IMCI: Integrate Multi-view Contextual Information for Fact Extraction and Verification

Proceedings of the 29th International Conference on Computational Linguistics, pages 1412–1421 October 12–17, 2022. 1412 IMCI: Integrate Multi-view Contextual Information for Fact Extraction and Verification Hao Wang†, Yangguang Li‡, Zhen Huang†∗, Yong Dou† † National University of Defense Technology, Changsha ‡ SenseTime, Beijing {hao.wang, huangzhen, yongdou}@nudt.edu.cn liyangguang@sensetime.com Abstract With the rapid development of automatic fake news detection technology, fact extraction and verification (FEVER) has been attracting more attention. The task aims to extract the most related fact evidences from millions of open- domain Wikipedia documents and then verify the credibility of corresponding claims. Al- though several strong models have been pro- posed for the task and they have made great progress, we argue that they fail to utilize multi-view contextual information and thus cannot obtain better performance. In this pa- per, we propose to integrate multi-view con- textual information (IMCI) for fact extraction and verification. For each evidence sentence, we define two kinds of context, i.e. intra- document context and inter-document con- text. Intra-document context consists of the document title and all the other sentences from the same document. Inter-document context consists of all other evidences which may come from different documents. Then we integrate the multi-view contextual infor- mation to encode the evidence sentences to handle the task. Our experimental results on FEVER 1.0 shared task show that our IMCI framework makes great progress on both fact extraction and verification, and achieves state-of-the-art performance with a winning FEVER score of 73.96% and label accuracy of 77.25% on the online blind test set. We also conduct ablation study to detect the impact of multi-view contextual information. Our codes will be released at https://github.com/ phoenixsecularbird/IMCI. 1 Introduction Fake news propagation is a severe social problem, which may cause great loss and lead to serious con- sequence, e.g. panic, quarrel, opposition and even war. The situation has become a general concern since Brexit and the U.S. President Campaign in ∗*Corresponding author 2016 and gets far more intense due to COVID-19 pandemic (Martino et al., 2020). In this condition, automatic fake news detection has been developing rapidly. According to Ruffo et al. (2021), auto- matic fake news detection mainly include textual- content based methods (Giachanou et al., 2019; Ghanem et al., 2020; Kaliyar et al., 2021), user- role based methods (Vo and Lee, 2019; Giachanou et al., 2020), multi-modal approaches (Zlatkova et al., 2019; Fung et al., 2021) and detection of bots and trolls (Stella et al., 2018; Sayyadiharikan- deh et al., 2020). Among textual-content based methods, fact extraction and verification (FEVER) (Thorne et al., 2018) has been attracting developing attention. As shown in Figure 1, for a given claim, the task aims to select at most 5 most related sen- tences as evidences from millions of open-domain Wikipedia documents for fact extraction, and com- bine the selected evidences to judge the claim as SUPPORTS, REFUTES or NOT ENOUGH INFO (NEI) for fact verification. Figure 1: An example from FEVER 1.0 shared task. (Underlined sentence is the unlabeled intra- document context. Words in …

Content truncated to 500 words