A Weakly Supervised Bayesian Model for Violence Detection in Social Media

International Joint Conference on Natural Language Processing, pages 109–117, Nagoya, Japan, 14-18 October 2013. A Weakly Supervised Bayesian Model for Violence Detection in Social Media Elizabeth Cano Yulan He School of Engineering and Applied Science Aston University, UK {a.cano basave,y.he9}@aston.ac.uk Kang Liu Jun Zhao Institute of Automation Chinese Academy of Sciences, China {kliu,jzhao}@nlpr.ia.ac.cn Abstract Social streams have proven to be the most up-to-date and inclusive information on cur- rent events. In this paper we propose a novel probabilistic modelling framework, called vi- olence detection model (VDM), which en- ables the identification of text containing vio- lent content and extraction of violence-related topics over social media data. The proposed VDM model does not require any labeled cor- pora for training, instead, it only needs the in- corporation of word prior knowledge which captures whether a word indicates violence or not. We propose a novel approach of deriving word prior knowledge using the relative en- tropy measurement of words based on the in- tuition that low entropy words are indicative of semantically coherent topics and therefore more informative, while high entropy words indicates words whose usage is more topical diverse and therefore less informative. Our proposed VDM model has been evaluated on the TREC Microblog 2011 dataset to identify topics related to violence. Experimental re- sults show that deriving word priors using our proposed relative entropy method is more ef- fective than the widely-used information gain method. Moreover, VDM gives higher vio- lence classification results and produces more coherent violence-related topics compared to a few competitive baselines. 1 Introduction Social media and in particular Twitter has proven to be a faster channel of communication when compared to traditional news media, as we have witnessed dur- ing events such as the Middle East revolutions and the 2011 Japan earthquake; acting as social sensors of real-time events (Sakaki et al., 2010). Therefore the identification of topics discussed in these channels could aid in different scenarios including violence de- tection and emergency response. In particular the task of classifying tweets as violence-related poses differ- ent challenges including: high topical diversity; ir- regular and ill-formed words; event-dependent vocab- ulary characterising violence-related content; and an evolving jargon emerging from violent events. Indeed, machine learning methods for classifica- tion present difficulty on short texts (Phan et al., 2008). A large body of work has been proposed for the task of topic classification of Tweets (Milne and Witten., 2008; Gabrilovich and Markovitch, 2006; Genc et al., 2011; Mu˜noz Garc´ıa et al., 2011; Ka- siviswanathan et al., 2011; Meij et al., 2012). Re- cent approaches have also been proposed (Michelson and Macskassy, 2010; Cano et al., 2013), to alle- viate microposts sparsity by leveraging existing so- cial knowledge sources (e.g Wikipedia). However, while the majority of these approaches rely on su- pervised classification techniques, others do not cater for the violence detection challenges. To the best of our knowledge very few have been devoted to vio- lent content analysis of Twitter, and none has car- ried out deep violence-related topic …

Content truncated to 500 words