Learning Disentangled Latent Topics for Twitter Rumour Veracity Classification

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 3902–3908 August 1–6, 2021. ©2021 Association for Computational Linguistics 3902 Learning Disentangled Latent Topics for Twitter Rumour Veracity Classification 1John Dougrez-Lewis 1,2,3Maria Liakata 2,3Elena Kochkina 1Department of Computer Science, University of Warwick, UK1Yulan He 2Queen-Mary University of London, UK 3Alan Turing Institute, UK {j.Dougrez-Lewis,yulan.he}@warwick.ac.uk {m.liakata,e.kochkina}@qmul.ac.uk Abstract With the rapid growth of social media in the past decade, the news are no longer controlled by just a few mainstream sources. Users them- selves create large numbers of potentially ficti- tious rumours, necessitating automated verac- ity classification systems. Here we present a novel approach towards automatically classi- fying rumours circulating on Twitter with re- spect to their veracity. We use a model built on Variational Autoencoder which disentan- gles the informational content of a tweet from the manner in which the information is writ- ten. This is achieved by obtaining latent topic vectors in an adversarial learning setting using the auxiliary task of stance classification. The latent vectors learnt in this way are used to pre- dict rumour veracity, obtaining state-of-the-art accuracy scores on the PHEME dataset.1 1 Introduction Anyone can publish rumours online with the po- tential to influence and pose as news. Since it is impossible to manually check the vast volume of circulating tweets, there is increasing need for ma- chine learning algorithms to assist with rumour veracity assessment. Given a rumour of unknown veracity introduced by a tweet in a conversation thread and the re- sponses to it, our goal is to automatically deter- mine the veracity of the rumour by assigning it one of the classes true, false, or unverified. Prior approaches to rumour veracity classification have primarily relied on careful feature engineering. For example, Li et al. (2019a) used meta-features such as user credibility together with more traditional features to top the leaderboard in SemEval 2019 Task 7 (Gorrell et al., 2019). This task encour- aged teams to use the stances of responses to the 1The code is available at: https://github.com/ JohnNLP/SAVED rumour to assist in veracity classification, which has previously been shown to be predictive of ru- mour veracity (Dungs et al., 2018). A number of approaches (Kochkina et al., 2018; Li et al., 2019b) also showed benefits of using stance classification as an auxiliary task in a multitask learning setup. Some recent approaches exploit the structure of the conversation discussing a rumour. Kochkina et al. (2018) used LSTM to model linear branches ex- tracted from the conversation tree, while Ma et al. (2018) and Bian et al. (2020) modelled a tree struc- ture to capture information from responses. Zeng et al. (2019) presented an unsupervised approach built on Variational Autoencoder (VAE) to jointly model topic content and discourse be- haviour in microblog conversations. We propose a novel architecture which incorporates a VAE with adversarial learning to disentangle topics which are informative for stance classification from those which are not. We then derive tweet representa- tions based on the word representations learned in the latent …

Content truncated to 500 words