A Factored Neural Network Model for Characterizing Online Discussions in Vector Space

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2296–2306 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics A Factored Neural Network Model for Characterizing Online Discussions in Vector Space Hao Cheng Hao Fang Mari Ostendorf University of Washington {chenghao,hfang,ostendorf}@uw.edu Abstract We develop a novel factored neural model that learns comment embeddings in an un- supervised way leveraging the structure of distributional context in online discussion forums. The model links different con- text with related language factors in the embedding space, providing a way to in- terpret the factored embeddings. Eval- uated on a community endorsement pre- diction task using a large collection of topic-varying Reddit discussions, the fac- tored embeddings consistently achieve im- provement over other text representations. Qualitative analysis shows that the model captures community style and topic, as well as response trigger patterns. 1 Introduction Massive user-generated content on social media has drawn interests in predicting community reac- tions in the form of virality (Guerini et al., 2011), popularity (Suh et al., 2010; Hong et al., 2011; Lakkaraju et al., 2013; Tan et al., 2014), commu- nity endorsement (Jaech et al., 2015; Fang et al., 2016), persuasive impact (Althoff et al., 2014; Tan et al., 2016; Wei et al., 2016), etc. Many of these studies have analyzed content-agnostic fac- tors such as submission timing and author social status, as well as language factors that underlie the textual content, e.g., the topic and idiosyncrasies of the community. In particular, there is an in- creasing amount of work on online discussion fo- rums such as Reddit that exploits the conversa- tional and community-centric nature of the user- generated content (Lakkaraju et al., 2013; Althoff et al., 2014; Jaech et al., 2015; Tan et al., 2016; Wei et al., 2016; He et al., 2016a; Fang et al., 2016), which contrasts with Twitter where the au- thor’s social status seems to play a larger role in popularity. This paper focuses on Reddit, using the karma score1 as a readily available measure of community endorsement. Some of the prior work on Reddit investi- gates specific linguistic phenomena (e.g. polite- ness, topic relevance, community style matching) using feature engineering to understand their role in predicting community reactions (Althoff et al., 2014; Jaech et al., 2015). In contrast, this pa- per explores methods for unsupervised text em- bedding learning using a model structured so as to provide some interpretability of the results when used in comment endorsement prediction. The model aims to characterize the interdependence of comment on its global context and subsequent responses that is characteristic of multi-party dis- cussions. Specifically, we propose a factored neu- ral model with separate mechanisms for represent- ing global context, comment content and response generation. By factoring the model, we hope un- supervised learning will pick up different compo- nents of interactive language in the resulting em- beddings, which will improve prediction of com- munity reactions. Distributed representations of text, or text em- beddings, have achieved great success in many language …

Content truncated to 500 words