Towards Debiasing Fact Verification Models

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 3419–3425, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 3419 Towards Debiasing Fact Verification Models Tal Schuster*,1, Darsh J Shah*,1, Yun Jie Serene Yeo2, Daniel Filizzola1, Enrico Santus1, Regina Barzilay1 1Computer Science and Artificial Intelligence Lab, MIT 2DSO National Laboratories, Singapore {tals, darsh, yeoyjs, danifili, esantus, regina}@csail.mit.edu Abstract Fact verification requires validating a claim in the context of evidence. We show, however, that in the popular FEVER dataset this might not necessarily be the case. Claim-only classi- fiers perform competitively with top evidence- aware models. In this paper, we investigate the cause of this phenomenon, identifying strong cues for predicting labels solely based on the claim, without considering any evidence. We create an evaluation set that avoids those id- iosyncrasies. The performance of FEVER- trained models significantly drops when evalu- ated on this test set. Therefore, we introduce a regularization method which alleviates the ef- fect of bias in the training data, obtaining im- provements on the newly created test set. This work is a step towards a more sound evalua- tion of reasoning capabilities in fact verifica- tion models.1 1 Introduction Creating quality datasets is essential for expand- ing NLP functionalities to new tasks. Today, such datasets are often constructed using crowdsourc- ing mechanisms. Prior research has demonstrated that artifacts of this data collection method often introduce idiosyncratic biases that impact perfor- mance in unexpected ways (Poliak et al., 2018; Gururangan et al., 2018). In this paper, we explore this issue using the FEVER dataset, designed for fact verification (Thorne et al., 2018). The task of fact verification involves assessing claim validity in the context of evidence, which can either support, refute or contain not enough information. Figure 1(A) shows an example of a FEVER claim and evidence. While validity of some claims may be asserted in isolation (e.g. Asterisk (*) denotes equal contribution. 1Data and code: https://github.com/ TalSchuster/FeverSymmetric Claim Evidence Refutes New Claim New Evidence Refutes Supports (A) ORIGINAL pair from the FEVER dataset Claim: Stanley Williams stayed in Cuba his whole life. Evidence: Stanley [...] was part of the West Side Crips, a street gang which has its roots in South Central Los Angeles. (B) Manually GENERATED pair Claim: Stanley Williams moved from Cuba to California when he was 15 years old. Evidence: Stanley [...] was born in Havana and didn’t leave the country until he died. Figure 1: An illustration of a REFUTES claim-evidence pair from the FEVER dataset (A) that is used to gener- ate a new pair (B). From the combination of the ORIGI- NAL and manually GENERATED pairs, we obtain a total of four pairs creating symmetry. through common sense knowledge), contextual verification is key for a fact-checking task (Al- hindi et al., 2018). Datasets should ideally eval- uate this ability. To assess whether this is the case for FEVER, we train a claim-only BERT (Devlin et …

Content truncated to 500 words