COMMIT at SemEval-2016 Task 5: Sentiment Analysis with Rhetorical Structure Theory

Proceedings of SemEval-2016, pages 356–360, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics COMMIT at SemEval-2016 Task 5: Sentiment Analysis with Rhetorical Structure Theory Kim Schouten and Flavius Frasincar Erasmus University Rotterdam P.O. Box 1738, NL-3000 DR Rotterdam, the Netherlands Abstract This paper reports our submission to the Aspect-Based Sentiment Analysis task of Se- mEval 2016. It covers the prediction of sen- timent for a given set of aspects (e.g., sub- task 1, slot 2) for the English language using discourse analysis. To that end, a discourse parser implementing the Rhetorical Structure Theory is employed and the resulting informa- tion is used to determine the context of each aspect, as well as to compute the expressed sentiment in that context by weighing the dis- course relations between words. While dis- course analysis yields high level linguistic in- formation that can be used to better predict sentiment, the proposed algorithm does not yet stack up to the high-performing machine learning approaches that are commonly ex- ploited for this task. 1 Introduction With sentiment analysis being at the forefront of re- search, many avenues are explored to find that one new algorithm that will outperform all others. This drive towards excellence is of no surprise given the high practical value this type of algorithms have and the added value they can yield for businesses and consumers alike. This is especially true for aspect- level sentiment analysis, where sentiment scores are assigned, not to a document or sentence, but to the various characteristics, or aspects, of the entity un- der consideration. Such a fine-grained analysis of, for instance, products or services, can provide many useful insights into consumer thinking. The majority of the algorithms for aspect-level sentiment analysis is centered around the use of a machine learning classifier and involves tasks such as feature construction and parameter estimation. While the past has shown that this kind of algorithms have strong performance, other directions are also explored (Schouten and Frasincar, 2016). One of these is the concept-driven approach (Cambria et al., 2015), which functions at the semantic level, built upon a layer of natural language processing compo- nents. Another direction, which is the research we describe here, is using discourse analysis to improve sentiment classification. Discourse analysis looks at how the various text segments interact with each other. For example, in “So even though this laptop is ugly, bulky and way overpriced, I still like it.”, the first half of the sen- tence gives an explanation for the second part of the sentence. As illustrated in the above example, know- ing which parts of the text are ancillary and which parts of the text form the core of the sentence can play a vital role in classifying sentiment. In this case, the sentiment for the laptop in general is pos- itive, even though all of the discussed aspects are negative. Rhetorical Structure Theory (RST) (Mann and Thompson, 1988) describes the various dis- course units and their relations, and multiple RST parsers …

Content truncated to 500 words