SISER: Semantic-Infused Selective Graph Reasoning for Fact Verification

Proceedings of the 29th International Conference on Computational Linguistics, pages 1367–1378 October 12–17, 2022. 1367 SISER: Semantic-Infused Selective Graph Reasoning for Fact Verification Eunhwan Park1♠, Jong-Hyeon Lee2♠∗, Donghyeon Jeon3, Seonhoon Kim3, Inho Kang3, Seung-Hoon Na1† 1Jeonbuk National University, 2NCSOFT, 3NAVER Corporation {judepark, nash}@jbnu.ac.kr, leejh1230@ncsoft.com {donghyeon.jeon, seonhoon.kim, once.ihkang}@navercorp.com Abstract This study proposes Semantic-Infused SElective Graph Reasoning (SISER) for fact verification, which newly presents semantic-level graph reasoning and injects its reasoning-enhanced representation into other types of graph-based and sequence-based reasoning methods. SISER combines three reasoning types: 1) semantic-level graph reasoning, which uses a semantic graph from evidence sentences, whose nodes are elements of a triple – <Subject, Verb, Object>, 2) “semantic-infused” sentence-level “selective” graph reasoning, which combine semantic- level and sentence-level representations and perform graph reasoning in a selective manner using the node selection mechanism, and 3) sequence reasoning, which concatenates all evidence sentences and performs attention- based reasoning. Experiment results on a large-scale dataset for Fact Extraction and VERification (FEVER) show that SISER out- performs the previous graph-based approaches and achieves state-of-the-art performance. 1 Introduction An ever-increasing number of unconfirmed false or misleading information spread on various so- cial media platforms has motivated the verifica- tion of textual information, referred to as fact verification. FEVER (Thorne et al., 2018a) pre- sented a large dataset for fact verification, initiat- ing a shared task that aims to automatically clas- sify a human-generated claim into ‘Supported’, ‘Refuted’, or ‘Not Enough Info’ based on retrieved evidence sentences from Wikipedia1. Claim verification, the final step of fact verifica- tion, is viewed as a task of natural language infer- ence (NLI) (Angeli and Manning, 2014). Specif- †Corresponding author ♠Equal contribution ∗This work was done during his M.S. degree at JBNU. 1https://competitions.codalab.org/ competitions/18814 ically, the NLI task for claim verification is for- mulated as the set-to-sentence entailment of infer- ring whether a claim (as the hypothesis) is logi- cally “entailed” from a set of retrieved evidence sentences (as the premise). Recently, graph reasoning for claim verifica- tion has been extensively explored (Zhou et al., 2019; Liu et al., 2020; Zhong et al., 2020), which creates a graph whose nodes are semantic units extracted from a set of evidence sentences or a claim, and applies graph neural networks (GNNs) such as (Veliˇckovi´c et al., 2018; Kipf and Welling, 2017) to infer the entailment relationship. How- ever, graph reasoning may be somehow restricted to unit-biased reasoning, when relying on a single type of semantic unit for nodes of a graph, such as sentences, entities, or words, meaning that the semantic interaction between claim and evidence is restricted to a single graph type and does not go beyond the coverage of the “given” semantic units. In addition, graph reasoning may suffer from over- smoothing inherited from GNNs (Gasteiger et al., 2019; Zhao and Akoglu, 2020; Chen et al., 2020a; Rong et al., 2020), likely causing all node repre- sentations to converge to a stationary point at the extreme, as reported by (Li et al., 2018). To address these limitations of graph …

Content truncated to 500 words