Elamur Research
☰
home / papers / 2021 / Explaining Answers with Entailment Trees

Explaining Answers with Entailment Trees

Written for the Research library (claude-opus-5-5) from the paper text. Original — arxiv.org/pdf/2104.08661

Summary

Most question-answering systems justify an answer by pointing to a supporting text snippet. They rarely show how known facts combine, step by step, to produce the conclusion. This paper frames explanation as a multistep entailment tree. The leaves are known facts, the intermediate nodes are derived conclusions, and the root is a hypothesis that restates the question together with its answer. The authors built ENTAILMENTBANK, the first dataset of such trees. Expert annotators used a custom drag-and-drop authoring tool to write the trees for ARC grade-school science questions, drawing on a WorldTree V2 knowledge corpus. They defined three tasks of rising difficulty, where the tree must be generated from (a) only the gold leaf facts, (b) the leaves mixed with distractors, or (c) the full corpus. As baselines, they trained generative T5-style models called EntailmentWriters, adapted from ProofWriter. These models solve the tasks partially and do best when the needed facts are supplied. The resource matters because it allows explanations to be judged on the correctness of their reasoning, and it opens the way to interactive debugging of model answers.

Key points

  • ENTAILMENTBANK contains 1,840 entailment trees, averaging about 6.6 nodes and 2.7 entailment steps per tree. The data, annotation tool and evaluation code are released at allenai.org/data/entailmentbank.
  • When given all relevant sentences (task a), the best EntailmentWriter produced a fully error-free tree in roughly 35% of cases. Scores fall sharply as the gold tree grows in steps.
  • Each inference step is fine-grained and draws on several premises. Compared with WorldTree V2, eQASC, HotpotQA, R4C and StrategyQA, it is the only dataset with explicitly ordered entailment trees.
  • In an out-of-domain test on eQASC with no fine-tuning, the model reached 67% leaf F1 and 26% Overall AllCorrect.
  • A version that receives no context fact set (it writes all tree sentences itself) produced valid, non-vacuous trees for about 35% of a manually checked sample of Challenge300. A one-step "shredded" variant supports interactive, drill-down explanation.

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: AI & machine learning

Cited by

121 papers of our library cite it (of the 140673 whose bibliographies we read); 5 of them have a page here, 116 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2104.08661 · arXiv 2104.08661

@misc{dalvi2021explaining,
  title = {{Explaining Answers with Entailment Trees}},
  author = {Dalvi, Bhavana and Jansen, Peter and Tafjord, Oyvind and Xie, Zhengnan and Smith, Hannah and Pipatanangkura, Leighanna and Clark, Peter},
  year = {2021},
  doi = {10.48550/arxiv.2104.08661},
  eprint = {2104.08661},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2104.08661}
}

References

34 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Roy Bar-Haim, I. Dagan, and Idan Szpektor. 2014. Benchmarking applied semantic inference: The pascal recognising textual entailment challenges. In Language, Culture, Computation. find
  2. Peter A. Jansen, Elizabeth Wainwright, Steven Marmorstein, and Clayton T. Morrison. 2018. WorldTree: A corpus of explanation graphs for elementary science questions supporting multi-hop inference. In LREC. find
  3. L. Bentivogli, Peter Clark, I. Dagan, and Danilo Giampiccolo. 2011. The seventh pascal recognizing textual entailment challenge. Theory and Applications of Categories. find
  4. Harsh Jhamtani and P. Clark. 2020. Learning to explain: Datasets and models for identifying valid reasoning chains in multihop question-answering. In EMNLP. find
  5. Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In EMNLP. find
  6. Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. 2020. UnifiedQA: Crossing format boundaries with a single QA system. In Findings-EMNLP. find our card
  7. Christos Christodoulopoulos, James Thorne, Andreas Vlachos, Oana Cocarascu, and Arpit Mittal, editors. 2020. Proc. 3rd Workshop on Fact Extraction and Verification. ACL. Https://aclanthology.org/events/fever-2020/. find
  8. Alice Lai, Yonatan Bisk, and J. Hockenmaier. 2017. Natural language inference from multiple premises. In IJCNLP. find
  9. Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. Think you have solved question answering? Try ARC, the AI2 reasoning challenge. ArXiv, abs/1803.05457. find
  10. Ido Dagan, Dan Roth, Mark Sammons, and Fabio Massimo Zanzotto. 2013. Recognizing Textual Entailment: Models and Applications. Morgan and Claypool. find
  11. Dorottya Demszky, Kelvin Guu, and Percy Liang. 2018. Transforming question answering datasets into natural language inference datasets. ArXiv, abs/1809.02922. find
  12. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HIT. find
  13. Jay DeYoung, Sarthak Jain, Nazneen Rajani, E. Lehman, Caiming Xiong, R. Socher, and Byron C. Wallace. 2019. ERASER: A benchmark to evaluate rationalized NLP models. In ACL. find
  14. Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, D. Roth, and Jonathan Berant. 2021. Did Aristotle use a laptop? A question answering benchmark with implicit reasoning strategies. ArXiv, abs/2101.02235. find
  15. Shuguang Han, Xuanhui Wang, Mike Bendersky, and Marc Najork. 2020. Learning-to-rank with BERT in TF-ranking. ArXiv, abs/2004.08476. — [arXiv:2004.08476](https://arxiv.org/abs/2004.08476) find
  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, abs/1907.11692. find
  17. T. Miller. 2019. Explanation in artificial intelligence: Insights from the social sciences. Artif. Intell., 267:1–38. find our card
  18. Ethan Perez, Siddharth Karamcheti, Rob Fergus, Jason Weston, Douwe Kiela, and Kyunghyun Cho. 2019. Finding generalizable evidence by learning to convince Q&A models. In EMNLP. find
  19. Stanislas Polu and Ilya Sutskever. 2020. Generative language modeling for automated theorem proving. ArXiv, abs/2009.03393. find our card
  20. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, M. Matena, Yanqi Zhou, W. Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1–140:67. find
  21. Nazneen Rajani, B. McCann, Caiming Xiong, and R. Socher. 2019. Explain yourself! Leveraging language models for commonsense reasoning. In ACL. find our card
  22. Swarnadeep Saha, Sayan Ghosh, Shashank Srivastava, and Mohit Bansal. 2020. PRover: Proof generation for interpretable reasoning over rules. In EMNLP. find
  23. Thibault Sellam, Dipanjan Das, and Ankur P Parikh. 2020. Bleurt: Learning robust metrics for text generation. ArXiv, abs/2004.04696. find
  24. Braden Hancock, Paroma Varma, Stephanie Wang, Martin Bringmann, Percy Liang, and Christopher Ré. 2018. Training classifiers with natural language explanations. In ACL. find
  25. Minjoon Seo, Aniruddha Kembhavi, Ali Farhadi, and Hannaneh Hajishirzi. 2016. Bidirectional attention flow for machine comprehension. In ICLR. find our card
  26. N. Inoue, Pontus Stenetorp, and Kentaro Inui. 2020. R4C: A benchmark for evaluating RC systems to get the right answer for the right reason. In ACL. find
  27. Oyvind Tafjord and Peter Clark. 2021. General-purpose question-answering with Macaw. ArXiv, abs/2109.02593. find
  28. Oyvind Tafjord, B. D. Mishra, and P. Clark. 2021. ProofWriter: Generating implications, proofs, and abductive statements over natural language. IJCAI. find our card
  29. James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. Fever: a large-scale dataset for fact extraction and verification. In NAACL. find our card
  30. Ming-Zhe Wang and Jun Deng. 2020. Learning to prove theorems by learning to generate theorems. In NeurIPS. find
  31. 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 LREC. find
  32. Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhutdinov, and Christopher D Manning. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. ArXiv, abs/1809.09600. find our card
  33. Qinyuan Ye, Xiaozhen Huang, and Xiang Ren. 2020. Teaching machine comprehension with compositional explanations. ArXiv, abs/2005.00806. find
  34. A Question: A student left a bar of chocolate in the sun on a hot day. As the chocolate melted, what property changed? Answer: its shape Relevant Fact Retrieval Algorithm find