Joint Inference and Disambiguation of Implicit Sentiments via Implicature Constraints

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 79–88, Dublin, Ireland, August 23-29 2014. Joint Inference and Disambiguation of Implicit Sentiments via Implicature Constraints Lingjia Deng1, Janyce Wiebe1,2, Yoonjung Choi2 1Intelligent Systems Program, University of Pittsburgh 2Department of Computer Science, University of Pittsburgh lid29@pitt.edu, wiebe@cs.pitt.edu, yjchoi@cs.pitt.edu Abstract This paper addresses implicit opinions expressed via inference over explicit sentiments and events that positively/negatively affect entities (goodFor/badFor, gfbf events). We incorporate the inferences developed by implicature rules into an optimization framework, to jointly improve sentiment detection toward entities and disambiguate components of gfbf events. The framework simultaneously beats the baselines by more than 10 points in F-measure on sentiment detection and more than 7 points in accuracy on gfbf polarity disambiguation. 1 Introduction Previous work in NLP on sentiment analysis has mainly focused on explicit sentiments. However, as noted in (Deng and Wiebe, 2014), many opinions are expressed implicitly, as shown by this example: Ex(1) The reform would lower health care costs, which would be a tremendous positive change across the entire health-care system. There is an explicit positive sentiment toward the event of “reform lower costs”. However, in expressing this sentiment, the writer also implies he is negative toward the “costs”, since he’s happy to see the costs being decreased. Moreover, the writer may be positive toward “reform” since it contributes to the “lower” event. Such inferences may be seen as opinion-oriented implicatures (i.e., defeasible inferences)1. We develop a set of rules for inferring and detecting implicit sentiments from explicit sentiments and events such as “lower” (Wiebe and Deng, 2014). In (Deng et al., 2013), we investigate such events, defining a badFor (bf) event to be an event that negatively affects the theme and a goodFor (gf) event to be an event that positively affects the theme of the event.2 Here, “lower” is a bf event. According to their annotation scheme, goodFor/badFor (gfbf) events have NP agents and themes (though the agent may be implicit), and the polarity of a gf event may be changed to bf by a reverser (and vice versa). The ultimate goal of this work is to utilize gfbf information to improve detection of the writer’s senti- ments toward entities mentioned in the text. However, this requires resolving several ambiguities: (Q1) Given a document, which spans are gfbf events? (Q2) Given a gfbf text span, what is its polarity, gf or bf? (Q3) Is the polarity of a gfbf event being reversed? (Q4) Which NP in the sentence is the agent and which is the theme? (Q5) What are the writer’s sentiments toward the agent and theme, positive or negative? Fortunately, the implicature rules in (Deng and Wiebe, 2014) define dependencies among these ambiguities. As in Ex(1), the sentiments toward the agent and theme, the sentiment toward the gfbf event (positive or negative), and the polarity of the gfbf event (gf or bf) are all interdependent. Thus, rather than having to take a pipeline approach, we are able to develop an …

Content truncated to 500 words