A user-centric model of voting intention from Social Media

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 993–1003, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics A user-centric model of voting intention from Social Media Vasileios Lampos, Daniel Preot¸iuc-Pietro and Trevor Cohn Computer Science Department University of Sheffield, UK {v.lampos,d.preotiuc,t.cohn}@dcs.shef.ac.uk Abstract Social Media contain a multitude of user opinions which can be used to predict real- world phenomena in many domains in- cluding politics, finance and health. Most existing methods treat these problems as linear regression, learning to relate word frequencies and other simple features to a known response variable (e.g., voting intention polls or financial indicators). These techniques require very careful fil- tering of the input texts, as most Social Media posts are irrelevant to the task. In this paper, we present a novel approach which performs high quality filtering au- tomatically, through modelling not just words but also users, framed as a bilin- ear model with a sparse regulariser. We also consider the problem of modelling groups of related output variables, us- ing a structured multi-task regularisation method. Our experiments on voting inten- tion prediction demonstrate strong perfor- mance over large-scale input from Twitter on two distinct case studies, outperform- ing competitive baselines. 1 Introduction Web Social Media platforms have ushered a new era in human interaction and communication. The main by-product of this activity is vast amounts of user-generated content, a type of information that has already attracted the interest of both marke- teers and scientists because it offers – for the first time at a large-scale – unmediated access to peo- ples’ observations and opinions. One exciting avenue of research concentrates on mining interesting signals automatically from this stream of text input. For example, by exploit- ing Twitter posts, it is possible to infer time series that correlate with financial indicators (Bollen et al., 2011), track infectious diseases (Lampos and Cristianini, 2010; Lampos et al., 2010; Paul and Dredze, 2011) and, in general, nowcast the magni- tude of events emerging in real-life (Sakaki et al., 2010; Lampos and Cristianini, 2012). Other stud- ies suggest ways for modelling opinions encap- sulated in this content in order to forge branding strategies (Jansen et al., 2009) or understand vari- ous socio-political trends (Tumasjan et al., 2010; O’Connor et al., 2010; Lansdall-Welfare et al., 2012). The main theme of the aforementioned works is linear regression between word frequen- cies and a real-world quantity. They also tend to incorporate hand-crafted lists of search terms to filter irrelevant content and use sentiment analy- sis lexicons for extracting opinion bias. Conse- quently, they are quite often restricted to a specific application and therefore, generalise poorly to new data sets (Gayo-Avello et al., 2011). In this paper, we propose a generic method that aims to be independent of the characteristics de- scribed above (use of search terms or sentiment analysis tools). Our approach is able to explore not only word frequencies, but also the space of users by introducing a bilinear formulation for this learning task. …

Content truncated to 500 words