A Concise Representation for Detailed Sentiment Analysis

This paper presents initial results in sentiment analysis classification, as an attempt to go beyond categorizing texts only by ‘positive’ or ‘negative’ orientation, using fine-grained features for this purpose. We present a method for sentiment classification based on a concise representation built from analyzing appraisal groups such as “very good” or “not terrible”. An appraisal group is represented as a set of attribute values anteceding an appraisal word (adjective). An appraisal lexicon is used to identify adjectives guiding the analysis. We performed experiments classifying movie reviews in Spanish using features based upon attitude taxonomy information, and report improvements on precision with eight dimensional vectors and a support vector machine algorithm.