Tree LSTMs with Convolution Units to Predict Stance and Rumor Veracity in Social Media Conversations

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5047–5058 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 5047 Tree LSTMs with Convolution Units to Predict Stance and Rumor Veracity in Social Media Conversations Sumeet Kumar Carnegie Mellon University 5000 Forbes Ave, Pittsburgh, PA 15213, USA sumeetku@cs.cmu.edu Kathleen M. Carley Carnegie Mellon University 5000 Forbes Ave, Pittsburgh, PA 15213, USA kathleen.carley@cs.cmu.edu Abstract Learning from social-media conversations has gained significant attention recently because of its applications in areas like rumor detec- tion. In this research, we propose a new way to represent social-media conversations as bi- narized constituency trees that allows compar- ing features in source-posts and their replies effectively. Moreover, we propose to use con- volution units in Tree LSTMs that are better at learning patterns in features obtained from the source and reply posts. Our Tree LSTM models employ multi-task (stance + rumor) learning and propagate the useful stance sig- nal up in the tree for rumor classification at the root node. The proposed models achieve state- of-the-art performance, outperforming the cur- rent best model by 12% and 15% on F1-macro for rumor-veracity classification and stance classification tasks respectively. 1 Introduction Online misinformation, commonly called ‘fake news’, has become a serious problem in society (Ferrara, 2015) to the extent that they are im- pacting election decisions (Allcott and Gentzkow, 2017). Many machine-learning approaches have been proposed to identify and contain the fake- news shared on online social-media platforms (Jin et al., 2016; Rubin et al., 2016; Rubin and Lukoianova, 2015; Schifferes et al., 2014; Tac- chini et al., 2017; Volkova et al., 2017; Vosoughi et al., 2018). One approach that combines machine-learning and human-intelligence by ex- ploiting stance in reply posts has gained significant attention recently (Zubiaga et al., 2016a, 2015). In this approach, we first identify the stance – cate- gorized as ‘supporting’, ‘denying’, ‘commenting’ and ‘querying’ – in the replies to the original post and then use the stance signal to find rumor ve- racity i.e. if a rumor is true or false. Prior work Putin is missing. www.abcnews.co.ir Source'Tweet' This is not verified It’s on TV as well Stance:'Deny' Stance:'Deny' He went missing last week Stance:'Favor' Rumor: False T1: R1: R11: R2: ' Reply'Tweet' Reply'Tweet' Reply'Tweet' Figure 1: Twitter threads with stance and rumor- veracity labels. The conversation tree shown above has two branches a) T1–R1–R11 and b) T1-R2. R1 and R2 are 1st level reply tweets and R11 is a 2nd level reply tweet. Stance labels for each reply is relative to the tweet it is replied to i.e. stance for R11 is with-respect- to R1. There is a rumor-veracity label on the root tweet (T1 in the example above). The goal of this research is to learn the root tweet’s veracity based on pattern in replies. has confirmed that replies to a ‘false’ (misleading) rumor contain specific patterns, e.g. more replies deny the claim made in the source post (Zubiaga et al., 2016b). This approach …

Content truncated to 500 words