{"title":"Effectively Modeling Sentence Interactions With Factorization Machines for Fact Verification","authors":"Zhen-Heng Chen, Fuzhen Zhuang, Lejian Liao, Meihuizi Jia, Jiaqi Li, Heyan Huang","doi":"10.1109/MIS.2023.3301170","DOIUrl":null,"url":null,"abstract":"Fact verification is a very challenging task that requires retrieving multiple evidence sentences from a reliable corpus to authenticate claims. Many claims require the simultaneous integration and reasoning of several pieces of evidence for verification. Existing models exhibit limitations in two aspects: 1) during the sentence selection stage, they only consider the interaction between the claim and the evidence, disregarding the intersentence information, and 2) most fusion strategies employed in current research, such as addition, concatenation, or simple neural networks, fail to capture the relationships and logical information among the evidence. To alleviate these problems, we propose select and fact verification modeling (SFVM). Our model utilizes a multihead self-attention mechanism combined with a gating mechanism to facilitate sentence interaction and enhance sentence embeddings. Then, we utilize factorization machines to effectively express the compressed alignment vectors, which are then used to expand the representations of the base evidence. To distinguish the importance of features, we use the evidence fusion network to determine the importance of various feature interactions. Results from experiments on the two public datasets showed that SFVM can utilize richer information between the claim and the evidence for fact verification and achieve competitive performance on the FEVER dataset.","PeriodicalId":13160,"journal":{"name":"IEEE Intelligent Systems","volume":null,"pages":null},"PeriodicalIF":5.6000,"publicationDate":"2023-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"IEEE Intelligent Systems","FirstCategoryId":"94","ListUrlMain":"https://doi.org/10.1109/MIS.2023.3301170","RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE","Score":null,"Total":0}
引用次数: 0
Abstract
Fact verification is a very challenging task that requires retrieving multiple evidence sentences from a reliable corpus to authenticate claims. Many claims require the simultaneous integration and reasoning of several pieces of evidence for verification. Existing models exhibit limitations in two aspects: 1) during the sentence selection stage, they only consider the interaction between the claim and the evidence, disregarding the intersentence information, and 2) most fusion strategies employed in current research, such as addition, concatenation, or simple neural networks, fail to capture the relationships and logical information among the evidence. To alleviate these problems, we propose select and fact verification modeling (SFVM). Our model utilizes a multihead self-attention mechanism combined with a gating mechanism to facilitate sentence interaction and enhance sentence embeddings. Then, we utilize factorization machines to effectively express the compressed alignment vectors, which are then used to expand the representations of the base evidence. To distinguish the importance of features, we use the evidence fusion network to determine the importance of various feature interactions. Results from experiments on the two public datasets showed that SFVM can utilize richer information between the claim and the evidence for fact verification and achieve competitive performance on the FEVER dataset.
期刊介绍:
IEEE Intelligent Systems serves users, managers, developers, researchers, and purchasers who are interested in intelligent systems and artificial intelligence, with particular emphasis on applications. Typically they are degreed professionals, with backgrounds in engineering, hard science, or business. The publication emphasizes current practice and experience, together with promising new ideas that are likely to be used in the near future. Sample topic areas for feature articles include knowledge-based systems, intelligent software agents, natural-language processing, technologies for knowledge management, machine learning, data mining, adaptive and intelligent robotics, knowledge-intensive processing on the Web, and social issues relevant to intelligent systems. Also encouraged are application features, covering practice at one or more companies or laboratories; full-length product stories (which require refereeing by at least three reviewers); tutorials; surveys; and case studies. Often issues are theme-based and collect articles around a contemporary topic under the auspices of a Guest Editor working with the EIC.