Towards Propagation Uncertainty: Edge-enhanced Bayesian Graph Convolutional Networks for Rumor Detection

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3845–3854 August 1–6, 2021. ©2021 Association for Computational Linguistics 3845 Towards Propagation Uncertainty: Edge-enhanced Bayesian Graph Convolutional Networks for Rumor Detection Lingwei Wei1,4, Dou Hu2, Wei Zhou1∗, Zhaojuan Yue3, Songlin Hu1,4∗ 1 Institute of Information Engineering, Chinese Academy of Sciences 2 National Computer System Engineering Research Institute of China 3 Computer Network Information Center, Chinese Academy of Sciences 4 School of Cyber Security, University of Chinese Academy of Sciences {weilingwei18, hudou18}@mails.ucas.edu.cn {zhouwei, husonglin}@iie.ac.cn yuezhaojuan@cnic.cn Abstract Detecting rumors on social media is a very critical task with significant implications to the economy, public health, etc. Previ- ous works generally capture effective features from texts and the propagation structure. How- ever, the uncertainty caused by unreliable re- lations in the propagation structure is com- mon and inevitable due to wily rumor pro- ducers and the limited collection of spread data. Most approaches neglect it and may seriously limit the learning of features. To- wards this issue, this paper makes the first attempt to explore propagation uncertainty for rumor detection. Specifically, we pro- pose a novel Edge-enhanced Bayesian Graph Convolutional Network (EBGCN) to capture robust structural features. The model adap- tively rethinks the reliability of latent relations by adopting a Bayesian approach. Besides, we design a new edge-wise consistency train- ing framework to optimize the model by en- forcing consistency on relations. Experiments on three public benchmark datasets demon- strate that the proposed model achieves better performance than baseline methods on both ru- mor detection and early rumor detection tasks. 1 Introduction With the ever-increasing popularity of social me- dia sites, user-generated messages can quickly reach a wide audience. However, social media can also enable the spread of false rumor infor- mation (Vosoughi et al., 2018). Rumors are now viewed as one of the greatest threats to democracy, journalism, and freedom of expression. Therefore, detecting rumors on social media is highly desir- able and socially beneficial (Ahsan et al., 2019). * Corresponding author. tweet relations Constructed Graph/Tree 1 5 2 4 Real Propagation 1 2 4 5 3 6 x Inaccurate relations Figure 1: An example of uncertain propagation struc- ture. It includes inaccurate relations, making con- structed graph inconsistent with the real propagation. Almost all the previous studies on rumor de- tection leverage text content including the source tweet and all user retweets or replies. As time goes on, rumors form their specific propagation struc- tures after being retweeted or replied to. Vosoughi (2015); Vosoughi et al. (2018) have confirmed ru- mors spread significantly farther, faster, deeper, and more broadly than the truth. They provide the possi- bility of detecting rumors through the propagation structure. Some works (Ma et al., 2016; Kochkina et al., 2018) typically learn temporal features alone from propagation sequences, ignoring the internal topology. Recent approaches (Ma et al., 2018; Khoo et al., 2020) model the propagation struc- ture as trees to capture structural features. Bian …

Content truncated to 500 words