Document-level Sentiment Inference with Social, Faction, and Discourse Context

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 333–343, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Document-level Sentiment Inference with Social, Faction, and Discourse Context Eunsol Choi Hannah Rashkin Luke Zettlemoyer Yejin Choi Computer Science & Engineering University of Washington {eunsol,hrashkin,lsz,yejin}@cs.washington.edu Abstract We present a new approach for document- level sentiment inference, where the goal is to predict directed opinions (who feels positively or negatively towards whom) for all entities mentioned in a text. To encour- age more complete and consistent predic- tions, we introduce an ILP that jointly models (1) sentence- and discourse-level sentiment cues, (2) factual evidence about entity factions, and (3) global constraints based on social science theories such as homophily, social balance, and reci- procity. Together, these cues allow for rich inference across groups of entities, includ- ing for example that CEOs and the com- panies they lead are likely to have simi- lar sentiment towards others. We evalu- ate performance on new, densely labeled data that provides supervision for all pairs, complementing previous work that only labeled pairs mentioned in the same sen- tence. Experiments demonstrate that the global model outperforms sentence-level baselines, by providing more coherent pre- dictions across sets of related entities. 1 Introduction Documents often present a complex web of facts and opinions that hold among the entities they de- scribe. Consider the international relations story in Figure 1. Representatives from three countries form factions and create a network of sentiment. While some opinions are relatively directly stated (e.g., Russia criticizes Belarus), many others must be inferred based on the factual ties among enti- ties (e.g., Moscow, Gryzlov, and Russia probably share the same sentiment towards other entities) and known social context (e.g., Russia probably Russia criticized Belarus for permitting Georgian Presi- dent Mikheil Saakhashvili to appear on Belorussian tele- vision. “The appearance was an unfriendly step towards Russia,” the speaker of Russian parliament Boris Gry- zlov said. . . . Saakhashvili announced Thursday that he did not understand Russia’s claims. Moscow refused to have any business with Georgia’s president after the armed conflict in 2008 . . . Figure 1: Example text excerpt paired with the document- level sentiment graph we aim to recover. The graph includes edges with direct textual support (e.g., from Russian to Be- larus given the verb “criticized”) as well as ones that must be inferred at the whole-document level (e.g., from Gryzlov to Saakhashvili given the web of relationships and opinions between them, Georgia, Russian, and Belarus). dislikes Saakhashvili since Russia criticized Be- larus for supporting him). In this paper, we show that jointly reasoning about all of these factors can provide more complete and consistent document- level sentiment predictions. More concretely, we present a global model for document-level entity-to-entity sentiment, i.e., who feels positively (or negatively) towards whom. Our goal is to make exhaustive predictions over all entity pairs, including those that require cross- sentence inference. We present a Integer Lin- ear Programming (ILP) model that combines …

Content truncated to 500 words