GLAF: Global-to-Local Aggregation and Fission Network for Semantic Level Fact Verification

Proceedings of the 29th International Conference on Computational Linguistics, pages 1801–1812 October 12–17, 2022. 1801 GLAF: Global-to-Local Aggregation and Fission Network for Semantic-Level Fact Verification Zhiyuan Ma1, Jianjun Li1∗, Guohui Li1, Yongjing Cheng2 1 Huazhong University of Science and Technology (HUST), China 2 National University of Defense Technology (NUDT), China {zhiyuanma,jianjunli,guohuili}@hust.edu.cn davidcheng1001@163.com Abstract Accurate fact verification depends on perform- ing fine-grained reasoning over crucial entities by capturing their latent logical relations hid- den in multiple evidence clues, which is gen- erally lacking in existing fact verification mod- els. In this work, we propose a novel Global- to-Local Aggregation and Fission network (GLAF) to fill this gap. Instead of treating en- tire sentences or all semantic elements within them as nodes to construct a coarse-grained or unstructured evidence graph as in previ- ous methods, GLAF constructs a fine-grained and structured evidence graph by parsing the rambling sentences into structural triple-level reasoning clues and regarding them as graph nodes to achieve fine-grained and interpretable evidence graph reasoning. Specifically, to cap- ture latent logical relations between the clues, GLAF first employs a local fission reasoning layer to conduct fine-grained multi-hop rea- soning, and then uses a global evidence ag- gregation layer to achieve information sharing and the interchange of evidence clues for fi- nal claim label prediction. Experimental re- sults on the FEVER dataset demonstrate the ef- fectiveness of GLAF, showing that it achieves the state-of-the-art performance by obtaining a 77.62% FEVER score. 1 Introduction The classic fact verification (FV) task is defined as retrieving relevant sentences as evidence and conducting joint reasoning over these evidence sen- tences to verify the correctness of a claim, and finally returning a result such as “SUPPORTS”, “REFUTES”, or “NOT ENOUGH INFO”. With the increasingly frequent internet fraud, political ru- mors, fake news and other false information on- line, fact verification is becoming more and more important. How to automatically verify the fake claims and prevent their spread is a vital problem. ∗Corresponding author. Claim C1: The Great Wall of China stretches from Lop Lake to Shenyang. Evidence #1: The Hushan Great Wall, the far eastern end of the Great Wall of China, is located Dandong. Evidence #2: Lop Lake, located in Xinjiang Province, is the starting side of the Great Wall of China at the western end. Evidence Clues Joint Reasoning: (the Great Wall of China, eastern end of, Dandong) (the Great Wall of China, western end of, Lop Lake) REFUTES Claim C2: The Rodney King riots took place in the most populous county in the USA. Evidence #1: Evidence #2: The 1992 Los Angeles riots, also known as the Rodney King riots were a series of riots, lootings, arsons, and civil disturbances that occurred in Los Angeles County, California in April and May 1992. Los Angeles County, officially the County of Los Angeles, is the most populous county in the USA. Evidence Clues 2-hop Reasoning: (the Rodney King riots,occurred in,Los Angeles County) (Los Angeles County,is, most populous county in the USA) SUPPORTS aggregate hop Figure …

Content truncated to 500 words