UC San Diego

A standard feature of the contemporary internet landscape is the ability for people to comment on published content and to interact with other individuals, discussing the issues at hand and engaging with each other in debate. In this thesis, I describe a method for the automatic detection of author stances in online forums with respect to discussions on divisive, polarizing social issues, such as gun control and marriage equality {--} a task which is often unproblematic for human readers of the discourse. The research investigates the linguistic and rhetorical devices used by discussion participants to express their topic stance in the context of multi-party, multi-threaded discourse. Along the way, I address necessary sub-tasks in the author stance detection problem, such as the classification of the topic stance of an individual contribution to the discourse, and the assessment of the level of agreement or disagreement between adjacent posts {--} which is crucial, given the highly interactive nature of this genre. I also identify features that provide evidence of an author's topic stance from the very structure of the discourse, without any information at all from the text of the comments posted. The final model is a collective classifier that is able to synthesize all of the stance indicators provided by these different sources, deal with the inconsistencies in this information that may arise, and arrive at a single prediction of the topic stance for every participant in the discussion. The model has many applications in industry and public life, including more tailored newsfeeds, social network suggestions, and use in political fundraising or advocacy campaigns.