Elamur Research⌕
☰
home / papers / 2022 / Uncovering More Shallow Heuristics: Probing the Natural Language Inference Capacities of Transformer-Based Pre-Trained Language Models Using Syllogistic Patterns

Uncovering More Shallow Heuristics: Probing the Natural Language Inference Capacities of Transformer-Based Pre-Trained Language Models Using Syllogistic Patterns

Paper · 2022 · Reto Gubelmann and Siegfried Handschuh

Summary

The paper asks whether transformer-based language models fine-tuned for natural language inference (NLI) actually reason or rely on surface cues. The authors build a synthetic test set from Aristotelian syllogistic, in which two quantified premises over three terms (subject, middle, predicate) either entail, contradict, or leave open a conclusion. Changing one quantifier can flip the label, so the set gives controlled contrasts. They test models fine-tuned on SNLI and MNLI and also critique MNLI itself. MNLI contains premises with no truth conditions, such as questions and conversational fragments, and it shows a strong negation bias. The best model, a DistilBERT checkpoint from textattack, reaches about 65% accuracy overall. Every model, however, scores below 10% on neutral items, far under the 33.3% chance level. Inspection shows that the models use symmetry as a cue. When premise and hypothesis match in negation and "some", they predict entailment. When they differ, they predict contradiction. The authors conclude that current models learn spurious heuristics rather than NLI. Scaling alone is unlikely to fix this.

Key points

  • The test data comes from syllogistic moods and figures. Aristotle identified 24 formally valid patterns, and the mood BARBARA is one example. Systematic variation produces many valid, contradictory and neutral samples.
  • Models evaluated include fine-tuned versions of BERT, RoBERTa, XLNet, DeBERTa, DistilBERT, DistilRoBERTa, MiniLM, Albert, T5 and BART, trained on SNLI and/or MNLI.
  • The top result was about 65% accuracy, from textattack's DistilBERT, one of the smallest models tested. No model exceeded 10% accuracy on neutral samples.
  • Two MNLI flaws are highlighted:
    • Alogical premises. Questions and fragments still received entailment or contradiction labels.
    • Negation bias. In about 48% of pairs containing a negation, the gold label is contradiction.
  • The diagnosed heuristic works as follows. Symmetry in negation and existential quantifier between premise and hypothesis leads to an "entailment" prediction. Asymmetry leads to "contradiction".

Interesting ideas

  • I1. Minimal edits that flip the gold label, such as swapping "All" for "No", separate real logical sensitivity from lexical overlap, because the surface stays nearly identical while the relation changes.
  • I2. Report accuracy separately for each label and slice the results by surface features (presence and symmetry of negation or quantifiers); a decent overall score can hide below-chance behaviour on one class.
  • I3. A statement can only enter entailment or contradiction relations if it has truth conditions, so questions and fragments should be filtered out or treated as neutral before inference.
  • I4. In syllogisms, validity is independent of whether the individual sentences are factually true, which makes them ideal for testing pure quantifier reasoning separately from world knowledge.
  • I5. Neutral items whose hypothesis almost entirely overlaps the premise are hard cases that defeat overlap heuristics and are good probes.

For Mova

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

  • I1 (learn): Generate such one-word contrast pairs as negative controls in evaluation, so that any shortcut ignoring the quantifier or negation is exposed.
  • I2 (answer): Break down test results by label and by structural features of the question, to catch readers that answer from cue patterns rather than from the world model.
  • I3 (frames): Use speech-act classification as a gate, so that only assertive frames feed logical inference while questions are routed to presupposition or implicature handling.
  • I4 (world): Build a set-inclusion and exclusion layer over entities and classes that checks All/No/Some relations formally, and test it on template-generated syllogisms with arbitrary or counterfactual terms.
  • I5 (learn): Include high-overlap neutral pairs in silver and evaluation data, so that repair rules or trained components cannot equate overlap with entailment.

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 St.GallenSaint Gallen · Switzerland3 works · 45 citations

📚 Shelves 5 Computer Science · 9,086 works Artificial Intelligence · 8,889 works Arts and Humanities · 9,296 works Language and Linguistics · 9,206 works Computer Vision and Pattern Recognition · 1,282 works

🧭 Topics 3 Language models & pretraining · 1,343 works Reasoning · 510 works Interpretability · 58 works

Cite

Cite as Reto Gubelmann and Siegfried Handschuh. 2022. Uncovering More Shallow Heuristics: Probing the Natural Language Inference Capacities of Transformer-Based Pre-Trained Language Models Using Syllogistic Patterns. arXiv:2201.07614.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2201.07614 · arXiv 2201.07614

@misc{gubelmann2022uncovering,
  title = {{Uncovering More Shallow Heuristics: Probing the Natural Language Inference Capacities of Transformer-Based Pre-Trained Language Models Using Syllogistic Patterns}},
  author = {Reto Gubelmann and Siegfried Handschuh},
  year = {2022},
  doi = {10.48550/arxiv.2201.07614},
  eprint = {2201.07614},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2201.07614}
}

References

39 references, 18 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Aristotle. 1984. Prior analytics. In Jonathan Barnes, editor, The Complete Works of Aristotle, pages 39– 113. Oxford University Press. find
  2. Dimion Asael, Zachary Ziegler, and Yonatan Belinkov. 2021. A generative approach for mitigating structural biases in natural language inference. arXiv preprint arXiv:2108.14006. ↗
  3. Jean-Philippe Bernardy and Stergios Chatzikyriakidis. 2019. What kind of natural language inference are nlp systems learning: Is this enough? In ICAART (2), pages 919–931. find
  4. Samuel R Bowman, Gabor Angeli, Christopher Potts, and Christopher D Manning. 2015. A large annotated corpus for learning natural language inference. In Conference on Empirical Methods in Natural Language Processing, EMNLP 2015, pages 632–642. Association for Computational Linguistics (ACL). ↗
  5. Ronan Le Bras, Swabha Swayamdipta, Chandra Bhagavatula, Rowan Zellers, Matthew E Peters, Ashish Sabharwal, and Yejin Choi. 2020. Adversarial filters of dataset biases. arXiv preprint arXiv:2002.04108. ↗
  6. Tiffany Chien and Jugal Kumar Kalita. 2020. Adversarial analysis of natural language inference systems. 2020 IEEE 14th International Conference on Semantic Computing (ICSC), pages 1–8. our card
  7. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT. our card
  8. Gottlob Frege. 1892. Über sinn und bedeutung. Zeitschrift für Philosophie und philosophische Kritik, 100:25–50. find
  9. Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. 2020. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665–673. ↗
  10. 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 Technologies, Volume 2 (Short Papers), pages 107–112. find
  11. He He, Sheng Zha, and Haohan Wang. 2019. Unlearn dataset bias in natural language inference by fitting the residual. arXiv preprint arXiv:1908.10763. ↗ our card
  12. Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2020. Deberta: Decoding-enhanced bert with disentangled attention. In International Conference on Learning Representations. ↗ our card
  13. Md Mosharaf Hossain, Venelin Kovatchev, Pranoy Dutta, Tiffany Kao, Elizabeth Wei, and Eduardo Blanco. 2020. An analysis of natural language inference benchmarks through the lens of negation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9106–9118. our card
  14. Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2019. Albert: A lite bert for self-supervised learning of language representations. In International Conference on Learning Representations. find
  15. Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. ArXiv, abs/1910.13461. ↗ our card
  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. Rabeeh Karimi Mahabadi, Yonatan Belinkov, and James Henderson. 2019. End-to-end bias mitigation by modelling biases in corpora. arXiv preprint arXiv:1909.06321. ↗ our card
  18. 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. our card
  19. Richard Moot and Christian Retoré. 2019. Natural language semantics and computability. Journal of Logic, Language and Information, 28(2):287–307. ↗
  20. John Morris, Eli Lifland, Jin Yong Yoo, Jake Grigsby, Di Jin, and Yanjun Qi. 2020. Textattack: A framework for adversarial attacks, data augmentation, and adversarial training in nlp. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 119–126. our card
  21. 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. find
  22. David S. Oderberg. 2005. Predicate logic and bare particulars. In David S. Oderberg, editor, The Old New Logic, pages 183–210. The MIT Press. find
  23. Jack W. Rae, Sebastian Borgeaud, and Trevor Cai et al. 2021. Scaling language models: Methods, analysis & insights from training gopher. DeepMind Company Publication. ↗
  24. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2019. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683. ↗ our card
  25. Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084. ↗ our card
  26. Marco Tulio Ribeiro, Tongshuang Wu, Carlos Guestrin, and Sameer Singh. 2020. Beyond accuracy: Behavioral testing of nlp models with checklist. arXiv preprint arXiv:2005.04118. ↗
  27. Kyle Richardson, Hai Hu, Lawrence Moss, and Ashish Sabharwal. 2020. Probing natural language inference models through semantic fragments. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8713–8721. ↗ our card
  28. Bertrand Russell. 1905. 14(56):479–493. On denoting. find
  29. Mind, Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108. ↗
  30. Prasetya Ajie Utama, Nafise Sadat Moosavi, and Iryna Gurevych. 2020. Towards debiasing nlu models from unknown biases. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7597–7610. find
  31. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In NIPS. our card
  32. Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. 2019. Universal adversarial triggers for attacking and analyzing NLP. 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 2153–2162. find
  33. Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. 2019. Superglue: A stickier benchmark for general-purpose language understanding systems. In NeurIPS. our card
  34. 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. ↗ our card
  35. Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. 2020. Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers. arXiv preprint arXiv:2002.10957. ↗
  36. Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112–1122. Association for Computational Linguistics. ↗
  37. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pier-ric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2019. Huggingface’s transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771. ↗ our card
  38. Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. 2019. Xlnet: Generalized autoregressive pretraining for language understanding. In Advances in neural information processing systems, pages 5753–5763. ↗ our card
  39. Xiang Zhou and Mohit Bansal. 2020. Towards robustifying nli models against lexical dataset biases. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8759– 8771. our card

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/gubelmann-2022-uncovering-more-shallow-heuristics-probing/