Stance Classification in Rumours as a Sequential Task Exploiting the Tree Structure of Social Media Conversations

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2438–2448, Osaka, Japan, December 11-17 2016. Stance Classification in Rumours as a Sequential Task Exploiting the Tree Structure of Social Media Conversations Arkaitz Zubiaga1, Elena Kochkina1, Maria Liakata1, Rob Procter1, Michal Lukasik2 1 University of Warwick, Coventry, UK 2 University of Sheffield, Sheffield, UK {a.zubiaga,e.kochkina,m.liakata,rob.procter}@warwick.ac.uk m.lukasik@sheffield.ac.uk Abstract Rumour stance classification, the task that determines if each tweet in a collection discussing a ru- mour is supporting, denying, questioning or simply commenting on the rumour, has been attract- ing substantial interest. Here we introduce a novel approach that makes use of the sequence of transitions observed in tree-structured conversation threads in Twitter. The conversation threads are formed by harvesting users’ replies to one another, which results in a nested tree-like struc- ture. Previous work addressing the stance classification task has treated each tweet as a separate unit. Here we analyse tweets by virtue of their position in a sequence and test two sequential clas- sifiers, Linear-Chain CRF and Tree CRF, each of which makes different assumptions about the conversational structure. We experiment with eight Twitter datasets, collected during breaking news, and show that exploiting the sequential structure of Twitter conversations achieves sig- nificant improvements over the non-sequential methods. Our work is the first to model Twitter conversations as a tree structure in this manner, introducing a novel way of tackling NLP tasks on Twitter conversations. 1 Introduction Rumour stance classification is a task that is increasingly gaining popularity in its application to tweets. While Twitter is a generous source of reports of breaking news, outpacing even news outlets (Kwak et al., 2010), it also comes with the caveat that some of those reports are still rumours at the time of posting and so are yet to be verified and corroborated (Mendoza et al., 2010; Procter et al., 2013b; Procter et al., 2013a). The rumour stance classification task intends to assist in this verification process by determining the type of support expressed in different tweets discussing the same rumour (Qazvinian et al., 2011). Aggregation of the stance of multiple tweets discussing a rumour can then be of help to determine its likely veracity, enabling – among other benefits – the flagging of highly disputed rumours that are likely to be false. Previous research on rumour stance classification for tweets has been limited to the tweet as the unit to be classified. However, such approaches ignore the additional context and knowledge that can be gained from the structure of Twitter interactions within conversational threads (Zubiaga et al., 2016; Procter et al., 2013b; Tolmie et al., 2015). The latter are formed as Twitter users reply to one another’s posts and ultimately users build on each others’ stance towards the rumour, leading to a potential consensus. For example, a tweet may report a rumour (source tweet) and others may reply to it by further supporting it or providing counter-evidence. Our objective here is to mine the sequence of stance types encountered …

Content truncated to 500 words