Fine-grained Fact Verification with Kernel Graph Attention Network

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7342–7351 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 7342 Fine-grained Fact Verification with Kernel Graph Attention Network Zhenghao Liu1 Chenyan Xiong2 Maosong Sun1 Zhiyuan Liu1 1Department of Computer Science and Technology, Tsinghua University, Beijing, China Institute for Artificial Intelligence, Tsinghua University, Beijing, China State Key Lab on Intelligent Technology and Systems, Tsinghua University, Beijing, China 2Microsoft Research AI, Redmond, USA Abstract Fact Verification requires fine-grained natural language inference capability that finds sub- tle clues to identify the syntactical and seman- tically correct but not well-supported claims. This paper presents Kernel Graph Attention Network (KGAT), which conducts more fine- grained fact verification with kernel-based at- tentions. Given a claim and a set of po- tential evidence sentences that form an evi- dence graph, KGAT introduces node kernels, which better measure the importance of the ev- idence node, and edge kernels, which conduct fine-grained evidence propagation in the graph, into Graph Attention Networks for more ac- curate fact verification. KGAT achieves a 70.38% FEVER score and significantly out- performs existing fact verification models on FEVER, a large-scale benchmark for fact ver- ification. Our analyses illustrate that, com- pared to dot-product attentions, the kernel- based attention concentrates more on relevant evidence sentences and meaningful clues in the evidence graph, which is the main source of KGAT’s effectiveness. All source codes of this work are available at https://github. com/thunlp/KernelGAT. 1 Introduction Online contents with false information, such as fake news, political deception, and online rumors, have been growing significantly and spread widely over the past several years. How to automatically “fact check” the integrity of textual contents, to pre- vent the spread of fake news, and to avoid the un- desired social influences of maliciously fabricated statements, is urgently needed for our society. Recent research formulates this problem as the fact verification task, which targets to automatically verify the integrity of statements using trustworthy corpora, e.g., Wikipedia (Thorne et al., 2018a). For example, as shown in Figure 1, a system could first Al Jardine is an American rhythm guitarist Claim Verification SUPPORTS REFUTES NOT ENOUGH INFO Evidence Reasoning Alan Charles Jardine (born September 3, 1942) is an American musician, singer and songwriter who co- founded the Beach Boys. He is best known as the band's rhythm guitarist, and for occasionally singing lead vocals on singles. Figure 1: An Example of Fact Verification System. retrieve related evidence sentences from the back- ground corpus, conduct joint reasoning over these sentences, and aggregate the signals to verify the claim integrity (Nie et al., 2019a; Zhou et al., 2019; Yoneda et al., 2018; Hanselowski et al., 2018). There are two challenges for evidence reasoning and aggregation in fact verification. One is that no ground truth evidence is given; the evidence sentences are retrieved from background corpora, which inevitably contain noise. The other is that the false claims are often deliberately fabricated; they may be semantically correct but are not supported. This makes …

Content truncated to 500 words