Learning Topics and Positions from Debatepedia

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1858–1868, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Learning Topics and Positions from Debatepedia Swapna Gottipati† Minghui Qiu† Yanchuan Sim‡ Jing Jiang† Noah A. Smith‡ †School of Information Systems, Singapore Management University, Singapore ‡Language Technologies Institute, Carnegie Mellon University, Pittsburgh, PA 15213, USA †{swapnag.2010,minghui.qiu.2010,jingjiang}@smu.edu.sg ‡{ysim,nasmith}@cs.cmu.edu Abstract We explore Debatepedia, a community- authored encyclopedia of sociopolitical de- bates, as evidence for inferring a low- dimensional, human-interpretable representa- tion in the domain of issues and positions. We introduce a generative model positing latent topics and cross-cutting positions that gives special treatment to person mentions and opin- ion words. We evaluate the resulting repre- sentation’s usefulness in attaching opinionated documents to arguments and its consistency with human judgments about positions. 1 Introduction The social web has evolved into a forum for large portions of the population to discuss and debate complex issues of societal importance. Websites like Debatepedia,1 an online, community-authored ency- clopedia of debates (§2), seek to organize some of this exchange into structured information resources that summarize arguments and link externally to texts (editorials, blog posts, etc.) that express and evoke them. Empirical NLP, we propose, has a role to play in creating a more compact and easily- interpretable way to understand the opinion space. In particular, we envision applications to computa- tional journalism, where there is high demand for transformation of and pattern discovery in unman- ageable, unstructured, evolving data (including text) to inform the public (Cohen et al., 2011). In this paper, we develop a generative model for discovering such a representation (§3), using De- batepedia as a corpus of evidence. We draw in- spiration from Lin et al. (2008) and Ahmed and 1http://dbp.idebate.org Xing (2010), who used generative models to infer topics—distributions over words—and other word- associated variables representing perspectives or ideologies. We view topics as lexicons, and propose that grounding a topic model with evidence beyond bags of words can lead to more lexicon-like repre- sentations. Specifically, our generative topic model grounds topics using the hierarchical organization of arguments within Debatepedia. Further, we use named entity recognition as a preprocessing step, an existing sentiment lexicon to construct an informed prior, and we incorporate a latent, discrete position variable that cuts across debates.2 We evaluate the model informally and formally (§4). Subjectively, the model identifies reasonable topic and perspective terms, and it associates topics sensibly with important public figures. In quanti- tative evaluations, we find the model’s representa- tion superior to topics from vanilla latent Dirichlet allocation (Blei et al., 2003) and the joint sentiment topic model (Lin and He, 2009) in matching external texts to debates. Further, the position variables can be used to infer the side of an argument within a de- bate; our model performs with an accuracy of 86% on position prediction of the debate argument. The cross-cutting position variable is not especially con- sistent with human judgments, suggesting that fur- ther knowledge sources may be required …

Content truncated to 500 words