Elamur Research⌕
☰
home / papers / 2024 / How Hard is this Test Set? NLI Characterization by Exploiting Training Dynamics

How Hard is this Test Set? NLI Characterization by Exploiting Training Dynamics

Paper · 2024 · Adrian Cosma, Stefan Ruseti, Mihai Dascalu et al.

Summary

Benchmark NLI datasets such as SNLI and MultiNLI contain annotation artifacts. Models can exploit these shortcuts, which makes reported accuracy look better than it should. As one example, a RoBERTa model that sees only the hypothesis still scores far above chance. The authors build a harder test set automatically, without writing artificial examples. They fine-tune an encoder directly on the test set in two settings: with the full premise+hypothesis input and with the hypothesis alone. For every example they record four training-dynamics statistics in each setting. A Gaussian Mixture Model then clusters these eight features into three tiers: easy, ambiguous and hard. This replaces the fixed percentile cut-offs used by Data Maps.

They apply the method to SNLI, MultiNLI and FEVER, the last recast as NLI. The hard tier shows much weaker spurious-correlation signals and much lower accuracy, and the split holds for both RoBERTa and DeBERTa. Used on training data, the same procedure keeps a fraction of the examples and still matches full-data performance on the Naik et al. stress test, beating Data Maps and AUM. This gives a cheaper and more honest view of NLU ability.

Key points

  • A hypothesis-only RoBERTa reaches 71.7% on SNLI and 61.4% on MultiNLI, against a 33% chance level.
  • Each example gets 8 training-dynamics features, 4 from the P+H model and 4 from the H-only model, clustered by a GMM into easy, ambiguous and hard. KMeans was rejected because it assumes spherical clusters.
  • Datasets are SNLI, MultiNLI and FEVER, and the encoders are RoBERTa and DeBERTa. Accuracy drops sharply on the "hard" split for both models, and difficulty assignments such as the "Contains Negation" heuristic stay consistent across them.
  • For training-set filtering, about 33% of SNLI and 59% of MultiNLI give results comparable to the full data on the Naik et al. (2018) stress test:
    • MultiNLI Amb+Hard scores 0.747 overall, versus 0.751 for all data, 0.722 for Data Maps and 0.712 for AUM.
    • The SNLI ambiguous subset scores 0.709, slightly above full data at 0.704.
  • Code is released at github.com/cosmaadrian/nli-stress-test.

Interesting ideas

  • I1. The authors fit a model on the evaluation set itself, only to watch how quickly and how stably each item is learned, and use that as a difficulty label instead of hand-building adversarial items.
  • I2. A deliberately handicapped shortcut model (hypothesis-only) serves as a probe: items it masters easily are suspected of carrying artifacts rather than requiring reasoning.
  • I3. A Gaussian Mixture Model over several per-example features replaces a single-feature percentile threshold, so tier boundaries come from the data rather than arbitrary cut-offs.
  • I4. Discarding the easy tier and training on a smaller, harder subset can match or beat training on everything, so data quality can matter more than quantity.
  • I5. The paper evaluates on stress-test categories (negation, numerical reasoning, antonyms, word overlap) and does not settle for a single accuracy figure.

For Mova

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

  • I1 (learn): Mova's evaluation sets could be split into easy, ambiguous and hard tiers this way, with scores reported per tier so gains on trivial items stop masking stagnation on hard ones.
  • I2 (answer): for school tests and fable QA, an "answer-options-only" or "question-only" baseline would flag questions solvable without reading the text, acting as a negative control.
  • I3 (learn): disagreement statistics among witness parsers on silver sentences could be clustered the same way to separate trustworthy, contested and noisy silver data.
  • I4 (tree): gparse or the perceptron parser could be retrained on treebank or silver sentences selected by their training dynamics, cutting training cost without losing LAS.
  • I5 (frames): frame and role extraction could be scored separately on sentences with negation, quantities or antonyms, to reveal which phenomena the pragmatics layer actually handles.

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 Illinois ChicagoChicago · United States61 works · 20,107 citations🎓Universitatea Națională de Știință și Tehnologie Politehnica BucureștiBucharest · Romania5 works · 265 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 1Research subjects and languages the work is about.Reasoning · 1,426 works

Cite

Cite as Adrian Cosma, Stefan Ruseti, Mihai Dascalu and Cornelia Caragea. 2024. How Hard is this Test Set? NLI Characterization by Exploiting Training Dynamics. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2024.emnlp-main.175

@inproceedings{cosma2024how,
  title = {{How Hard is this Test Set? NLI Characterization by Exploiting Training Dynamics}},
  author = {Adrian Cosma, Stefan Ruseti, Mihai Dascalu and Cornelia Caragea},
  booktitle = {Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing},
  year = {2024},
  doi = {10.18653/v1/2024.emnlp-main.175},
  url = {https://doi.org/10.18653/v1/2024.emnlp-main.175}
}

References

34 references, 21 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Lasha Abzianidze, Johannes Bjerva, Kilian Evang, Hessel Haagsma, Rik van Noord, Pierre Ludmann, Duc-Duy Nguyen, and Johan Bos. 2017. The Parallel Meaning Bank: Towards a multilingual corpus of translations annotated with compositional meaning representations. In Proceedings of EACL, pages 242– 247, Valencia, Spain. Association for Computational Linguistics. ↗
  2. Our method is unsuitable for automatically identifying mislabeled examples in a dataset. While it does incorporate measures such as Area Under Margin (Pleiss et al., 2020), designed with this purpose, proper manual verification is needed to increase annotation quality. find
  3. 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 (EMNLP). Association for Computational Linguistics. ↗
  4. Ido Dagan, Bill Dolan, Bernardo Magnini, and Dan Roth. 2009. Recognizing textual entailment: Rational, evaluation and approaches. Natural Language Engineering, 15(4):i–xvii. ↗
  5. Ido Dagan, Oren Glickman, and Bernardo Magnini. 2005. The pascal recognising textual entailment challenge. In Proceedings of the First International Conference on Machine Learning Challenges: Evaluating Predictive Uncertainty Visual Object Classification, and Recognizing Textual Entailment, MLCW’05, page 177–190, Berlin, Heidelberg. Springer-Verlag. find
  6. Dorottya Demszky, Kelvin Guu, and Percy Liang. 2018. Transforming question answering datasets into natural language inference datasets. CoRR, abs/1809.02922. ↗
  7. Christiane Fellbaum, editor. 1998. WordNet: An Electronic Lexical Database. Language, Speech, and Communication. MIT Press, Cambridge, MA. find
  8. Atticus Geiger, Ignacio Cases, Lauri Karttunen, and Christopher Potts. 2018. Stress-testing neural models of natural language inference with multiply-quantified sentences. arXiv preprint arXiv:1810.13033. ↗
  9. Max Glockner, Vered Shwartz, and Yoav Goldberg. 2018. Breaking NLI systems with sentences that require simple lexical inferences. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 650–655, Melbourne, Australia. Association for Computational Linguistics. ↗
  10. Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. 2017. On calibration of modern neural networks. In International conference on machine learning, pages 1321–1330. PMLR. ↗
  11. 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 NA-ACL, pages 107– 112, New Orleans, Louisiana. Association for Computational Linguistics.
  12. Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2022. Debertav3: Improving deberta using electra-style pre-training with gradient-disentangled embedding sharing. In The Eleventh International Conference on Learning Representations. ↗
  13. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685. ↗
  14. Noah Lee, Na Min An, and James Thorne. 2023. Can large language models capture dissenting human voices? In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 4569–4585. find
  15. Tianyu Liu, Xin Zheng, Baobao Chang, and Zhifang Sui. 2020. Hyponli: Exploring the artificial patterns of hypothesis-only bias in natural language inference. arXiv preprint arXiv:2003.02756. ↗
  16. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. ↗
  17. Henry B Mann and Donald R Whitney. 1947. On a test of whether one of two random variables is stochastically larger than the other. The annals of mathematical statistics, pages 50–60. ↗
  18. Aakanksha Naik, Abhilasha Ravichander, Norman Sadeh, Carolyn Rose, and Graham Neubig. 2018. Stress test evaluation for natural language inference. In Proceedings of the 27th International Conference on Computational Linguistics, pages 2340–2353, Santa Fe, New Mexico, USA. Association for Computational Linguistics. ↗
  19. Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2019. Adversarial nli: A new benchmark for natural language understanding. arXiv preprint arXiv:1910.14599. ↗
  20. Geoff Pleiss, Tianyi Zhang, Ethan Elenberg, and Kilian Q Weinberger. 2020. Identifying mislabeled data using the area under the margin ranking. Advances in Neural Information Processing Systems, 33:17044– 17056. ↗
  21. 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 Computational Linguistics. ↗
  22. Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of EMNLP, pages 2383–2392, Austin, Texas. Association for Computational Linguistics. ↗
  23. Douglas A. Reynolds. 2009. Gaussian mixture models. In Encyclopedia of Biometrics. ↗
  24. Mobashir Sadat and Cornelia Caragea. 2022. SciNLI: A corpus for natural language inference on scientific text. In Proceedings of ACL, pages 7399–7409, Dublin, Ireland. Association for Computational Linguistics. ↗
  25. Swarnadeep Saha, Yixin Nie, and Mohit Bansal. 2020. Conjnli: Natural language inference over conjunctive sentences. arXiv preprint arXiv:2010.10418. ↗
  26. Christian Stab, Tristan Miller, and Iryna Gurevych. 2018. Cross-topic argument mining from heterogeneous sources using attention-based neural networks. arXiv preprint arXiv:1802.05758. ↗
  27. Shane Storks, Qiaozi Gao, and Joyce Y Chai. 2019. Recent advances in natural language inference: A survey of benchmarks, resources, and approaches. arXiv preprint arXiv:1904.01172. ↗
  28. Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie, Yizhong Wang, Hannaneh Hajishirzi, Noah A Smith, and Yejin Choi. 2020. Dataset cartography: Mapping and diagnosing datasets with training dynamics. arXiv preprint arXiv:2009.10795. ↗
  29. James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. FEVER: a large-scale dataset for fact extraction and VERification. In Proceedings of NA-ACL, pages 809–819, New Orleans, Louisiana. Association for Computational Linguistics.
  30. Masatoshi Tsuchiya. 2018. Performance impact caused by hidden bias of training data for recognizing textual entailment. In Proceedings of LREC 2018, Miyazaki, Japan. European Language Resources Association (ELRA).
  31. Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353–355, Brussels, Belgium. Association for Computational Linguistics. ↗
  32. Sean Welleck, Jason Weston, Arthur Szlam, and Kyunghyun Cho. 2019. Dialogue natural language inference. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3731–3741, Florence, Italy. Association for Computational Linguistics. ↗
  33. Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of NA-ACL, pages 1112–1122. Association for Computational Linguistics.
  34. Hitomi Yanaka, Koji Mineshima, Daisuke Bekki, Kentaro Inui, Satoshi Sekine, Lasha Abzianidze, and Johan Bos. 2019. Help: A dataset for identifying shortcomings of neural models in monotonicity reasoning. arXiv preprint arXiv:1904.12166. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/cosma-2024-how-hard-is-this-test/