Elamur Research⌕
☰
home / papers / 2019 / Posing Fair Generalization Tasks for Natural Language Inference

Posing Fair Generalization Tasks for Natural Language Inference

Paper · 2019 · Atticus Geiger, Ignacio Cases, Lauri Karttunen et al.

Summary

Adversarial and synthetic test sets often show that NLI models with strong scores on naturalistic corpora break down on systematic cases. A failure only means something, though, if the training data could in principle have taught the required generalization. Geiger et al. make this precise. They model a task as a composition tree, where each node has a domain and a function and the answer is computed bottom-up. A trivial baseline learner memorizes every node function from fully labelled trees, and this learner serves as the yardstick for fairness: a split between training and test data counts as fair when the memorizer can reproduce the target function without error. A sufficient condition is that each node function meets every possible input during training. Using the seven natural-logic relations of MacCartney and Manning, the authors build difficult but provably fair artificial NLI datasets. Standard neural architectures fail to generalize on them. Only the task-specific CompTreeNN and CompTreeNTN, which compose aligned premise and hypothesis jointly, perform well, and even they are not perfect. The authors conclude that encoding the two sentences separately creates a bottleneck that blocks inference.

Key points

  • Intermediate values are semantic relations between phrase pairs, drawn from the seven basic relations: equivalence, forward entailment, reverse entailment, negation, alternation, cover and independence. The composing functions are operators such as quantifiers and negation.
  • The models compared are CBoW, TreeNN, LSTM Encoder and Attention LSTM (standard), plus CompTreeNN and CompTreeNTN (task-specific, with a hard-coded aligned tree). Only the last two succeed on the true generalization task.
  • When supervision of intermediate nodes is withheld, every model fails, which is expected because the task is then unfair. In the hardest setting, CBoW is the only standard model that improves significantly with intermediate predictions.
  • The authors vary the splits from "hardest" to "easiest", where the training data redundantly encodes the test set (mean of 5 runs).
    • Standard models succeed only near pure memorization.
    • CBoW cannot even memorize, because it ignores word order.
  • Growing representations from 100 to 200 dimensions changes nothing, so the authors attribute the failure to architecture, not capacity. They also argue (Appendix A) that earlier held-out-pattern tasks, such as Bowman 2013 and the De Morgan tests of Veldhoen and Zuidema, are unfair.

Interesting ideas

  • I1. Before blaming a learner for failing a held-out split, run a simple symbolic memorizer on the same split to prove the split is solvable at all. This separates model limits from data limits.
  • I2. A dataset is fair when every compositional operator has seen every possible combination of intermediate inputs, which gives a checkable coverage criterion for designing training data.
  • I3. Inference can be treated as computing a small set of natural-logic relations between aligned phrase pairs and composing them upward through quantifiers and negation. This keeps every step inspectable.
  • I4. Models that encode premise and hypothesis separately lose the interactions that inference needs, while joint, aligned composition recovers them.
  • I5. A graded series of splits, from true generalization to near memorization, shows where a learner stops generalizing and starts merely remembering.

For Mova

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

  • I1 (learn): Add a "fairness control" to evaluation runs: if a lookup-based oracle over the training frames or trees cannot solve the test split, the split is reported as unfair rather than counted as a model failure.
  • I2 (answer): When generating school-test items or maths word problems, enumerate operator × argument-type combinations and verify each appears in training before testing novel compositions.
  • I3 (answer): Yes/no and entailment questions could be answered by aligning the question's UD tree with the text's tree and propagating relation labels node by node, with each label kept as an explanation.
  • I4 (frames): When matching a question frame against text frames, compare them role by role through explicit alignment rather than reducing each to one summary score.
  • I5 (learn): Evaluate gparse or the self-induced repair rules on splits with controlled overlap to training, to see whether gains come from real generalization or memorized patterns.

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

🎓Stanford UniversityStanford · United States310 works · 284,480 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 1Research subjects and languages the work is about.Reasoning · 510 works

Cite

Cite as Atticus Geiger, Ignacio Cases, Lauri Karttunen and Christopher Potts. 2019. Posing Fair Generalization Tasks for Natural Language Inference. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/d19-1456

@inproceedings{geiger2019posing,
  title = {{Posing Fair Generalization Tasks for Natural Language Inference}},
  author = {Atticus Geiger, Ignacio Cases, Lauri Karttunen and Christopher Potts},
  booktitle = {Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)},
  year = {2019},
  doi = {10.18653/v1/d19-1456},
  url = {https://doi.org/10.18653/v1/d19-1456}
}

References

30 references, 9 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Johan van Benthem. 2008. A brief history of natural logic. In Logic, Navya-Nyaya and Applications: Homage to Bimal Matilal. find
  2. Samuel R. Bowman. 2013. Can recursive neural tensor networks learn logical reasoning? CoRR, abs/1312.6192. ↗
  3. Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015a. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632–642, Lisbon, Portugal. Association for Computational Linguistics. ↗
  4. Samuel R. Bowman, Christopher Potts, and Christopher D. Manning. 2015b. Recursive neural networks can learn logical semantics. In Proceedings of the 3rd Workshop on Continuous Vector Space Models and their Compositionality, pages 12–21. Association for Computational Linguistics. find
  5. Ishita Dasgupta, Demi Guo, Andreas Stuhlmüller, Samuel J. Gershman, and Noah D. Goodman. 2018. Evaluating compositionality in sentence embeddings. CoRR, abs/1802.04302. ↗
  6. Jeffrey L. Elman. 1990. Finding structure in time. Cognitive Science, 14(2):179–211. ↗
  7. Richard Evans, David Saxton, David Amos, Pushmeet Kohli, and Edward Grefenstette. 2018. Can neural networks understand logical entailment? CoRR, abs/1802.08535. ↗
  8. Atticus Geiger, Ignacio Cases, Lauri Karttunen, and Christopher Potts. 2018. Stress-testing neural models of natural language inference with multiply-quantified sentences. Ms., Stanford University. 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. Association for Computational Linguistics.
  10. Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. 2015. Explaining and harnessing adversarial examples. In ICLR. ↗
  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 the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 107–112. Association for Computational Linguistics. find
  12. Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural Comput., 9(8):1735– 1780. ↗
  13. Thomas F. Icard and Lawrence S. Moss. 2013. Recent progress on monotonicity. Linguistic Issues in Language Technology, 9(7):1–31. ↗
  14. Robin Jia and Percy Liang. 2017. Adversarial examples for evaluating reading comprehension systems. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2021–2031, Copenhagen, Denmark. Association for Computational Linguistics. ↗
  15. Diederik P. Kingma and Jimmy Ba. 2014. A method for stochastic optimization. abs/1412.6980. Adam: CoRR, Brenden M. Lake and Marco Baroni. 2017. Still not systematic after all these years: On the compositional skills of sequence-to-sequence recurrent networks. CoRR, abs/1711.00350. ↗
  16. Bill MacCartney and Christopher D. Manning. 2007. Natural logic for textual inference. In Proceedings of the ACL-PASCAL Workshop on Textual Entailment and Paraphrasing, RTE ’07, pages 193–200, Stroudsburg, PA, USA. Association for Computational Linguistics. find
  17. Bill MacCartney and Christopher D. Manning. 2009. An extended model of natural logic. In Proceedings of the Eight International Conference on Computational Semantics, pages 140–156. Association for Computational Linguistics. find
  18. W. McCune. 2005–2010. Prover9 and Mace4. http://www.cs.unm.edu/˜mccune/prover9/. ↗
  19. Aakanksha Naik, Abhilasha Ravichander, Norman Sadeh, Carolyn Rose, and Graham Neubig. 2018. Stress test evaluation for natural language inference. arXiv preprint arXiv:1806.00692. ↗
  20. Yixin Nie, Yicheng Wang, and Mohit Bansal. 2018. Analyzing compositionality-sensitivity of NLI models. CoRR, abs/1811.07033. ↗
  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. Association for Computational Linguistics. find
  22. Tim Rocktäschel, Edward Grefenstette, Karl Moritz Hermann, Tomás Kociský, and Phil Blunsom. 2015. Reasoning about entailment with neural attention. CoRR, abs/1509.06664. ↗
  23. V.M.S. Sánchez-Valencia. 1991. Studies on Natural Logic and Categorial Grammar. Universiteit van Amsterdam. find
  24. Richard Socher, Cliff Chiung-Yu Lin, Andrew Y. Ng, and Christopher D. Manning. 2011a. Parsing natural scenes and natural language with recursive neural networks. In Proceedings of the 28th International Conference on International Conference on Machine Learning, ICML’11, pages 129–136, USA. Omnipress. find
  25. Richard Socher, Jeffrey Pennington, Eric H. Huang, Andrew Y. Ng, and Christopher D. Manning. 2011b. Semi-supervised recursive autoencoders for predicting sentiment distributions. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, EMNLP ’11, pages 151–161, Stroudsburg, PA, USA. Association for Computational Linguistics.
  26. Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631–1642, Seattle, Washington, USA. Association for Computational Linguistics. ↗
  27. Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. 2014. Intriguing properties of neural networks. In International Conference on Learning Representations. ↗
  28. Masatoshi Tsuchiya. 2018. Performance impact caused by hidden bias of training data for recognizing textual entailment. CoRR, abs/1804.08117. ↗
  29. Sara Veldhoen and Willem Zuidema. 2018. Can neural networks learn logical reasoning? In Proceedings of the Conference on Logic and Machine Learning in Natural Language (LaML 2017), pages 35–41. Association for Computational Linguistics. find
  30. 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:1904.12166. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/geiger-2019-posing-fair-generalization-tasks-natural/