USNA: A Dual-Classifier Approach to Contextual Sentiment Analysis

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 390–394, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics USNA: A Dual-Classifier Approach to Contextual Sentiment Analysis Ganesh Harihara and Eugene Yang and Nathanael Chambers United States Naval Academy Annapolis, MD 21401, USA nchamber@usna.edu Abstract This paper describes a dual-classifier ap- proach to contextual sentiment analysis at the SemEval-2013 Task 2. Contextual analysis of polarity focuses on a word or phrase, rather than the broader task of identifying the senti- ment of an entire text. The Task 2 definition includes target word spans that range in size from a single word to entire sentences. How- ever, the context of a single word is depen- dent on the word’s surrounding syntax, while a phrase contains most of the polarity within it- self. We thus describe separate treatment with two independent classifiers, outperforming the accuracy of a single classifier. Our system ranked 6th out of 19 teams on SMS message classification, and 8th of 23 on twitter data. We also show a surprising result that a very small amount of word context is needed for high-performance polarity extraction. 1 Introduction A variety of approaches to sentiment analysis have been proposed in the literature. Early work sought to identify the general sentiment of entire documents, but a recent shift to social media has provided a large quantity of publicly available data, and private orga- nizations are increasingly interested in how a pop- ulation “feels” toward its products. Identifying the polarity of language toward a particular topic, how- ever, no longer requires identifying the sentiment of an entire text, but rather the contextual sentiment surrounding a target phrase. Identifying the polarity of text toward a phrase is significantly different from a sentence’s overall po- larity, as seen in this example from the SemEval- 2013 Task 2 (Wilson et al., 2013) training set: I had a severe nosebleed last night. I think my iPad caused it as I was browsing for a few hours on it. Anyhow, its stopped, which is good. An ideal sentiment classifier would classify this text as overall positive (the nosebleed stopped!), but this short snippet actually contains three types of po- larity (positive, negative, and neutral). The middle sentence about the iPad is not positive, but neutral. The word ‘nosebleed’ has a very negative polarity in this context, and the phrase ‘its stopped’ is posi- tive. Someone interested in specific health concerns, such as nosebleeds, needs a contextual classifier to identify the desired polarity in this context. This example also illustrates how phrases of dif- ferent sizes require unique handling. Single token phrases, such as ‘nosebleed’, are highly dependent on the surrounding context for its polarity. How- ever, the polarity of the middle iPad sentence is con- tained within the phrase itself. The surrounding con- text is not as important. This paper thus proposes a dual-classifier that trains two separate classifiers, one for single words, and another for …

Content truncated to 500 words