Elamur Research
☰
home / papers / 2019 / Probing Natural Language Inference Models through Semantic Fragments

Probing Natural Language Inference Models through Semantic Fragments

Written for the Research library (claude-opus-5-5) from the paper text. Original — arxiv.org/pdf/1909.07521

Summary

The paper asks whether strong NLI models can handle, or quickly acquire, reasoning skills such as boolean coordination, quantifiers, conditionals, comparatives, counting, and monotonicity. Existing benchmarks do not clearly measure these skills. The authors propose "semantic fragments": synthetic challenge sets produced from formal specifications, so that each set isolates one phenomenon and its difficulty can be controlled. They build eight fragments. Six cover elementary logic and are reworked from data by Salvatore, Finger, and Hirata Jr (2019). Two are new monotonicity sets based on van Benthem's monotonicity calculus and automatic polarity projection. Models trained on SNLI and MNLI, including BERT, score poorly on these fragments, even though the targeted phenomena are basic to inference. A short extra fine-tuning step fixes this. The authors call their variant of inoculation "lossless inoculation": it picks the configuration with the best combined score on the fragment and the original benchmark. With it, BERT learns every fragment in minutes and keeps its MNLI accuracy. The work offers a cheap, controlled way to find gaps in NLI models and patch them with generated data.

Key points

  • The fragment library has 8 sets. Six are logic fragments (negation, boolean, quantifier, counting, conditionals, comparatives) with 5,000 pairs each, mostly in a countries/travel domain. Two are monotonicity fragments in an animals domain.
  • The evaluation is organized around three questions:
    • Can a fragment be learned from scratch?
    • How do models pre-trained on SNLI+MNLI (about 942K pairs) perform on it?
    • Can these models be re-tuned quickly without forgetting the original task?
  • Existing architectures can learn the fragments from scratch, but the resulting models are of little practical value. Pre-trained models such as BERT perform badly on the new test sets.
  • After lossless inoculation, BERT masters all logic and monotonicity fragments with no real drop on MNLI. Decomp-Attn masters the monotonicity fragments but not the logic ones, losing about 6.7% on MNLI on average. Its failure on comparatives probably reflects its insensitivity to word order.
  • How costly re-training is depends on the phenomenon. Comparatives can hurt original performance, while monotonicity learning stays stable across models.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages

Cited by

86 papers of our library cite it (of the 140673 whose bibliographies we read); 5 of them have a page here, 81 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.1909.07521 · arXiv 1909.07521

@misc{richardson2019probing,
  title = {{Probing Natural Language Inference Models through Semantic Fragments}},
  author = {Richardson, Kyle and Hu, Hai and Moss, Lawrence S. and Sabharwal, Ashish},
  year = {2019},
  doi = {10.48550/arxiv.1909.07521},
  eprint = {1909.07521},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.1909.07521}
}

References

44 references, 11 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Bowman, S. R.; Angeli, G.; Potts, C.; and Manning, C. D. 2015. A Large Annotated Corpus for Learning Natural Language Inference. In EMNLP. find
  2. Chen, Q.; Zhu, X.; Ling, Z.-H.; Wei, S.; Jiang, H.; and Inkpen, D. 2017. Enhanced LSTM for Natural Language Inference. In ACL. find
  3. Chomsky, N. 1965. Aspects of the Theory of Syntax. MIT press. find
  4. Chrupała, G., and Alishahi, A. 2019. Correlating Neural and Symbolic Representations of Language. In ACL. find
  5. Conneau, A.; Kiela, D.; Schwenk, H.; Barrault, L.; and Bordes, A. 2017. Supervised Learning of Universal Sentence Representations from Natural Language Inference Data. In EMNLP. find our card
  6. Cooper, R.; Crouch, D.; Van Eijck, J.; Fox, C.; Van Genabith, J.; Jaspars, J.; Kamp, H.; Milward, D.; Pinkal, M.; Poesio, M.; et al. 1996. Using the Framework. Technical report, LRE 62-051 D-16, The FraCaS Consortium. find
  7. Dagan, I.; Glickman, O.; and Magnini, B. 2005. The PAS-CAL Recognising Textual Entailment Challenge. In Machine Learning Challenges Workshop. find
  8. Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In NAACL. find
  9. Gardner, M.; Grus, J.; Neumann, M.; Tafjord, O.; Dasigi, P.; Liu, N.; Peters, M.; Schmitz, M.; and Zettlemoyer, L. 2018. AllenNLP: A Deep Semantic Natural Language Processing Platform. In Workshop for NLP Open Source Software (NLP-OSS). find our card
  10. Geiger, A.; Cases, I.; Karttunen, L.; and Potts, C. 2018. Stress-Testing Neural Models of Natural Language Inference with Multiply-Quantified Sentences. arXiv:1810.13033. ↗
  11. Geiger, A.; Cases, I.; Karttunen, L.; and Potts, C. 2019. Posing Fair Generalization Tasks for Natural Language Inference. In EMNLP. find
  12. Glockner, M.; Shwartz, V.; and Goldberg, Y. 2018. Breaking NLI Systems with Sentences that Require Simple Lexical Inferences. In ACL. find our card
  13. Goldberg, Y. 2019. Assessing BERT’s Syntactic Abilities. arXiv:1901.05287. ↗ our card
  14. Gururangan, S.; Swayamdipta, S.; Levy, O.; Schwartz, R.; Bowman, S.; and Smith, N. A. 2018. Annotation Artifacts in Natural Language Inference Data. In NAACL. find
  15. Hu, H., and Moss, L. S. 2018. Polarity Computations in Flexible Categorial Grammar. In *SEM. find
  16. Hu, H.; Chen, Q.; Richardson, K.; Mukherjee, A.; Moss, L. S.; and Kuebler, S. 2019. Monalog: a Lightweight System for Natural Language Inference Based on Monotonicity. arXiv:1910.08772. ↗
  17. Hu, H.; Chen, Q.; and Moss, L. S. 2019. Natural Language Inference with Monotonicity. In IWCS. find
  18. Icard, T. F., and Moss, L. S. 2013. A Complete Calculus of Monotone and Antitone Higher-Order Functions. TACL. find
  19. Icard, T. F., and Moss, L. S. 2014. Recent Progress on Monotonicity. Linguistic Issues in Language Technology 9(7):167–194. find
  20. Kalouli, A.-L.; de Paiva, V.; and Real, L. 2017. Correcting Contradictions. In the Computing Natural Language Inference Workshop. find
  21. Lake, B. M., and Baroni, M. 2017. Generalization Without Systematicity: On the Compositional Skills of Sequence-to-Sequence Recurrent Networks. In ICML. find
  22. Liu, N. F.; Schwartz, R.; and Smith, N. A. 2019. Inoculation by Fine-Tuning: A Method for Analyzing Challenge Datasets. In NAACL. find
  23. LoBue, P., and Yates, A. 2011. Types of Common-sense Knowledge Needed for Recognizing Textual Entailment. In ACL. find
  24. McCoy, R. T.; Pavlick, E.; and Linzen, T. 2019. Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inference. In ACL. find our card
  25. Montague, R. 1973. The proper treatment of quantification in ordinary English. In Approaches to natural language. Springer. find
  26. Naik, A.; Ravichander, A.; Sadeh, N.; Rose, C.; and Neubig, G. 2018. Stress Test Evaluation for Natural Language Inference. In COLING. find
  27. Parikh, A. P.; Täckström, O.; Das, D.; and Uszkoreit, J. 2016. A Decomposable Attention Model for Natural Language Inference. In EMNLP. find our card
  28. Pavlick, E., and Kwiatkowski, T. 2019. Inherent Disagreements in Human Textual Inferences. TACL 7:677–694. find our card
  29. Poliak, A.; Haldar, A.; Rudinger, R.; Hu, J. E.; Pavlick, E.; White, A. S.; and Van Durme, B. 2018a. Collecting Diverse Natural Language Inference Problems for Sentence Representation Evaluation. In EMNLP. find our card
  30. Poliak, A.; Naradowsky, J.; Haldar, A.; Rudinger, R.; and Van Durme, B. 2018b. Hypothesis Only Baselines in Natural Language Inference. In *SEM. find
  31. Pratt-Hartmann, I. 2004. Fragments of Language. Journal of Logic, Language and Information 13(2):207–223. find
  32. Rozen, O.; Shwartz, V.; Aharoni, R.; and Dagan, I. 2019. Analyzing Generalization in Natural Language Inference via Controlled Variance in Adversarial Datasets. In CoNLL. find
  33. Salvatore, F.; Finger, M.; and Hirata Jr, R. 2019. Using Syntactical and Logical Forms to Evaluate Textual Inference Competence. arXiv:1905.05704. ↗
  34. Steedman, M. 2000. The Syntactic Process. MIT Press. find
  35. van Benthem, J. 1986. Essays in Logical Semantics, volume 29 of Studies in Linguistics and Philosophy. Dordrecht: D. Reidel Publishing Co. find
  36. Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, Ł.; and Polosukhin, I. 2017. Attention is all you need. In NIPS, 5998–6008. find
  37. Wang, A.; Singh, A.; Michael, J.; Hill, F.; Levy, O.; and Bowman, S. R. 2018. Glue: A Multi-task Benchmark and Analysis Platform for Natural Language Understanding. In EMNLP Workshop BlackboxNLP. find our card
  38. Warstadt, A.; Cao, Y.; Grosu, I.; Peng, W.; Blix, H.; Nie, Y.; Alsop, A.; Bordia, S.; Liu, H.; Parrish, A.; et al. 2019. Investigating BERT’s Knowledge of Language: Five Analysis Methods with NPIs. In EMNLP. find
  39. Weston, J.; Bordes, A.; Chopra, S.; Rush, A. M.; van Merriënboer, B.; Joulin, A.; and Mikolov, T. 2015. Towards AI-Complete Question Answering: A Set of Prerequisite Toy Tasks. arXiv:1502.05698. ↗ our card
  40. White, A. S.; Rastogi, P.; Duh, K.; and Van Durme, B. 2017. Inference is Everything: Recasting Semantic Resources into a Unified Evaluation Framework. In IJCNLP. find our card
  41. Williams, A.; Nangia, N.; and Bowman, S. 2018. A Broad-Coverage Challenge Corpus for Sentence Understanding through Inference. In NAACL. find
  42. Yanaka, H.; Mineshima, K.; Bekki, D.; Inui, K.; Sekine, S.; Abzianidze, L.; and Bos, J. 2019a. Can Neural Networks Understand Monotonicity Reasoning? In ACL Workshop BlackboxNLP. find
  43. Yanaka, H.; Mineshima, K.; Bekki, D.; Inui, K.; Sekine, S.; Abzianidze, L.; and Bos, J. 2019b. HELP: A Dataset for Identifying Shortcomings of Neural Models in Monotonicity Reasoning. In *SEM. find
  44. Zaenen, A.; Karttunen, L.; and Crouch, R. 2005. Local Textual Inference: Can it be Defined or Circumscribed? In the ACL workshop on empirical modeling of semantic equivalence and entailment. find