Preprint / Version 1

What Can a Model Learn Without the Premise? Explaining Hypothesis-Only Shortcuts in Natural Language Inference

Authors

  • Sai Sharan Nellutla Sierra High School

DOI:

https://doi.org/10.58445/6se5mt57

Keywords:

natural language inference, annotation artifacts, shortcut learning, explainable AI, SHAP, dataset bias, feature attribution, interpretability, Artificial Intelligence

Abstract

Natural Language Inference (NLI) focuses on predicting whether the relationship between two sentences is one of contradiction, neutrality, or entailment. A good classifier needs to use both sentences, but in many cases the hypothesis alone is enough to infer the label. This paper examines how well classifiers that use only the hypothesis can solve the problem, which words they use to make decisions, and whether those words causally affect the decision. With only TF-IDF features from the hypothesis, logistic regression reaches 64.0% test accuracy, compared to a majority baseline of 34.3%, which is 83% of the gain achieved by a similar model that also sees the premise (70.1%). Most of the information comes from single words: unigrams alone reach 63.9%. The key cues are negations and words such as "sleeping" for contradiction, vague words such as "outdoors" for entailment, and added detail such as "tall" for neutral.

Four explanation methods disagree on the most important words, since SHAP and random forest importances prioritize frequent words while Naive Bayes ratios prioritize rare ones. Removing the five highest-weighted words (nobody, nothing, no, because, sad) from test hypotheses lowers accuracy 0.77 percentage points more than removing frequency-matched random words, and this extra drop grows to 9.1 points at 100 words. On the 36% of test examples where the hypothesis-only classifier makes mistakes, the classifier that also sees the premise reaches only 31.7% accuracy, compared to 91.6% on the rest. These shortcuts transfer poorly to MultiNLI. A significant amount of success on NLI can therefore come from shortcuts in how the data was written, and the choice of explanation method strongly affects which shortcuts are found.

References

I. Dagan, O. Glickman, and B. Magnini, "The PASCAL recognising textual entailment challenge," in Machine Learning Challenges: Evaluating Predictive Uncertainty, Visual Object Classification, and Recognising Textual Entailment, Lecture Notes in Computer Science, vol. 3944. Berlin, Germany: Springer, 2006, pp. 177–190, doi: 10.1007/11736790_9.

S. R. Bowman, G. Angeli, C. Potts, and C. D. Manning, "A large annotated corpus for learning natural language inference," in Proc. 2015 Conf. Empirical Methods in Natural Language Processing (EMNLP), Lisbon, Portugal, 2015, pp. 632–642, doi: 10.18653/v1/D15-1075.

A. Williams, N. Nangia, and S. R. Bowman, "A broad-coverage challenge corpus for sentence understanding through inference," in Proc. 2018 Conf. North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), New Orleans, LA, USA, 2018, pp. 1112–1122, doi: 10.18653/v1/N18-1101.

S. Gururangan, S. Swayamdipta, O. Levy, R. Schwartz, S. R. Bowman, and N. A. Smith, "Annotation artifacts in natural language inference data," in Proc. 2018 Conf. North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), Vol. 2 (Short Papers), New Orleans, LA, USA, 2018, pp. 107–112, doi: 10.18653/v1/N18-2017.

A. Poliak, J. Naradowsky, A. Haldar, R. Rudinger, and B. Van Durme, "Hypothesis only baselines in natural language inference," in Proc. Seventh Joint Conf. Lexical and Computational Semantics (*SEM), New Orleans, LA, USA, 2018, pp. 180–191, doi: 10.18653/v1/S18-2023.

M. Tsuchiya, "Performance impact caused by hidden bias of training data for recognizing textual entailment," in Proc. Eleventh Int. Conf. Language Resources and Evaluation (LREC), Miyazaki, Japan, 2018. [Online]. Available: https://aclanthology.org/L18-1239/

R. Geirhos, J.-H. Jacobsen, C. Michaelis, R. Zemel, W. Brendel, M. Bethge, and F. A. Wichmann, "Shortcut learning in deep neural networks," Nature Machine Intelligence, vol. 2, no. 11, pp. 665–673, 2020, doi: 10.1038/s42256-020-00257-z.

T. McCoy, E. Pavlick, and T. Linzen, "Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference," in Proc. 57th Annu. Meeting of the Association for Computational Linguistics (ACL), Florence, Italy, 2019, pp. 3428–3448, doi: 10.18653/v1/P19-1334.

Y. Nie, A. Williams, E. Dinan, M. Bansal, J. Weston, and D. Kiela, "Adversarial NLI: A new benchmark for natural language understanding," in Proc. 58th Annu. Meeting of the Association for Computational Linguistics (ACL), 2020, pp. 4885–4901, doi: 10.18653/v1/2020.acl-main.441.

M. T. Ribeiro, S. Singh, and C. Guestrin, "‘Why should I trust you?’ Explaining the predictions of any classifier," in Proc. 22nd ACM SIGKDD Int. Conf. Knowledge Discovery and Data Mining (KDD), San Francisco, CA, USA, 2016, pp. 1135–1144, doi: 10.1145/2939672.2939778.

S. M. Lundberg and S.-I. Lee, "A unified approach to interpreting model predictions," in Advances in Neural Information Processing Systems 30 (NeurIPS), Long Beach, CA, USA, 2017, pp. 4765–4774. [Online]. Available: https://arxiv.org/abs/1705.07874

S. Krishna, T. Han, A. Gu, S. Wu, S. Jabbari, and H. Lakkaraju, "The disagreement problem in explainable machine learning: A practitioner's perspective," arXiv preprint arXiv:2202.01602, 2022. [Online]. Available: https://arxiv.org/abs/2202.01602

J. Adebayo, J. Gilmer, M. Muelly, I. Goodfellow, M. Hardt, and B. Kim, "Sanity checks for saliency maps," in Advances in Neural Information Processing Systems 31 (NeurIPS), Montréal, Canada, 2018, pp. 9505–9515. [Online]. Available: https://arxiv.org/abs/1810.03292

J. Mu and J. Andreas, "Compositional explanations of neurons," in Advances in Neural Information Processing Systems 33 (NeurIPS), 2020, pp. 17153–17163. [Online]. Available: https://arxiv.org/abs/2006.14032

D. Kaushik and Z. C. Lipton, "How much reading does reading comprehension require? A critical investigation of popular benchmarks," in Proc. 2018 Conf. Empirical Methods in Natural Language Processing (EMNLP), Brussels, Belgium, 2018, pp. 5010–5015, doi: 10.18653/v1/D18-1546.

T. Schuster, D. Shah, Y. J. S. Yeo, D. Roberto Filizzola Ortiz, E. Santus, and R. Barzilay, "Towards debiasing fact verification models," in Proc. 2019 Conf. Empirical Methods in Natural Language Processing and 9th Int. Joint Conf. Natural Language Processing (EMNLP-IJCNLP), Hong Kong, China, 2019, pp. 3419–3425, doi: 10.18653/v1/D19-1341.

Y. Belinkov, A. Poliak, S. Shieber, B. Van Durme, and A. Rush, "Don't take the premise for granted: Mitigating artifacts in natural language inference," in Proc. 57th Annu. Meeting of the Association for Computational Linguistics (ACL), Florence, Italy, 2019, pp. 877–891, doi: 10.18653/v1/P19-1084.

C. Clark, M. Yatskar, and L. Zettlemoyer, "Don't take the easy way out: Ensemble based methods for avoiding known dataset biases," in Proc. 2019 Conf. Empirical Methods in Natural Language Processing and 9th Int. Joint Conf. Natural Language Processing (EMNLP-IJCNLP), Hong Kong, China, 2019, pp. 4069–4082, doi: 10.18653/v1/D19-1418.

L. S. Shapley, "A value for n-person games," in Contributions to the Theory of Games, Vol. II, H. W. Kuhn and A. W. Tucker, Eds. Princeton, NJ, USA: Princeton Univ. Press, 1953, pp. 307–317, doi: 10.1515/9781400881970-018.

L. Breiman, "Random forests," Machine Learning, vol. 45, no. 1, pp. 5–32, 2001, doi: 10.1023/A:1010933404324.

S. Jain and B. C. Wallace, "Attention is not explanation," in Proc. 2019 Conf. North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), Minneapolis, MN, USA, 2019, pp. 3543–3556, doi: 10.18653/v1/N19-1357.

S. Hooker, D. Erhan, P.-J. Kindermans, and B. Kim, "A benchmark for interpretability methods in deep neural networks," in Advances in Neural Information Processing Systems 32 (NeurIPS), Vancouver, Canada, 2019, pp. 9737–9748. [Online]. Available: https://arxiv.org/abs/1806.10758

D. Bau, B. Zhou, A. Khosla, A. Oliva, and A. Torralba, "Network dissection: Quantifying interpretability of deep visual representations," in Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, USA, 2017, pp. 6541–6549, doi: 10.1109/CVPR.2017.354.

Q. Lhoest et al., "Datasets: A community library for natural language processing," in Proc. 2021 Conf. Empirical Methods in Natural Language Processing: System Demonstrations, Punta Cana, Dominican Republic, 2021, pp. 175–184, doi: 10.18653/v1/2021.emnlp-demo.21.

F. Pedregosa et al., "Scikit-learn: Machine learning in Python," Journal of Machine Learning Research, vol. 12, pp. 2825–2830, 2011. [Online]. Available: https://jmlr.org/papers/v12/pedregosa11a.html

G. Salton and C. Buckley, "Term-weighting approaches in automatic text retrieval," Information Processing & Management, vol. 24, no. 5, pp. 513–523, 1988, doi: 10.1016/0306-4573(88)90021-0.

A. McCallum and K. Nigam, "A comparison of event models for naive Bayes text classification," in Proc. AAAI-98 Workshop on Learning for Text Categorization, Madison, WI, USA, 1998, pp. 41–48. [Online]. Available: https://www.cs.tufts.edu/~nr/cs257/archive/andrew-mccallum/event-models.pdf

C. Cortes and V. Vapnik, "Support-vector networks," Machine Learning, vol. 20, no. 3, pp. 273–297, 1995, doi: 10.1007/BF00994018.

C. Spearman, "The proof and measurement of association between two things," The American Journal of Psychology, vol. 15, no. 1, pp. 72–101, 1904, doi: 10.2307/1412159.

Downloads

Posted

2026-11-15