University of Pittsburgh

Latent Variable Models for Viewpoint Discovery in Online Discussions

Visiting PhD Student
Friday, March 7, 2014 - 12:30pm - 1:30pm


Online discussion forums provide an important social media platform that allows netizens to express their opinions, to ask for advice, and to form online communities. In particular, responses to major sociopolitical events and issues can often be found in discussion forums, which serve as an important source of public feedback. How to automatically mine user viewpoints from the forum threads is the focus of this talk.
We study three latent variable models for modeling and discovering different viewpoints in forum threads. Our first model takes two important characteristics of threaded forum data into consideration, namely, user consistency and user interactions. The former refers to the observation that a user’s opinion on an issue usually remains unchanged during a certain time period. The latter refers to the observation that in a forum thread, like in conversations, users interact with each other by commenting on each other’s posts. Thus, modeling the agreement/disagreement expressions among users can help find different viewpoints.

Our second model addresses the sparsity of user interactions in online discussions. We first make use of the advances in sentiment analysis to extract user opinions in online user interactions. As this user interactions can be very sparse, we propose to apply collaborative filtering through matrix factorization to generalize and improve the extracted opinion matrices from forum posts. The resulting low-rank latent factor representations of users makes it feasible to cluster users by their viewpoints.

Last but not least, in our ongoing work, we study viewpoint discovery for a new issue which has a low online participation rate of Internet users in online discussion forums. We propose an integrated model that jointly models texts, user viewpoints and social networks. We consider hidden factor models to model user viewpoints, and user texts to give a human-interpretable explanation for each hidden factor. We also incorporate social context to boost the model performance. Our experiments show promising results on the viewpoint discovery task.

Minghui is a PhD candidate in the School of Information Systems, Singapore Management University, under the supervision of Jing Jiang. He works in the area of text mining, social network mining and recommender systems. His primary research interest focuses on mining controversial topics and finding viewpoints. Minghui has published several papers in top venues such as CIKM, NAACL and SDM, including a co-authored paper that was a runner-up for the best paper award in SocInfo’13. He is currently visiting Language Technologies Institute, Carnegie Mellon University, working with Noah Smith and Alex Smola.

Copyright 2009 | Web site by UMC Web Team