Dual-Channel Evidence Fusion for Fact Verification over Texts and Tables

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5232 - 5242 July 10-15, 2022 ©2022 Association for Computational Linguistics Dual-Channel Evidence Fusion for Fact Verification over Texts and Tables Nan Hu♠, Zirui Wu♠, Yuxuan Lai♡♠, Xiao Liu♠, 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,lxlisa,fengyansong}@pku.edu.cn laiyx@ouchn.edu.cn Abstract Different from previous fact extraction and ver- ification tasks that only consider evidence of a single format, FEVEROUS brings further chal- lenges by extending the evidence format to both plain text and tables. Existing works convert all candidate evidence into either sentences or tables, thus often failing to fully capture the rich context in their original format from the converted evidence, let alone the context in- formation lost during conversion. In this pa- per, we propose a Dual Channel Unified For- mat fact verification model (DCUF), which uni- fies various evidence into parallel streams, i.e., natural language sentences and a global evi- dence table, simultaneously. With carefully- designed evidence conversion and organization methods, DCUF makes the most of pre-trained table/language models to encourage each ev- idence piece to perform early and thorough interactions with other pieces in its original format. Experiments show that our model can make better use of existing pre-trained models to absorb evidence of two formats, thus outper- forming previous works by a large margin. Our code and models are publicly available1. 1 Introduction The task of fact extraction and verification aims to extract evidence and verify a given claim. Previ- ous efforts focus on dealing with text format ev- idence from unstructured documents (Nie et al., 2019; Zhong et al., 2020; Kruengkrai et al., 2021) or evidence from a single given table (Chen et al., 2020; Yang et al., 2020; Eisenschlos et al., 2020). Recently, Aly et al. (2021) propose a new realis- tic setting, FEVEROUS, i.e., fact extraction and verification over unstructured and structured infor- mation. In FEVEROUS, models should not only extract evidence sentences/table cells from millions ∗Corresponding Author. 1https://github.com/lanlanabcd/dual_ channel_feverous Actor Award Result Cartwright Primetime Emmy Won Roswell Annie Nominated Sentence Evidence Table/Cell Evidence The Simpsons is …… stars. SUPPORTS? Both Fox and …… artists. Castellaneta …… won one. Roswell won an Annie Award for her performance in The Simpsons. Claim Figure 1: An excerpt example from FEVEROUS. of passages, but also combine the evidence in dif- ferent formats to verify a given claim. Previous works on FEVEROUS generally con- vert all evidence pieces into either plain text (Aly et al., 2021; Saeed et al., 2021; Malon, 2021) or several tables (Bouziane et al., 2021). However, format conversions inevitably lose rich context in- formation for the converted evidence, thus may mislead the subsequent encoding and interaction steps. For example, in Figure 1, the entire top two rows are indispensable to understand the table cell Won. It is difficult to identify all related …

Content truncated to 500 words