishnamoorthy, Girish Malkarnenkar, Subhashini Venugopalan, Raymond Mooney, Trevor Darrell, and Kate Saenko. 2013. find
  12. Youtube2text: Recognizing and describing arbitrary activities using semantic hierarchies and zero-shot recognition. In Proceedings of the IEEE international conference on computer vision, pages 2712– 2719. find
  13. Karen Hambardzumyan, Hrant Khachatrian, and Jonathan May. 2021. Warp: Word-level adversarial reprogramming. arXiv preprint arXiv:2101.00121. find
  14. Ari Holtzman, Jan Buys, Maxwell Forbes, and Yejin Choi. 2020. The curious case of neural text degeneration. International Conference on Learning Representations. find
  15. Tim Florian Jaeger. 2006. Redundancy and syntactic reduction in spontaneous speech. Ph.D. thesis, Stanford University Stanford, CA. find
  16. Zhengbao Jiang, Jun Araki, Haibo Ding, and Graham Neubig. 2020a. How can we know when language models know? arXiv preprint arXiv:2012.00955. find
  17. Zhengbao Jiang, Frank F. Xu, Jun Araki, and Graham Neubig. 2020b. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423–438. find our card
  18. Daniel Khashabi, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. 2020. Unifiedqa: Crossing format boundaries with a single qa system. arXiv preprint arXiv:2005.00700. find our card
  19. Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard Hovy. 2017. Race: Large-scale reading comprehension dataset from examinations. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 785– 794. find our card
  20. Roger Levy. 2018. Communicative efficiency, uniform information density, and the rational speech act theory. In CogSci. find
  21. Roger Levy and T Florian Jaeger. 2007. Speakers optimize information density through syntactic reduction. Advances in neural information processing systems, 19:849. find
  22. Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and William B Dolan. 2016. 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… find our card
  23. Xin Li and Dan Roth. 2002. Learning question classifiers. In COLING 2002: The 19th International Conference on Computational Linguistics. find
  24. Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. 2021. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. arXiv preprint arXiv:2104.08786. find our card
  25. Huanru Henry Mao, Bodhisattwa Prasad Majumder, Julian McAuley, and Garrison Cottrell. 2019. Improving neural story generation by targeted common sense grounding. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Na… find
  26. Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. 2018. Can a suit of armor conduct electricity? a new dataset for open book question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2381–2391. find
  27. Nasrin Mostafazadeh, Michael Roth, Annie Louis, Nathanael Chambers, and James Allen. 2017. find
  28. Lsdsem 2017 shared task: The story cloze test. In Proceedings of the 2nd Workshop on Linking Models of Lexical, Sentential and Discourse-level Semantics, pages 46–51. find
  29. Lili Mou, Yiping Song, Rui Yan, Ge Li, Lu Zhang, and Zhi Jin. 2016. Sequence to backward and forward sequences: A content-introducing approach to generative short-text conversation. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages… find
  30. Olabiyi Oluwatobi and Erik Mueller. 2020. DLGNet: A transformer-based model for dialogue response generation. In Proceedings of the 2nd Workshop on Natural Language Processing for Conversational AI, pages 54–62, Online. Association for Computational Linguistics. find
  31. Patrick Pantel, Rahul Bhagat, Bonaventura Coppola, Timothy Chklovski, and Eduard Hovy. 2007. ISP: Learning inferential selectional preferences. In Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings of the Main… find
  32. Ethan Perez, Douwe Kiela, and Kyunghyun Cho. 2021. True few-shot learning with language models. arXiv preprint arXiv:2105.11447. find our card
  33. Ben Peters, Vlad Niculae, and André FT Martins. 2019. Sparse sequence-to-sequence models. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1504–1519. find
  34. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9. find
  35. Laria Reynolds and Kyle McDonell. 2021. Prompt programming for large language models: Beyond the few-shot paradigm. arXiv preprint arXiv:2102.07350. find our card
  36. Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S Gordon. 2011. Choice of plausible alternatives: An evaluation of commonsense causal reasoning. In AAAI Spring Symposium: Logical Formalizations of Commonsense Reasoning, pages 90–95. find
  37. Bernardino Romera-Paredes and Philip Torr. 2015. An embarrassingly simple approach to zero-shot learning. In International conference on machine learning, pages 2152–2161. PMLR. find
  38. Timo Schick and Hinrich Schütze. 2020a. Exploiting cloze questions for few-shot text classification and natural language inference. arXiv preprint arXiv:2001.07676. find our card
  39. Timo Schick and Hinrich Schütze. 2020b. Few-shot text generation with pattern-exploiting training. arXiv preprint arXiv:2012.11926. — [arXiv:2012.11926](https://arxiv.org/abs/2012.11926) find
  40. Timo Schick and Hinrich Schütze. 2020c. It’s not just size that matters: Small language models are also few-shot learners. arXiv preprint arXiv:2009.07118. find our card
  41. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715– 1725. find
  42. Taylor Shin, Yasaman Razeghi, Robert L. Logan IV, Eric Wallace, and Sameer Singh. 2020. AutoPrompt: Eliciting Knowledge from Language Models with Automatically Generated Prompts. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4222–4235, Onlin… find our card
  43. Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631… find
  44. Robyn Speer, Joshua Chin, and Catherine Havasi. 2017. Conceptnet 5.5: An open multilingual graph of general knowledge. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 31. find
  45. Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2019. Commonsenseqa: A question answering challenge targeting commonsense knowledge. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,… find
  46. Jianheng Tang, Tiancheng Zhao, Chenyan Xiong, Xiaodan Liang, Eric Xing, and Zhiting Hu. 2019. Target-guided open-domain conversation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5624–5634. find
  47. Ruud Van der Helm. 2006. Towards a clarification of probability, possibility and plausibility: how semantics could help futures practice to improve. Foresight. find
  48. Changhan Wang, Kyunghyun Cho, and Jiatao Gu. 2020. Neural machine translation with byte-level subwords. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 9154– 9160. find
  49. Su Wang, Greg Durrett, and Katrin Erk. 2018. Modeling semantic plausibility by injecting world knowledge. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 303–308, New Orl… find
  50. Tony Z Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. arXiv preprint arXiv:2102.09690. find our card
  51. Ruiqi Zhong, Kristy Lee, Zheng Zhang, and Dan Klein. 2021. Adapting language models for zero-shot learning by meta-tuning on dataset and prompt collections. arXiv preprint arXiv:2109.01652. find
  52. Kun Zhou, Kai Zhang, Yu Wu, Shujie Liu, and Jingsong Yu. 2019. Unsupervised context rewriting for open domain conversation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNL… find
  53. Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. 2021. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652. find
  54. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pier-ric Cistac, Tim Rault, Remi Louf, Morgan Funtow-icz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lh… find our card
  55. Lili Yao, Yaoyuan Zhang, Yansong Feng, Dongyan Zhao, and Rui Yan. 2017. Towards implicit content-introducing for generative short-text conversation systems. In Proceedings of the 2017 conference on empirical methods in natural language processing, pages 2190–2199. find
  56. Qinyuan Ye, Bill Yuchen Lin, and Xiang Ren. 2021. Crossfit: A few-shot learning challenge for cross-task generalization in nlp. arXiv preprint arXiv:2104.08835. find
  57. Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. Hellaswag: Can a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4791–4800. find
  58. Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. Advances in Neural Information Processing Systems, 28:649–657. find our card