UnifEE: Unified Evidence Extraction for Fact Verification

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1150–1160 May 2-6, 2023 ©2023 Association for Computational Linguistics UNIFEE: Unified Evidence Extraction for Fact Verification Nan Hu♠∗, Zirui Wu♠∗, Yuxuan Lai♡♠, Chen Zhang♠, Yansong Feng♠♢† ♠Wangxuan Institute of Computer Technology, Peking University, China ♢The MOE Key Laboratory of Computational Linguistics, Peking University, China ♡Department of Computer Science, The Open University of China {hunan,ziruiwu,erutan,zhangch,fengyansong}@pku.edu.cn laiyx@ouchn.edu.cn Abstract FEVEROUS is a fact extraction and verification task that requires systems to extract evidence of both sentences and table cells from a Wikipedia dump, then predict the veracity of the given claim accordingly. Existing works extract evi- dence in the two formats separately, ignoring potential connections between them. In this pa- per, we propose a Unified Evidence Extraction model (UNIFEE), which uses a mixed evidence graph to extract the evidence in both formats. With the carefully-designed unified evidence graph, UNIFEE allows evidence interactions among all candidates in both formats at sim- ilar granularity. Experiments show that, with information aggregated from related evidence candidates in the fusion graph, UNIFEE can make better decisions about which evidence should be kept, especially for claims requiring multi-hop reasoning or a combination of tables and texts. Thus it outperforms all previous evi- dence extraction methods and brings significant improvement in the subsequent claim verifica- tion step. 1 Introduction FEVEROUS (Aly et al., 2021) is a fact extrac- tion and verification task over structured and un- structured information. Models should extract fine- grained evidence in two formats, namely, sentences and table cells, from a Wikipedia dump and verify the given claim accordingly. Many previous works focus on the claim veri- fication procedure. They propose various meth- ods to fuse evidence in different formats while leaving the evidence extraction within each for- mat separately (Kotonya et al., 2021; Malon, 2021; Bouziane et al., 2021; Hu et al., 2022). For those ef- forts on evidence extraction, Saeed et al. (2021) use a neural re-ranker to refine page retrieval. Bouziane ∗The first author and the second author contribute equally to this work. † Corresponding Author. Claim: John Laurie has collaborated with many directors such as Alfred Hitchcock, Michael Powell, and Laurence Olivier, but his most famous role is Private Frazer in a film written and directed by Alfred Hitchcock. [John Laurie] Evidence Cells Evidence Sentences [John Laurie] Laurie appeared in scores of feature films with directors such as Alfred Hitchcock, Michael Powell, and Laurence Olivier, generally playing bit-parts or supporting roles rather than leading roles. [Dad's Army (1971 film)] Directed by Norman Cohen, Dad's Army was filmed between series three and four and was based upon material from the early episodes of the television series. Year Film Role 1970 Step Laughing Into the Grave 1971 Dad's Army Private Frazer The Abominable Dr. Phibes Darrow 1976 Crime Casebook George Winterman / Sellens 1979 The Prisoner of Zenda Archbishop Figure 1: An example in FEVEROUS. The green rect- angle shows the gold evidence sentences, while the blue one is …

Content truncated to 500 words