Detecting Subgroups in Online Discussions by Modeling Positive and Negative Relations among Participants

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 59–70, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Detecting Subgroups in Online Discussions by Modeling Positive and Negative Relations among Participants Ahmed Hassan Microsoft Research Redmond, WA hassanam@microsoft.com Amjad Abu-Jbara University of Michigan Ann Arbor, MI amjbara@umich.edu Dragomir Radev University of Michigan Ann Arbor, MI radev@umich.edu Abstract A mixture of positive (friendly) and nega- tive (antagonistic) relations exist among users in most social media applications. However, many such applications do not allow users to explicitly express the polarity of their interac- tions. As a result most research has either ig- nored negative links or was limited to the few domains where such relations are explicitly expressed (e.g. Epinions trust/distrust). We study text exchanged between users in online communities. We find that the polarity of the links between users can be predicted with high accuracy given the text they exchange. This allows us to build a signed network represen- tation of discussions; where every edge has a sign: positive to denote a friendly relation, or negative to denote an antagonistic relation. We also connect our analysis to social psy- chology theories of balance. We show that the automatically predicted networks are consis- tent with those theories. Inspired by that, we present a technique for identifying subgroups in discussions by partitioning singed networks representing them. 1 Introduction Most online communities involve a mixture of pos- itive and negative relations between users. Positive relations may indicate friendship, agreement, or ap- proval. Negative relations usually indicate antago- nism, opposition, or disagreement. Most of the research on relations in social media applications has almost exclusively focused on pos- itive links between individuals (e.g. friends, fans, followers, etc.). We think that one of the main rea- sons, of why the interplay of positive and negative links did not receive enough attention, is the lack of a notion for explicitly expressing negative interac- tions. Recently, this problem has received increas- ing attention. However, all studies have been limited to a handful of datasets from applications that allow users to explicitly label relations as either positive or negative (e.g. trust/distrust on Epinion (Leskovec et al., 2010b) and friends/foes on Slashdot (Kunegis et al., 2009)). Predicting positive/negative relations between discussants is related to another well studied prob- lem, namely debate stance recognition. The ob- jective of this problem is to identify which partic- ipants are supporting and which are opposing the topic being discussed. This line of work does not pay enough attention to the relations between par- ticipants, rather it focuses on participant’s stance to- ward the topic. It also assumes that every partici- pant either supports or opposes the topic being dis- cussed. This is a simplistic view that ignore the nature of complex topics that has many aspects in- volved which may result in more than two subgroups with different opinions. In this work, we apply Natural Language Pro- cessing techniques to text …

Content truncated to 500 words