Learning Hierarchical Discourse-level Structure for Fake News Detection

Proceedings of NAACL-HLT 2019, pages 3432–3442 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 3432 Learning Hierarchical Discourse-level Structure for Fake News Detection Hamid Karimi Computer Science and Engineering Michigan State University karimiha@msu.edu Jiliang Tang Computer Science and Engineering Michigan State University tangjili@msu.edu Abstract On the one hand, nowadays, fake news arti- cles are easily propagated through various on- line media platforms and have become a grand threat to the trustworthiness of information. On the other hand, our understanding of the language of fake news is still minimal. In- corporating hierarchical discourse-level struc- ture of fake and real news articles is one cru- cial step toward a better understanding of how these articles are structured. Nevertheless, this has rarely been investigated in the fake news detection domain and faces tremendous chal- lenges. First, existing methods for captur- ing discourse-level structure rely on annotated corpora which are not available for fake news datasets. Second, how to extract out useful information from such discovered structures is another challenge. To address these chal- lenges, we propose Hierarchical Discourse- level Structure for Fake news detection. HDSF learns and constructs a discourse-level struc- ture for fake/real news articles in an auto- mated and data-driven manner. Moreover, we identify insightful structure-related proper- ties, which can explain the discovered struc- tures and boost our understating of fake news. Conducted experiments show the effective- ness of the proposed approach. Further struc- tural analysis suggests that real and fake news present substantial differences in the hierarchi- cal discourse-level structures. 1 Introduction In this work, we focus on detecting fake news ar- ticles (hereafter referred to as documents) based on their contents. Many existing linguistic ap- proaches for fake news detection (Feng et al., 2012; Pennebaker et al., 2015; Ott et al., 2011) overlook a crucial linguistic aspect of fake/real news documents i.e., the hierarchical discourse- level structure. Usually, in a document, discourse units (e.g., sentences) are organized in a hierarchi- cal structure e.g., a tree. The importance of con- sidering the hierarchical discourse-level structure for fake news detection is three-fold. First, pre- vious studies (Bachenko et al., 2008; Rubin and Lukoianova, 2015) explored discourse-level struc- ture in fake news detection and discovered that the way two discourse units of a document are connected could be quite revealing and insightful about its truthfulness. For instance, (Rubin and Lukoianova, 2015) applied Rhetorical Structure Theory (RST) (Mann and Thompson, 1988) and noted that fake stories lack “evidence” as a defined inter-discourse relation. Second, fake news is typ- ically produced by connecting disjoint pieces of news and unlike well-established journalism (e.g., New York Times) fake news production lacks a meticulous editorial board. Therefore, by incor- porating the hierarchical discourse-level structure, we can investigate the coherence of fake/real news documents (we will show this later). Third, a sub- stantial number of studies have shown that using hierarchical structures yields a better document representation in various downstream tasks whose predictions depend on the entire text (Bhatia et al., 2015; Morey …

Content truncated to 500 words