Elamur Research
☰
home / papers / 2019 / Towards Debiasing Fact Verification Models

Towards Debiasing Fact Verification Models

Written for the Research library (claude-opus-5-5) from the paper text. Original — aclanthology.org/D19-1341.pdf

Summary

The paper examines whether the FEVER fact-checking benchmark actually forces models to reason over evidence. The authors found that a BERT classifier that sees only the claim reaches 61.7% accuracy, far above the 33.3% majority baseline. Using Local Mutual Information (LMI), they trace this to annotation artifacts: negation-style bigrams in crowd-written claims strongly predict REFUTES. Experiments with InferSent suggest that world knowledge in embeddings is not the main cause.

To measure evidence-based reasoning without these cues, they built a SYMMETRIC TEST SET. Each original claim-evidence pair receives a hand-written counterpart that keeps the same label but states the opposite fact. Crossing the original and new pairs adds two pairs with inverted labels, so cues from the claim alone carry no information. FEVER-trained models, including the top system NSMN, fall below 60% accuracy on this set.

The authors also propose a model-agnostic regularizer that reweights training instances to neutralize n-gram–label correlations. It gives modest gains on the symmetric set. The work matters because it exposes how dataset bias can inflate reported fact-verification performance.

Key points

  • Claim-only BERT achieves 61.7% on FEVER, versus 33.3% for the majority baseline. Claim-only InferSent scores 57.3% with GloVe and 54.1% with random embeddings.
  • The SYMMETRIC TEST SET contains 956 pairs derived from 99 SUPPORTS and 140 REFUTES FEVER pairs. Two annotators reviewed 285 sampled pairs, and their judgments matched the assigned labels 94% of the time (Cohen κ = 0.88).
  • NSMN drops from 81.8% on FEVER DEV to 58.7% on the generated pairs. ESIM falls from 80.8% to 55.9%, and BERT from 86.2% to 58.3%.
  • Instance reweighting is computed as a preprocessing step that minimizes per-n-gram label bias with an L2 penalty. After reweighting, top give-away bigrams have p(l|w) of roughly 0.30–0.35.
  • Reweighting raises symmetric-set accuracy by 3.4 points for ESIM (to 59.3%) and 3.3 points for BERT (to 61.6%). It lowers FEVER DEV accuracy to 76.0% and 84.6%, respectively.
  • Data and code are available at github.com/TalSchuster/FeverSymmetric.

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

127 papers of our library cite it (of the 140673 whose bibliographies we read); 5 of them have a page here, 122 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/d19-1341

@inproceedings{schuster2019debiasing,
  title = {{Towards Debiasing Fact Verification Models}},
  author = {Schuster, Tal and Shah, Darsh and Yeo, Yun Jie Serene and Ortiz, Daniel Roberto Filizzola and Santus, Enrico and Barzilay, Regina},
  booktitle = {Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)},
  year = {2019},
  doi = {10.18653/v1/d19-1341},
  url = {https://doi.org/10.18653/v1/d19-1341}
}

References

24 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2021–2031. Association for Computational Linguistics. find
  2. Tariq Alhindi, Savvas Petridis, and Smaranda Muresan. 2018. Where is your evidence: Improving fact-checking by justification modeling. In Proceedings of the First Workshop on Fact Extraction and VERification (FEVER), pages 85–90, Brussels, Belgium. Association for Computational Linguistics. find
  3. Heinrich Jiang and Ofir Nachum. 2019. Identifying and correcting label bias in machine learning. arXiv preprint arXiv:1901.04966. — [arXiv:1901.04966](https://arxiv.org/abs/1901.04966) find
  4. Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632–642. Association for Computational Linguistics. find
  5. Divyansh Kaushik and Zachary C. Lipton. 2018. How much reading does reading comprehension require? a critical investigation of popular benchmarks. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 5010–5015. Association for Computational Linguistics. find our card
  6. Qian Chen, Xiaodan Zhu, Zhen-Hua Ling, Si Wei, Hui Jiang, and Diana Inkpen. 2017. Enhanced lstm for natural language inference. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1657–1668. find
  7. Tom McCoy, Ellie Pavlick, and Tal Linzen. 2019. Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3428–3448, Florence, Italy. Association for Computational Ling… find our card
  8. Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and psychological measurement, 20(1):37–46. find
  9. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,… find
  10. Stefan Evert. 2005. The statistics of word cooccurrences: word pairs and collocations. find
  11. Matt Gardner, Joel Grus, Mark Neumann, Oyvind Tafjord, Pradeep Dasigi, Nelson F. Liu, Matthew Peters, Michael Schmitz, and Luke S. Zettlemoyer. 2017. Allennlp: A deep semantic natural language processing platform. find our card
  12. Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel Bowman, and Noah A. Smith. 2018. Annotation artifacts in natural language inference data. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Techno… find
  13. Felix Hill, Antoine Bordes, Sumit Chopra, and Jason Weston. 2016. The goldilocks principle: Reading children’s books with explicit memory representations. In International Conference on Learning Representations 2016. find our card
  14. Robin Jia and Percy Liang. 2017. Adversarial examples for evaluating reading comprehension systems. find our card
  15. Yixin Nie, Haonan Chen, and Mohit Bansal. 2019. Combining fact extraction and verification with neural semantic matching networks. In Association for the Advancement of Artificial Intelligence. find
  16. Timothy Niven and Hung-Yu Kao. 2019. Probing neural network comprehension of natural language arguments. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4658–4664, Florence, Italy. Association for Computational Linguistics. find
  17. Takeshi Onishi, Hai Wang, Mohit Bansal, Kevin Gimpel, and David A. McAllester. 2016. Who did what: A large-scale person-centered cloze dataset. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2230– 2235. The Association for Computational Linguistics. find
  18. Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing, pages 1532–1543. find
  19. Matthew Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technolog… find
  20. Adam Poliak, Jason Naradowsky, Aparajita Haldar, Rachel Rudinger, and Benjamin Van Durme. 2018. Hypothesis only baselines in natural language inference. In Proceedings of the Seventh Joint Conference on Lexical and Computational Semantics, pages 180–191, New Orleans, Louisiana. Association for Compu… find
  21. Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don’t know: Unanswerable questions for squad. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 784–789. Association for Computational Linguistics. find our card
  22. 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
  23. James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. Fever: a large-scale dataset for fact extraction and verification. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Vo… find our card
  24. Jason Weston, Antoine Bordes, Sumit Chopra, Alexander M Rush, Bart van Merriënboer, Armand Joulin, and Tomas Mikolov. 2016. Towards ai-complete question answering: A set of prerequisite toy tasks. In International Conference on Learning Representations. find our card