Joint Prediction for Entity/Event-Level Sentiment Analysis using Probabilistic Soft Logic Models

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 179–189, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Joint Prediction for Entity/Event-Level Sentiment Analysis using Probabilistic Soft Logic Models Lingjia Deng Intelligent Systems Program University of Pittsburgh lid29@pitt.edu Janyce Wiebe Intelligent Systems Program Department of Computer Science University of Pittsburgh wiebe@cs.pitt.edu Abstract In this work, we build an entity/event-level sentiment analysis system, which is able to recognize and infer both explicit and im- plicit sentiments toward entities and events in the text. We design Probabilistic Soft Logic models that integrate explicit senti- ments, inference rules, and +/-effect event information (events that positively or neg- atively affect entities). The experiments show that the method is able to greatly im- prove over baseline accuracies in recog- nizing entity/event-level sentiments. 1 Introduction There are increasing numbers of opinions ex- pressed in various genres, including reviews, newswire, editorials, and forums. While much early work was at the document or sentence level, to fully understand and utilize opinions, re- searchers are increasingly carrying out more fine- grained sentiment analysis to extract components of opinion frames: the source (whose sentiment is it), the polarity, and the target (what is the senti- ment toward). Much fine-grained analysis is span or aspect based (Yang and Cardie, 2014; Pontiki et al., 2014). In contrast, this work contributes to entity/event-level sentiment analysis. A sys- tem that could recognize sentiments toward enti- ties and events would be valuable in an application such as Automatic Question Answering, to sup- port answering questions such as “Who is nega- tive/positive toward X?” (Stoyanov et al., 2005), where X could be any entity or event. Let us consider an example from the MPQA opinion annotated corpus (Wiebe et al., 2005a; Wilson, 2007; Deng and Wiebe, 2015). Ex(1) When the Imam ( may God be satisfied with him 1) issued the fatwa against 2 Salman Rushdie for insulting 3 the Prophet ( peace be upon him 4), the countries that are so-called 5 supporters of human rights protested against 6 the fatwa. There are several sentiment expressions anno- tated in MPQA. In the first clause, the writer is positive toward Imam and Prophet as expressed by may God be satisfied with him (1) and peace be upon him (4), respectively. Imam is negative toward Salman Rushdie and the insulting event, as revealed by the expression issued the fatwa against (2). And Salman Rushdie is negative to- ward Prophet, as revealed by the expression insult- ing (3). In the second clause, the writer is negative toward the countries, as expressed by so-called (5). And the countries are negative toward fatwa, as revealed by the expression protested against (6). Using the source and the target, we summa- rize the positive opinions above in a set P, and the negative opinions above in another set N. Thus, P contains {(writer, Imam), (writer, Prophet)}, and N contains {(Imam, Rushdie), (Imam, insulting), (Rushdie, Prophet), (writer, countries), (countries, fatwa)}.1 An (ideal) explicit sentiment analysis system is …

Content truncated to 500 words