A Re-evaluation of Knowledge Graph Completion Methods

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 505–514 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 505 GCAN: Graph-aware Co-Attention Networks for Explainable Fake News Detection on Social Media Yi-Ju Lu Department of Statistics National Cheng Kung University Tainan, Taiwan l852888@gmail.com Cheng-Te Li Institute of Data Science National Cheng Kung University Tainan, Taiwan chengte@mail.ncku.edu.tw Abstract This paper solves the fake news detection prob- lem under a more realistic scenario on so- cial media. Given the source short-text tweet and the corresponding sequence of retweet users without text comments, we aim at pre- dicting whether the source tweet is fake or not, and generating explanation by highlight- ing the evidences on suspicious retweeters and the words they concern. We develop a novel neural network-based model, Graph-aware Co- Attention Networks (GCAN), to achieve the goal. Extensive experiments conducted on real tweet datasets exhibit that GCAN can signifi- cantly outperform state-of-the-art methods by 16% in accuracy on average. In addition, the case studies also show that GCAN can produce reasonable explanations. 1 Introduction Social media is indispensable in people’s daily life, where users can express themselves, access news, and interact with each other. Information can fur- ther spread through the social network. Opinions and sentiments on source stories can be reflected by user participation and interaction. The conve- nient and low-cost essence of social networking brings collective intelligence, but at the same time leads to a negative by-product, the propagation of misinformation such as fake news. Fake news is a kind of news story possess- ing intentionally false information on social me- dia (Rashkin et al., 2017; Allcott and Gentzkow, 2017). The widespread of fake news can mislead the public, and produce unjust political, economic, or psychological profit for some parties (Horne and Adali, 2017; Allcott and Gentzkow, 2017). Data mining and machine learning techniques were uti- lized to detect fake news (Shu et al., 2017; Cha et al., 2020). Typical approaches rely on the con- tent of new articles to extract textual features, such as n-gram and bag of words, and apply supervised learning (e.g., random forest and support vector ma- chine) for binary classification (Shu et al., 2017). NLP researchers also learn advanced linguistic fea- tures, such as factive/assertive verbs and subjec- tivity (Popat, 2017) and writing styles and consis- tency (Potthast et al., 2018). Multi-modal context information is also investigated, such as user pro- files (Yang et al., 2012; Liu and Wu, 2018) and retweet propagation (Ruchansky et al., 2017; Shu et al., 2019a). Nevertheless, there are still critical challenges in detecting fake news online. First, existing content- based approaches (Castillo et al., 2011; Potthast et al., 2018; Shu et al., 2019a) require documents to be long text, e.g., news articles, so that the rep- resentation of words and sentences can be better learned. However, tweets on social media are usu- ally short text (Yan et al., 2015), which produces severe data sparsity problem. Second, some state- of-the-art models (Ruchansky et al., …

Content truncated to 500 words