Elamur Research⌕
☰
home / papers / 2020 / Towards Robustifying NLI Models Against Lexical Dataset Biases

Towards Robustifying NLI Models Against Lexical Dataset Biases

Paper · 2020 · Xiang Zhou and Mohit Bansal

Summary

NLI models trained on MNLI often score well by exploiting lexical shortcuts instead of understanding meaning. The paper studies two such shortcuts. In Contradiction Word Bias (CWB), words like negations in the hypothesis signal contradiction. In Word Overlapping Bias (WOB), heavy lexical overlap between premise and hypothesis signals entailment. The authors test three levels of debiasing:

  • Data: repeat the rare training examples that go against a bias, or add synthetic examples built like stress-test items.
  • Embeddings: strip label-specific information from the embeddings, borrowing a technique from gender-bias work.
  • Classifier: attach a bag-of-words sub-model that soaks up shallow lexical cues, then use HEX projection to force the main model's representation to be orthogonal to it.

Evaluation uses new balanced subsets extracted from MNLI plus the NLI stress tests of Naik et al. Changing the data helps only a little, and embedding debiasing barely helps while costing accuracy. BoW orthogonality clearly improves robustness at a small overall cost. The lesson is that lexical bias lives mostly in how a model compares and composes words, not in the word vectors themselves.

Key points

  • Data enhancement: counter-bias MNLI training samples were repeated in amounts from 500 up to 50,000. These were non-contradictions containing contradiction words, and non-entailments with high Jaccard word overlap. Bias dropped only slightly and never vanished.
  • Embedding debiasing: removing label bias at the embedding level, as done for gender bias, gave little debiasing and lowered overall accuracy.
  • BoW sub-model with HEX projection (Wang et al., 2019a):
    • It gave large gains on the hard-subset accuracy (Acc hr) of both the balanced and stress-test sets for CWB and WOB.
    • Overall accuracy was only 1.4% (CWB) and 1.3% (WOB) below the baseline.
    • The gains held when the number of layers was changed.
  • What did not help: gradient reversal against the BoW sub-model worked worse than HEX in preliminary runs. Combining HEX with data-level debiasing added nothing.
  • Feature analysis: LIME showed the baseline predicting contradiction almost entirely from "no". After HEX, the model also weighted matches such as the repeated "bad" and tense-bearing words.

Interesting ideas

  • I1. Deliberately build a weak shallow model to absorb shortcut features, then force the real model to learn something orthogonal to it; this needs no prior list of which biases exist.
  • I2. Build evaluation sets by mining balanced, counter-bias examples from the original corpus instead of only synthetic stress tests, so the text distribution stays natural.
  • I3. The finding that bias sits in the comparison/composition layer, not in word representations, suggests auditing the interaction step of a system rather than its lexicon.
  • I4. Using LIME-style feature attribution before and after a fix shows qualitatively whether a model switched from a cue word to meaningful matches.
  • I5. Ensembling several differently initialized shallow sub-models is a way to cover a more diverse set of lexical biases at once.

For Mova

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

  • I1 (learn): A bag-of-words reader could serve as a standing negative control, and any trained Mova component whose predictions track it closely gets flagged as shortcut-driven.
  • I2 (learn): Evaluation of readers could include mined subsets where, for example, negation words or lexical overlap do not predict the answer, separating real inference from cue matching.
  • I3 (frames): When frames are compared to decide entailment or contradiction, checks should test whether the verdict depends on role structure rather than on shared lemmas or a bare "not".
  • I4 (answer): Since Mova's steps are explainable, answers can be audited for how often the justification rests on a single trigger token, giving a cheap shortcut detector for school tests and QA.
  • I5 (learn): Several cheap baselines, such as BoW and overlap counters, could jointly act as witnesses of what not to learn when filtering silver data.

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

🎓University of North Carolina at Chapel HillChapel Hill · United States37 works · 14,332 citations

📚 Shelves 4 Computer Science · 9,086 works Artificial Intelligence · 8,889 works Arts and Humanities · 9,296 works Language and Linguistics · 9,206 works

🧭 Topics 1 Reasoning · 510 works

Cite

Cite as Xiang Zhou and Mohit Bansal. 2020. Towards Robustifying NLI Models Against Lexical Dataset Biases. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2020.acl-main.773

@inproceedings{zhou2020robustifying,
  title = {{Towards Robustifying NLI Models Against Lexical Dataset Biases}},
  author = {Xiang Zhou and Mohit Bansal},
  booktitle = {Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
  year = {2020},
  doi = {10.18653/v1/2020.acl-main.773},
  url = {https://doi.org/10.18653/v1/2020.acl-main.773}
}

References

40 references, 20 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Yonatan Belinkov, Adam Poliak, Stuart Shieber, Benjamin Van Durme, and Alexander Rush. 2019a. On adversarial removal of hypothesis-only bias in natural language inference. In Proceedings of the Eighth Joint Conference on Lexical and Computational Semantics (*SEM 2019), pages 256–262, Minneapolis, Minnesota. Association for Computational Linguistics. our card
  2. Yonatan Belinkov, Adam Poliak, Stuart M Shieber, Benjamin Van Durme, and Alexander M Rush. 2019b. Dont take the premise for granted: Mitigating artifacts in natural language inference. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 877–891. our card
  3. Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In NeurIPS, pages 4349–4357. ↗ our card
  4. Samuel R Bowman, Gabor Angeli, Christopher Potts, and Christopher D Manning. 2015. A large annotated corpus for learning natural language inference. In EMNLP, pages 632–642. ↗
  5. Aylin Caliskan, Joanna J Bryson, and Arvind Narayanan. 2017. Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334):183–186. ↗
  6. Qian Chen, Xiaodan Zhu, Zhen-Hua Ling, Si Wei, Hui Jiang, and Diana Inkpen. 2017. Enhanced lstm for natural language inference. In ACL, volume 1, pages 1657–1668. find
  7. Christopher Clark, Mark Yatskar, and Luke Zettlemoyer. 2019. Dont take the easy way out: Ensemble based methods for avoiding known dataset biases. 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 4060– 4073. ↗
  8. Alexis Conneau, Douwe Kiela, Holger Schwenk, Loı̈c Barrault, and Antoine Bordes. 2017. Supervised learning of universal sentence representations from natural language inference data. In EMNLP, pages 670–680. our card
  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, Volume 1 (Long and Short Papers), pages 4171–4186. ↗ our card
  10. Yanai Elazar and Yoav Goldberg. 2018. Adversarial removal of demographic attributes from text data. In EMNLP, pages 11–21. our card
  11. Yaroslav Ganin and Victor Lempitsky. 2015. Unsupervised domain adaptation by backpropagation. In ICML, pages 1180–1189. ↗
  12. Max Glockner, Vered Shwartz, and Yoav Goldberg. 2018. Breaking nli systems with sentences that require simple lexical inferences. In ACL, volume 2, pages 650–655. our card
  13. Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel Bowman, and Noah A Smith. 2018. Annotation artifacts in natural language inference data. In NAACL-HLT, volume 2, pages 107–112. find
  14. Lieve Hamers, Yves Hemeryck, Guido Herweyers, Marc Janssen, Hans Keters, Ronald Rousseau, and Andr Vanhoutte. 1989. Similarity measures in scientometric research: The jaccard index versus salton’s cosine formula. Information Processing & Management, 25(3):315 – 318. ↗
  15. He He, Sheng Zha, and Haohan Wang. 2019. Unlearn dataset bias in natural language inference by fitting the residual. In Proceedings of the 2nd Workshop on Deep Learning Approaches for Low-Resource NLP (DeepLo 2019), pages 132–142. our card
  16. Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural computation, 9(8):1735–1780. ↗
  17. Robin Jia and Percy Liang. 2017. Adversarial examples for evaluating reading comprehension systems. In EMNLP, pages 2021–2031. our card
  18. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings. ↗
  19. Yitong Li, Timothy Baldwin, and Trevor Cohn. 2018. Towards robust and privacy-preserving text representations. In ACL, volume 2, pages 25–30. our card
  20. Marco Marelli, Luisa Bentivogli, Marco Baroni, Raffaella Bernardi, Stefano Menini, and Roberto Zamparelli. 2014. Semeval-2014 task 1: Evaluation of compositional distributional semantic models on full sentences through semantic relatedness and textual entailment. In SemEval, pages 1–8. our card
  21. R Thomas McCoy, Ellie Pavlick, and Tal Linzen. 2019. Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. arXiv preprint arXiv:1902.01007. ↗ our card
  22. Pasquale Minervini and Sebastian Riedel. 2018. Adversarially regularising neural nli models to integrate logical background knowledge. CoNLL 2018, page 65. ↗
  23. Lili Mou, Rui Men, Ge Li, Yan Xu, Lu Zhang, Rui Yan, and Zhi Jin. 2016. Natural language inference by tree-based convolution and heuristic matching. In ACL, volume 2, pages 130–136. ↗ our card
  24. Aakanksha Naik, Abhilasha Ravichander, Norman Sadeh, Carolyn Rose, and Graham Neubig. 2018. Stress test evaluation for natural language inference. In COLING, pages 2340–2353. ↗
  25. Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In NAACL-HLT, volume 1, pages 1112–1122. find
  26. Yixin Nie, Yicheng Wang, and Mohit Bansal. 2019. Analyzing compositionality-sensitivity of nli models. In AAAI, volume 33, pages 6867–6874. our card
  27. Qizhe Xie, Zihang Dai, Yulun Du, Eduard Hovy, and Graham Neubig. 2017. Controllable invariance through adversarial feature learning. In NeurIPS, pages 585–596. ↗
  28. Tong Niu and Mohit Bansal. 2018. Adversarial over-sensitivity and over-stability strategies for dialogue models. In CoNLL, pages 486–496. find
  29. Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. Glove: Global vectors for word representation. In EMNLP, pages 1532–1543. our card
  30. Adam Poliak, Jason Naradowsky, Aparajita Haldar, Rachel Rudinger, and Benjamin Van Durme. 2018. Hypothesis only baselines in natural language inference. In *SEM, pages 180–191. find
  31. Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. ” why should i trust you?” explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135–1144. our card
  32. Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: a simple way to prevent neural networks from overfitting. JMLR, 15(1):1929–1958. find
  33. Aarne Talman and Stergios Chatzikyriakidis. 2019. Testing the generalization power of neural network models across nli benchmarks. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 85–94. find
  34. 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 NeurIPS, pages 5998–6008. our card
  35. Haohan Wang, Zexue He, Zachary L. Lipton, and Eric P. Xing. 2019a. Learning robust representations by projecting superficial statistics out. In ICLR. ↗
  36. Haohan Wang, Da Sun, and Eric P Xing. 2019b. What if we simply swap the two text fragments? a straightforward yet effective way to test the robustness of methods to confounding signals in nature language inference tasks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 7136–7143. find
  37. Tianlu Wang, Jieyu Zhao, Mark Yatskar, Kai-Wei Chang, and Vicente Ordonez. 2019c. Balanced datasets are not enough: Estimating and mitigating gender bias in deep image representations. In Proceedings of the IEEE International Conference on Computer Vision, pages 5310–5319. find
  38. Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2017. Men also like shopping: Reducing gender bias amplification using corpus-level constraints. In EMNLP. our card
  39. Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2018a. Gender bias in coreference resolution: Evaluation and debiasing methods. In NAACL-HLT, volume 2. ↗ our card
  40. Jieyu Zhao, Yichao Zhou, Zeyu Li, Wei Wang, and Kai-Wei Chang. 2018b. Learning gender-neutral word embeddings. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4847–4853. our card

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhou-2020-robustifying-nli-models-against-lexical/