Mining User Relations from Online Discussions using Sentiment Analysis and Probabilistic Matrix Factorization

Proceedings of NAACL-HLT 2013, pages 401–410, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Mining User Relations from Online Discussions using Sentiment Analysis and Probabilistic Matrix Factorization Minghui Qiu†, Liu Yang†,‡, Jing Jiang† † School of Information Systems, Singapore Management University, Singapore ‡ School of Software and Microelectronics, Peking University, China {minghui.qiu.2010,jingjiang}@smu.edu.sg, yang.liu@pku.edu.cn Abstract Advances in sentiment analysis have enabled extraction of user relations implied in online textual exchanges such as forum posts. How- ever, recent studies in this direction only con- sider direct relation extraction from text. As user interactions can be sparse in online dis- cussions, we propose to apply collaborative filtering through probabilistic matrix factor- ization to generalize and improve the opinion matrices extracted from forum posts. Exper- iments with two tasks show that the learned latent factor representation can give good per- formance on a relation polarity prediction task and improve the performance of a subgroup detection task. 1 Introduction The fast growth of the social Web has led to a large amount of interest in online social network analysis. Most existing work on social network analysis re- lies on explicit links among users such as undirected friendship relations (Liben-Nowell and Kleinberg, 2003), directed following relations (Hopcroft et al., 2011) and trust/distrust relations (Leskovec et al., 2010). However, besides these explicit social rela- tions, the various kinds of interactions between on- line users often suggest other implicit relations. In particular, in online discussion forums, users inter- act through textual posts and these exchanged texts often reveal whether two users are friends or foes, or whether two users share the same viewpoint towards a given issue. To uncover such implicit relations requires text analysis and particularly sentiment analysis. Re- cently, Hassan et al. (2012) studied predicting the polarity of user interactions in online discussions based on textual exchanges. They found that the au- tomatically predicted signed relations had an accu- racy above 80%. The extracted signed network was further used to detect ideological subgroups. This is a piece of pioneering work that extracts online social relations based on text analysis. In this paper, we further extend the idea of mining social relations from online forum posts by incorpo- rating collaborative filtering. Our work is motivated by the observation that direct textual exchanges be- tween users are sparse. For example, in the data set we use, only around 13% of user-user pairs have di- rect interactions. Collaborative filtering is a com- monly used technique in recommender systems to predict missing ratings. The key assumption is that if two people have the same opinion on an item A, they are likely to also have the same opinion on a different item B. In online discussion forums, users express their opinions about each other as well as the various aspects of the topic under discussion, but not every user comments on every aspect or every other user. Collaborative filtering allows us to iden- tify users with the same opinion even if they have not directly interacted with each …

Content truncated to 500 words