Sentiment Polarity Classification Using Structural Features

Abstract

This work investigates the role of contrasting discourse relations signaled by cue phrases, together with phrase positional information, in predicting sentiment at the phrase level. Two domains of online reviews were chosen. The first domain is of nutritional supplement reviews, which are often poorly structured yet also allow certain simplifying assumptions to be made. The second domain is of hotel reviews, which have somewhat different characteristics. A corpus is built from these reviews, and manually tagged for polarity. We propose and evaluate a few new features that are realized through a lightweight method of discourse analysis, and use these features in a hybrid lexicon and machine learning based classifier. Our results show that these features may be used to obtain an improvement in classification accuracy compared to other traditional machine learning approaches.