Classifying Tweet Level Judgements of Rumours in Social Media

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2590–2595, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Classifying Tweet Level Judgements of Rumours in Social Media Michal Lukasik,1 Trevor Cohn2 and Kalina Bontcheva1 1Computer Science 2Computing and Information Systems The University of Sheffield The University of Melbourne {m.lukasik,k.bontcheva}@shef.ac.uk t.cohn@unimelb.edu.au Abstract Social media is a rich source of rumours and corresponding community reactions. Rumours reflect different characteristics, some shared and some individual. We for- mulate the problem of classifying tweet level judgements of rumours as a super- vised learning task. Both supervised and unsupervised domain adaptation are con- sidered, in which tweets from a rumour are classified on the basis of other annotated rumours. We demonstrate how multi-task learning helps achieve good results on ru- mours from the 2011 England riots. 1 Introduction There is an increasing need to interpret and act upon rumours spreading quickly through social media, especially in circumstances where their ve- racity is hard to establish. For instance, during an earthquake in Chile rumours spread through Twit- ter that a volcano had become active and that there was a tsunami warning in Valparaiso (Mendoza et al., 2010). Other examples, from the riots in Eng- land in 2011, were that rioters were going to attack Birmingham’s children hospital and that animals had escaped from the zoo (Procter et al., 2013). Social scientists (Procter et al., 2013) analysed manually a sample of tweets expressing different judgements towards rumours and categorised them manually in supporting, denying or questioning. The goal here is to carry out tweet-level judge- ment classification automatically, in order to assist in (near) real-time rumour monitoring by journal- ists and authorities (Procter et al., 2013). In ad- dition, information about tweet-level judgements has been used as a first step for early rumour de- tection by (Zhao et al., 2015). The focus here is on tweet-level judgement clas- sification on unseen rumours, based on a training text position Birmingham Children’s hospital has been attacked. F***ing morons. #UKRiots support Girlfriend has just called her ward in Birmingham Children’s Hospital & there’s no sign of any trouble #Birm- inghamriots deny Birmingham children’s hospital guarded by police? Really? Who would target a childrens hospital #disgusting #Birminghamriots question Table 1: Tweets on a rumour about hospital being attacked during 2011 England Riots. set of other already annotated rumours. Previous work on this problem either considered unrealis- tic settings ignoring temporal ordering and rumour identities (Qazvinian et al., 2011) or proposed reg- ular expressions as a solution (Zhao et al., 2015). We expect posts expressing similar opinions to ex- hibit many similar characteristics across different rumours. Based on the assumption of a common underlying linguistic signal, we build a transfer learning system that labels newly emerging ru- mours for which we have little or no annotated data. Results demonstrate that Gaussian Process- based multi task learning allows for significantly improved performance. The novel contributions of this paper are: 1. Formulating the problem of classifying judge- ments …

Content truncated to 500 words