Non-entailed subsequences as a challenge for natural language inference
Paper · 2018 · R. Thomas McCoy and Tal Linzen
Summary
The paper asks whether strong NLI models actually understand sentences or simply exploit shallow cues in crowdsourced data. The authors name one suspected shortcut, the subsequence heuristic: a model assumes that any word sequence taken from the premise is entailed by it. To test this, they build a small challenge set from NP/S sentences. These use verbs such as believed that can take either a noun phrase or a whole clause as complement, so the hypothesis cuts the clause short (e.g., believing "the story was awful" does not imply believing "the story"). Premises come from a template grounded in the MNLI training corpus and are validated by Mechanical Turk workers. Five BiLSTM models are tested, each pre-trained on a different task and then trained on MNLI. All of them score far below chance. Most of their few correct answers disappear once negation words are removed, which shows that a second shortcut is at work. The study shows that high benchmark accuracy can hide reliance on systematic heuristics, and it argues for controlled diagnostic test sets.
Key points
- Five bidirectional LSTM encoders follow the two-stage setup of Wang et al. (2018). Each is first pre-trained on one of five options: MNLI, CCG supertagging (CCGbank), caption-to-image generation (MS COCO), language modeling (WikiText-103), or nothing (random encoder). Each is then trained on MNLI with a three-way classifier.
- The data set is built from the template NP₁ V₁ S₁ with the verbs heard, believed, felt and claimed. Every subject–verb, verb–object and clause combination was attested in the MNLI training data.
- The authors generated 200 pairs, and three MTurk annotators judged each one. 88 were kept: those where two annotators agreed the pair made sense and was not entailment.
- On MNLI dev, accuracy ranged from 0.61 to 0.75 (chance 0.33). On the NP/S set it ranged from 0.03 to 0.24, against a chance level of 0.50.
- After negation words were stripped from the premises, NP/S accuracy fell to 0.00–0.03. The earlier correct answers had mostly come from a negation cue.
Interesting ideas
- I1. Adversarial items are assembled only from words and dependencies the model saw in training, so failure cannot be blamed on unfamiliar vocabulary or domain.
- I2. The authors first write down the suspected shortcut as an explicit rule and then design items where that rule is always wrong. A near-zero score then directly exposes the shortcut.
- I3. When a model is right, they check why by deleting a suspected irrelevant cue (negation) and re-scoring. Correct labels earned for the wrong reason then vanish.
- I4. NP/S verbs make the string-level subsequence and the true constituent diverge. This is a precise probe of whether a system reasons over structure or over linear order.
For Mova
How these ideas could serve Mova, our small open language model — its files, stages and tests:
- I1 (learn): Build negative-control test items from lemmas and dependency patterns already in our treebanks, so that a failure points to reasoning rather than coverage.
- I2 (answer): For each reader, list the cheap strategies it might take, such as string overlap or keyword matching, and keep a test set where each strategy is guaranteed to fail.
- I3 (learn): Add cue-ablation runs to evaluation, for example removing negators or hedges, to confirm that correct answers still follow from the tree and frames.
- I4 (tree): Use NP/S verbs as a regression suite checking that the parser labels the clause as ccomp of believe/hear/claim rather than attaching the noun as obj.
Institutions
United States📚 Shelves 5The 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 worksComputer Vision and Pattern Recognition · 1,282 works
🧭 Topics 1Research subjects and languages the work is about.Reasoning · 510 works
Cite
Cite as R. Thomas McCoy and Tal Linzen. 2018. Non-entailed subsequences as a challenge for natural language inference. arXiv:1811.12112.
@misc{mccoy2018non,
title = {{Non-entailed subsequences as a challenge for natural language inference}},
author = {R. Thomas McCoy and Tal Linzen},
year = {2018},
doi = {10.48550/arxiv.1811.12112},
eprint = {1811.12112},
archivePrefix = {arXiv},
url = {https://doi.org/10.48550/arxiv.1811.12112}
}References
- 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, pages 632–642. Association for Computational Linguistics. ↗
- Cleo Condoravdi, Dick Crouch, Valeria de Paiva, Reinhard Stolle, and Daniel G. Bobrow. 2003. Entailment, intensionality and text understanding. In Proceedings of the HLT-NAACL 2003 Workshop on Text Meaning. find
- Ido Dagan, Oren Glickman, and Bernardo Magnini. 2006. The PASCAL Recognising Textual Entailment Challenge. In Joaquin Quiñonero-Candela, Ido Dagan, Bernardo Magnini, and Florence d’Alché Buc, editors, Machine Learning Challenges. Evaluating Predictive Uncertainty, Visual Object Classification, and Recognising Textual Entailment, pages 177–190. Springer Berlin Heidelberg, Berlin, Heidelberg. Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel Bowman, and Noah A. find
- 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
- Julia Hockenmaier and Mark Steedman. 2007. CCGbank: A Corpus of CCG Derivations and Dependency Structures Extracted from the Penn Treebank. Computational Linguistics, 33(3):355–396. ↗
- Seonhoon Kim, Jin-Hyuk Hong, Inho Kang, and Nojun Kwak. 2018. Semantic sentence matching with densely-connected recurrent and co-attentive information. arXiv preprint arXiv:1805.11360. ↗
- Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. 2014. Microsoft COCO: Common objects in context. In European Conference on Computer Vision, pages 740–755. Springer. ↗
- Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2016. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843. ↗
- 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
- Bradley L. Pritchett. 1988. Garden path phenomena and the grammatical basis of language processing. Language, 64(3):539–576. ↗
- Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. 2018. Improving language understanding by generative pre-training. find
- Alex Wang, Amapreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461. ↗
- 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. ↗
Updated
This page in Markdown → https://md.research.elamur.ai/free/en/papers/mccoy-2018-non-entailed-subsequences-challenge-natural/