Microblog Conversation Recommendation via Joint Modeling of Topics and Discourse

Proceedings of NAACL-HLT 2018, pages 375–385 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Microblog Conversation Recommendation via Joint Modeling of Topics and Discourse Xingshan Zeng1,3, Jing Li2, Lu Wang4, Nicholas Beauchamp5,6, Sarah Shugars6, Kam-Fai Wong1,3 1The Chinese University of Hong Kong, 2Tencent AI Lab 3MoE Key Laboratory of High Confidence Software Technologies, China 4College of Computer and Information Science, Northeastern University, US 5Department of Political Science, Northeastern University, US 6Network Science Institute, Northeastern University, US 1,3{xszeng,kfwong}@se.cuhk.edu.hk 2ameliajli@tencent.com, 5n.beauchamp@northeastern.edu,4luwang@ccs.neu.edu 6shugars.s@husky.neu.edu Abstract Millions of conversations are generated ev- ery day on social media platforms. With lim- ited attention, it is challenging for users to se- lect which discussions they would like to par- ticipate in. Here we propose a new method for microblog conversation recommendation. While much prior work has focused on post- level recommendation, we exploit both the conversational context, and user content and behavior preferences. We propose a statistical model that jointly captures: (1) topics for rep- resenting user interests and conversation con- tent, and (2) discourse modes for describing user replying behavior and conversation dy- namics. Experimental results on two Twitter datasets demonstrate that our system outper- forms methods that only model content with- out considering discourse. 1 Introduction Online platforms have revolutionized the way in- dividuals collect and share information (O’Connor et al., 2010; Lee and Ma, 2012; Bakshy et al., 2015), but the vast bulk of online content is ir- relevant or unpalatable to any given individual. A user interested in political discussion, for in- stance, might prefer content concerning a specific candidate or issue, and only then if discussed in a positive light without controversy (Adamic and Glance, 2005; Bakshy et al., 2015). How do individuals facing such large quanti- ties of superfluous material select which conver- sations to engage in, and how might we better al- gorithmically recommend conversations suited to individual users? We approach this problem from a microblog conversation recommendation frame- work. Where prior work has focused on the con- tent of individual posts for recommendation (Chen ...Conversation 1 [U1]: The sheer cognitive dissonance required for a “lib- eral” to say Clinton is as bad as Trump is just staggering. [U2]: Hillarists, Troll; they insult Liberals trying to dis- tract from Hillary’s Conseratism. [U3]: I still prefer Hillarist b/c it describes their Cultish and ideological aspects. ...—————————— ...Conversation 2 [U4]: I do not like trump at all, but Comey left her in place knowing Bernie is much stronger. [U1]: If you’re going to actively start rooting against the Democrats, get off my mentions. I have enough GOP doing that. [U5]: Your tweets are an example of why open primaries are stupid. You’re not a Dem, you’re just for one guy. —————————— [U1]: No offense, but you’ve been wrong about pretty much everything so far. Why would I trust your prognos- tication now? ... Figure 1: Two snippets of conversations on Twitter. [Ui]: The message is posted by user Ui. “—” is the dividing line between training history …

Content truncated to 500 words