Elamur Research⌕
☰
home / papers / 2020 / Towards the Necessity for Debiasing Natural Language Inference Datasets

Towards the Necessity for Debiasing Natural Language Inference Datasets

Paper · 2020 · Mithun Paul Panenghat, Sandeep Suntwal, Faiz Rafique et al.

Summary

Neural NLI models often succeed by exploiting dataset artifacts, such as particular words or entities that co-occur with a label, rather than by reasoning over the premise and hypothesis. The authors test whether masking lexical content reduces this reliance. They use two delexicalization schemes. OA-NER replaces named and numeric entities, found by the CoreNLP NER, with class tags that carry an index showing whether the entity first appeared in the claim or only in the evidence. OA-NER+SS also replaces common nouns and verbs with WordNet super sense tags. They retrain Decomposable Attention and ESIM, without tuning, on masked versions of MNLI, FEVER, FNC and MedNLI.

In-domain accuracy stays close to the lexicalized baselines under OA-NER. Cross-domain transfer improves, notably between FEVER and FNC. Super sense lifting, however, hurts accuracy, and an error analysis shows the masking tags introduce their own label biases. The masked datasets and code are released. The work matters because it offers a cheap way to test, and partly force, reasoning that does not depend on memorizing lexical items.

Key points

  • Masking schemes:
    • OA-NER produces tags like organization-c1 (first seen in the claim) or date-e2 (evidence only), and an entity keeps the same tag wherever it recurs.
    • OA-NER+SS additionally maps nouns and verbs to coarse WordNet classes, e.g. "enter" becomes motion and "Airbus A380" becomes artifact.
  • MNLI (Table 2):
    • DA improves with delexicalized data: 60.95→64.52% on matched and 61.43→64.86% on mismatched.
    • ESIM barely moves: 68.84→68.14% and 69.40→69.10%.
  • Out-of-domain transfer with DA (Table 3, lexicalized vs OA-NER):
    • FEVER→FNC: 48.86→53.59%.
    • FNC→FEVER: 41.16→46.47%.
    • MNLI→MedNLI: essentially flat at 51.47→51.57%.
  • In-domain cost (Table 4):
    • FEVER drops slightly, from 83.43 to 82.31% (−1.12).
    • FNC drops from 68.99 to 65.85%.
    • Adding super senses causes a sharp fall, to 75.26% on FEVER and 45.51% on FNC.
  • Error analysis and release:
    • The neutral FNC class discuss is misclassified most often.
    • Tags such as PERSON-c1 lean toward agree more than the base label rate (62.39% vs 55.05%).
    • Masked MNLI, FEVER and FNC are released on OSF, with code on GitHub (clulab/releases).

Interesting ideas

  • I1. Replacing entities with typed placeholders that keep a shared index across premise and hypothesis preserves the overlap signal while hiding the identity of the entity, so a model must reason about "the same thing" rather than about "Clinton".
  • I2. Training on masked data can serve as a probe: if accuracy barely drops when the words are hidden, the model never truly used lexical meaning, and if transfer improves, the lexicon was carrying spurious cues.
  • I3. Measuring how strongly each placeholder tag co-occurs with each label, relative to the base label distribution, shows that debiasing can itself create new biases.
  • I4. Semantic lifting to coarse WordNet classes can lose too much information, and the right level of abstraction is task-dependent rather than "the more general, the better".
  • I5. To evaluate across datasets with different label sets, the authors align labels explicitly, for example splitting "not enough info" into "unrelated" and "discuss", so that each target's official scoring can still be used.

For Mova

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

  • I1 (world): Entity slots in the text world could be checked by re-running answering with the names swapped for indexed type tokens; if answers change, the reader is relying on lexical identity instead of the graph.
  • I2 (learn): Delexicalized copies of the evaluation sets could be added as a standard negative control, so that a learned component scoring equally well on them is flagged as relying on structure-free shortcuts.
  • I3 (learn): Silver data and self-induced repair rules could be audited with the same per-feature versus prior label tables, to catch rules that fire mainly because of a tag artifact.
  • I4 (world): When mapping frame arguments onto the WordNet/FrameNet ontology, the hypernym level could be chosen per task and validated empirically, instead of always collapsing to supersenses.
  • I5 (answer): School-test and QA readers with differing answer schemes could share one internal verdict set, with a documented mapping to each benchmark's labels, keeping cross-benchmark comparisons honest.

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 ArizonaTucson · United States19 works · 10,269 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 2Research subjects and languages the work is about.Reasoning · 510 worksBias, fairness & safety · 209 works

Cite

Cite as Mithun Paul Panenghat, Sandeep Suntwal, Faiz Rafique, Rebecca Sharp and Mihai Surdeanu. 2020. Towards the Necessity for Debiasing Natural Language Inference Datasets. Proceedings of the Twelfth Language Resources and Evaluation Conference.

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{panenghat2020necessity,
  title = {{Towards the Necessity for Debiasing Natural Language Inference Datasets}},
  author = {Mithun Paul Panenghat, Sandeep Suntwal, Faiz Rafique, Rebecca Sharp and Mihai Surdeanu},
  booktitle = {Proceedings of the Twelfth Language Resources and Evaluation Conference},
  year = {2020},
  doi = {10.17605/osf.io/szdkn},
  url = {https://doi.org/10.17605/osf.io/szdkn}
}

References

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

  1. Bos, J. and Markert, K. (2005). Recognising textual entailment with logical inference. In Proceedings of the conference on Human Language Technology and Empirical Methods in Natural Language Processing, pages 628– 635. Association for Computational Linguistics.
  2. Bowman, S. R., Angeli, G., Potts, C., and Manning, C. D. (2015). A large annotated corpus for learning natural language inference. arXiv preprint arXiv:1508.05326. ↗
  3. Chen, Q., Zhu, X., Ling, Z., Wei, S., Jiang, H., and Inkpen, D. (2016). Enhanced lstm for natural language inference. arXiv preprint arXiv:1609.06038. ↗
  4. Ciaramita, M. and Johnson, M. (2003). Supersense tagging of unknown nouns in wordnet. In Proceedings of the 2003 conference on Empirical methods in natural language processing, pages 168–175. Association for Computational Linguistics.
  5. Condoravdi, C., Crouch, D., De Paiva, V., Stolle, R., and Bobrow, D. G. (2003). Entailment, intensionality and text understanding. In Proceedings of the HLT-NAACL 2003 workshop on Text meaning, pages 38–45. find
  6. Dagan, I., Roth, D., Sammons, M., and Zanzotto, F. M. (2013). Recognizing textual entailment: Models and applications. Synthesis Lectures on Human Language Technologies, 6(4):1–220. ↗
  7. Fyodorov, Y., Winter, Y., and Francez, N. (2000). A natural logic inference system. In Proceedings of the 2nd Workshop on Inference in Computational Semantics (ICoS-2). Citeseer. find
  8. Grishman, R. and Sundheim, B. (1996). Message understanding conference-6: A brief history. In COLING 1996 Volume 1: The 16th International Conference on Computational Linguistics. find
  9. Gururangan, S., Swayamdipta, S., Levy, O., Schwartz, R., Bowman, S. R., and Smith, N. A. (2018). Annotation artifacts in natural language inference data. arXiv preprint arXiv:1803.02324. ↗
  10. MacCartney, B. and Manning, C. D. (2009). An extended model of natural logic. In Proceedings of the eighth international conference on computational semantics, pages 140–156. Association for Computational Linguistics. find
  11. Manning, C., Surdeanu, M., Bauer, J., Finkel, J., Bethard, S., and McClosky, D. (2014). The stanford corenlp natural language processing toolkit. In Proceedings of 52nd annual meeting of the association for computational linguistics: system demonstrations, pages 55–60.
  12. Miller, G. A., Beckwith, R., Fellbaum, C., Gross, D., and Miller, K. J. (1990). Introduction to wordnet: An on-line lexical database. International journal of lexicography, 3(4):235–244. ↗
  13. Parikh, A. P., Täckström, O., Das, D., and Uszkoreit, J. (2016). A decomposable attention model for natural language inference. arXiv preprint arXiv:1606.01933. ↗
  14. Peyrard, M. (2019). A simple theoretical model of importance for summarization. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1059–1073. ↗
  15. Poliak, A., Naradowsky, J., Haldar, A., Rudinger, R., and Van Durme, B. (2018). Hypothesis only baselines in natural language inference. arXiv preprint arXiv:1805.01042. ↗
  16. Pomerleau, D. and Rao, D. (2017). Fake news challenge. find
  17. Romanov, A. and Shivade, C. (2018). Lessons from natural language inference in the clinical domain. arXiv preprint arXiv:1808.06752. ↗
  18. Schuster, T., Shah, D. J., Yeo, Y. J. S., Filizzola, D., Santus, E., and Barzilay, R. (2019). Towards debiasing fact verification models. arXiv preprint arXiv:1908.05267. ↗
  19. Suntwal, S., Paul, M., Sharp, R., and Surdeanu, M. (2019). On the importance of delexicalization for fact verification. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, (Short Papers), Hong Kong, November. Association for Computational Linguistics. find
  20. Thorne, J., Vlachos, A., Christodoulopoulos, C., and Mittal, A. (2018). Fever: a large-scale dataset for fact extraction and verification. arXiv preprint arXiv:1803.05355. ↗
  21. Williams, A., Nangia, N., and Bowman, S. R. (2017). A broad-coverage challenge corpus for sentence understanding through inference. arXiv preprint arXiv:1704.05426. ↗
  22. Yadav, V., Bethard, S., and Surdeanu, M. (2019). Alignment over heterogeneous embeddings for question answering. 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 2681–2691. find
  23. Zeman, D. and Resnik, P. (2008). Cross-language parser adaptation between related languages. In Proceedings of the IJCNLP-08 Workshop on NLP for Less Privileged Languages.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/panenghat-2020-necessity-debiasing-natural-language-inference/