Elamur Research⌕
☰
home / papers / 2021 / Do Natural Language Explanations Represent Valid Logical Arguments? Verifying Entailment in Explainable NLI Gold Standards

Do Natural Language Explanations Represent Valid Logical Arguments? Verifying Entailment in Explainable NLI Gold Standards

Paper · 2021 · Marco Valentino, Ian Pratt-Hartmann and André Freitas

Summary

Datasets that pair NLI items with human-written explanations are often treated as ground truth for training and judging models that reason step by step. This paper tests whether those explanations actually work as logical arguments, meaning whether the gold answer really follows from them. The authors propose Explanation Entailment Verification (EEV), a manual annotation protocol with three stages:

  1. Problem definition. Each item, either Textual Entailment or multiple-choice QA, is recast as a set of premises and one conclusion.
  2. Formalisation. Trained annotators translate the sentences into event-based logical forms.
  3. Verification. Annotators assign one of five mutually exclusive validity labels.

They apply EEV to samples from e-SNLI, Worldtree and QASC. Most explanations turn out not to entail the answer, even though they read as coherent. They fail mainly because a needed premise is missing, because no argument can be discerned at all, or because they contain outright logical errors. The finding matters because models are being supervised and scored against explanations whose inferential quality was never checked. It also suggests that annotators' intuitions about what makes an argument "complete" differ from formal validity.

Key points

  • The five EEV labels. An explanation can be judged valid with no superfluous premises, valid yet carrying unneeded premises, lacking a reasonable premise the argument requires, flawed in its reasoning, or offering no identifiable argument whatsoever.

  • How tasks are recast. For TE, only items labelled entailment are used, and the original premise becomes part of the premise set. For MCQA, the conclusion is built by inserting the correct answer into the question.

  • Formalisation. It uses event-based semantics: verbs become event types and their arguments become semantic roles. The aim is to reduce ambiguity and raise inter-annotator agreement.

  • Overall label distribution:

    LabelShare
    Valid and non-redundant20.42%
    Valid, redundant premises21.91%
    Missing plausible premise26.00%
    Logical error11.19%
    No discernible argument20.47%

  • Per-dataset results.

    • QASC is the weakest: 62.74% of its explanations are invalid, including 35.29% with no discernible argument and 17.65% with logical errors.
    • e-SNLI has the most valid arguments, at 31.37%.
    • Worldtree's quality is linked to its non-crowd-sourced authoring.
    • Explanations judged non-contrastive (they fail to rule out other plausible options) make up 26.53% for Worldtree and 49.02% for QASC.

Interesting ideas

  • I1. Recasting different task formats (TE pairs, multiple-choice questions) into one uniform "premises entail conclusion" problem lets a single checker judge all of them, independent of the original format.
  • I2. A small, mutually exclusive taxonomy of argument failures separates "redundant but valid" from "missing premise" and from "real logical error". This turns a vague quality judgement into countable diagnostics.
  • I3. Formalising explanations in event-based semantics, with verbs as event types and arguments as roles, before judging entailment makes hidden gaps visible that fluent surface text conceals.
  • I4. A large share of human explanations leave out "obvious" world knowledge, so the missing-premise rate measures how much background knowledge a task silently assumes.
  • I5. Checking whether an explanation is contrastive tests a further property beyond support for the gold answer. It asks whether the explanation also rules out the competing options, which exposes questions with several plausible answers.

For Mova

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

  • I1 (answer): School-test, word-problem and fable-QA readers could all emit their justification as premises plus a conclusion, so one shared validity checker audits every answer type.
  • I2 (learn): Evaluation reports could tag each failed derivation with these five classes, showing whether errors come from knowledge gaps or from broken inference.
  • I3 (frames): Once frames are built from VerbNet roles, an explanation's sentences can be compared role by role against the conclusion frame to detect which participant or event is unsupported.
  • I4 (world): Missing-premise cases found when checking gold explanations can serve as a to-do list of facts to add to the compiled ontology and knowledge graph.
  • I5 (answer): Multiple-choice readers could be required to show why each distractor fails, and questions where no such reason exists could be flagged as ambiguous rather than counted as errors.

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 ManchesterManchester · United Kingdom30 works · 43,942 citations

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 9,085 worksArtificial Intelligence · 8,888 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 Marco Valentino, Ian Pratt-Hartmann and André Freitas. 2021. Do Natural Language Explanations Represent Valid Logical Arguments? Verifying Entailment in Explainable NLI Gold Standards. Proceedings of the 14th International Conference on Computational Semantics (IWCS).

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{valentino2021do,
  title = {{Do Natural Language Explanations Represent Valid Logical Arguments? Verifying Entailment in Explainable NLI Gold Standards}},
  author = {Marco Valentino, Ian Pratt-Hartmann and André Freitas},
  booktitle = {Proceedings of the 14th International Conference on Computational Semantics (IWCS)},
  year = {2021},
  url = {https://aclanthology.org/2021.iwcs-1.8/}
}

References

31 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. pers), pages 3543–3556, Minneapolis, Minnesota. Association for Computational Linguistics. Chandra Bhagavatula, Ronan Le Bras, Chaitanya Malaviya, Keisuke Sakaguchi, Ari Holtzman, Hannah Rashkin, Doug Downey, Wen tau Yih, and Yejin Choi. 2020. Abductive commonsense reasoning. In International Conference on Learning Representations.
  2. Peter Jansen, Elizabeth Wainwright, Steven Marmorstein, and Clayton Morrison. 2018. WorldTree: A corpus of explanation graphs for elementary science questions supporting multi-hop inference. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA). ↗
  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, pages 632–642, Lisbon, Portugal. Association for Computational Linguistics. ↗
  4. Harsh Jhamtani and Peter Clark. 2020. Learning to explain: Datasets and models for identifying valid reasoning chains in multihop question-answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 137–150, Online. Association for Computational Linguistics. find
  5. Samuel R. Bowman and George E. Dahl. 2021. What will it take to fix benchmarking in natural language understanding? find
  6. Oana-Maria Camburu, Tim Rocktäschel, Thomas Lukasiewicz, and Phil Blunsom. 2018. e-snli: Natural language inference with natural language explanations. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems 31, pages 9539–9549. Curran Associates, Inc. ↗
  7. Tushar Khot, Peter Clark, Michal Guerquin, Peter Jansen, and Ashish Sabharwal. 2020. Qasc: A dataset for question answering via sentence composition. Proceedings of the AAAI Conference on Artificial Intelligence, 34(05):8082–8090. ↗
  8. Sawan Kumar and Partha Talukdar. 2020. NILE : Natural language inference with faithful natural language explanations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8730–8742, Online. Association for Computational Linguistics.
  9. Oana-Maria Camburu, Brendan Shillingford, Pasquale Minervini, Thomas Lukasiewicz, and Phil Blunsom. 2020. Make up your mind! adversarial generation of inconsistent natural language explanations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4157– 4165, Online. Association for Computational Linguistics. find
  10. Patrick Lewis, Pontus Stenetorp, and Sebastian Riedel. 2021. Question and answer test-train overlap in open-domain question answering datasets. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1000–1008, Online. Association for Computational Linguistics. find
  11. Ruben Cartuyvels, Graham Spinks, and Marie-Francine Moens. 2020. Autoregressive reasoning over chains of facts with transformers. In Proceedings of the 28th International Conference on Computational Linguistics, pages 6916–6930, Barcelona, Spain (Online). International Committee on Computational Linguistics. ↗
  12. Peter Lipton. 1990. Contrastive explanation. Royal Institute of Philosophy Supplement, 27:247–266. find
  13. Tim Miller. 2019. Explanation in artificial intelligence: Insights from the social sciences. Artificial Intelligence, 267:1–38. ↗
  14. Deborah Ferreira and André Freitas. 2020a. Natural language premise selection: Finding supporting statements for mathematical text. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 2175–2182, Marseille, France. European Language Resources Association. ↗
  15. Sewon Min, Eric Wallace, Sameer Singh, Matt Gardner, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2019. Compositional questions do not necessitate multi-hop reasoning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4249–4257, Florence, Italy. Association for Computational Linguistics.
  16. Deborah Ferreira and André Freitas. 2020b. Premise selection in natural language mathematical texts. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7365– 7374, Online. Association for Computational Linguistics. find
  17. Ellie Pavlick and Tom Kwiatkowski. 2019. Inherent Disagreements in Human Textual Inferences. Transactions of the Association for Computational Linguistics, 7:677–694. ↗
  18. Sarthak Jain and Byron C. Wallace. 2019. Attention is not Explanation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Pa-Nazneen Fatema Rajani, Bryan McCann, Caiming Xiong, and Richard Socher. 2019. Explain yourself! ↗
  19. Sarah Wiegreffe and Yuval Pinter. 2019. Attention is not not explanation. 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 11–20, Hong Kong, China. Association for Computational Linguistics. leveraging language models for commonsense reasoning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4932–4942, Florence, Italy. Association for Computational Linguistics.
  20. Marco Tulio Ribeiro, Tongshuang Wu, Carlos Guestrin, and Sameer Singh. 2020. Beyond accuracy: Behavioral testing of NLP models with CheckList. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4902– 4912, Online. Association for Computational Linguistics. ↗
  21. Zhengnan Xie, Sebastian Thiem, Jaycie Martin, Elizabeth Wainwright, Steven Marmorstein, and Peter Jansen. 2020. WorldTree v2: A corpus of science-domain structured explanations and inference patterns supporting multi-hop inference. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 5456–5473, Marseille, France. European Language Resources Association. find
  22. Viktor Schlegel, Marco Valentino, Andre Freitas, Goran Nenadic, and Riza Batista-Navarro. 2020. A framework for evaluation of machine reading comprehension gold standards. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 5359–5369, Marseille, France. European Language Resources Association. ↗
  23. Dongfang Xu, Peter Jansen, Jaycie Martin, Zhengnan Xie, Vikas Yadav, Harish Tayyar Madabushi, Oyvind Tafjord, and Peter Clark. 2020. Multi-class hierarchical question classification for multiple choice science exams. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 5370–5382, Marseille, France. European Language Resources Association. find
  24. Sanjay Subramanian, Ben Bogin, Nitish Gupta, Tomer Wolfson, Sameer Singh, Jonathan Berant, and Matt Gardner. 2020. Obtaining faithful interpretations from compositional neural networks. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5594–5608, Online. Association for Computational Linguistics. find
  25. Mokanarangan Thayaparan, Marco Valentino, and André Freitas. 2020a. Explanationlp: Abductive reasoning for explainable science question answering. ↗
  26. Mokanarangan Thayaparan, Marco Valentino, and André Freitas. 2020b. A survey on explainability in machine reading comprehension. ↗
  27. Marco Valentino, Mokanarangan Thayaparan, and André Freitas. 2021. Unification-based reconstruction of multi-hop explanations for science questions. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 200–211, Online. Association for Computational Linguistics. find
  28. Marco Valentino, Mokanarangan Thayaparan, and André Freitas. 2020. Explainable natural language reasoning via conceptual unification. find
  29. Douglas Walton. 2004. A new dialectical theory of explanation. Philosophical Explorations, 7(1):71–89. ↗
  30. Cunxiang Wang, Shuailong Liang, Yili Jin, Yilong Wang, Xiaodan Zhu, and Yue Zhang. 2020. SemEval-2020 task 4: Commonsense validation and explanation. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pages 307–321, Barcelona (online). International Committee for Computational Linguistics.
  31. Sarah Wiegreffe and Ana Marasović. 2021. Teach me to explain: A review of datasets for explainable nlp. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/valentino-2021-do-natural-language-explanations-represent/