Sentiment Propagation via Implicature Constraints

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 377–385, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Sentiment Propagation via Implicature Constraints Lingjia Deng Intelligent Systems Program University of Pittsburgh lid29@pitt.edu Janyce Wiebe Department of Computer Science University of Pittsburgh wiebe@cs.pitt.edu Abstract Opinions may be expressed implicitly via inference over explicit sentiments and events that positively/negatively affect en- tities (goodFor/badFor events). We in- vestigate how such inferences may be exploited to improve sentiment analysis, given goodFor/badFor event information. We apply Loopy Belief Propagation to propagate sentiments among entities. The graph-based model improves over explicit sentiment classification by 10 points in precision and, in an evaluation of the model itself, we find it has an 89% chance of propagating sentiments correctly. 1 Introduction Previous research in sentiment analysis and opinion extraction has largely focused on the interpretation of explicitly stated opinions. How- ever, many opinions are expressed implicitly via opinion implicature (i.e., opinion-oriented defeasible inference). Consider the following sentence: EX(1) The bill would lower health care costs, which would be a tremendous positive change across the entire health-care system. The writer is clearly positive toward the idea of lowering health care costs. But how does s/he feel about the costs? If s/he is positive toward the idea of lowering them, then, presumably, she is nega- tive toward the costs themselves (specifically, how high they are). The only explicit sentiment expres- sion, tremendous positive change, is positive, yet we can infer a negative attitude toward the object of the event itself (i.e., health care costs). Going further, since the bill is the agent of an event toward which the writer is positive, we may (defeasibly) infer that the writer is positive toward the bill, even though there are no explicit senti- ment expressions describing it. Now, consider The bill would curb skyrocketing health care costs. The writer expresses an explicit negative sentiment (skyrocketing) toward the ob- ject (health care costs) of the event. Note that curbing costs, like lowering them, is bad for them (the costs are reduced). We can reason that, be- cause the event is bad for something toward which the writer is negative, the writer is positive toward the event. We can reason from there, as above, that the writer is positive toward the bill, since it is the agent of the positive event. These examples illustrate how explicit sen- timents toward one entity may be propagated to other entities via opinion implicature rules. The rules involve events that positively or nega- tively affect entities. We call such events good- For/badFor (hereafter gfbf)events. This work investigates how gfbf event interac- tions among entities, combined with opinion in- ferences, may be exploited to improve classifica- tion of the writer’s sentiments toward entities men- tioned in the text. We introduce four rule schemas which reveal sentiment constraints among gfbf events and their agents and objects. Those con- straints are incorporated into a graph-based model, where a node represents an entity (agent/object), …

Content truncated to 500 words