Exploiting rhetorical structures to improve feature-based sentiment analysis

Sentiment analysis is an interesting application in natural language processing, aiming at identifying emotional expressions attached to speeches or texts. In this paper, simple yet effective strategies to extract feature-based segments and combine sentiment scores were studied. The strategies exploit textual structures to improve the segmentation quality. Each relevant set of segmented texts is subsequently passed to a lexical-based sentiment classification to obtain the polarity of a product feature. By using textual structures, the proposed strategies can improve accuracy of the sentiment classification. Especially, the accuracy on feature reviews with negation terms is improved by 86.4%. Moreover, for positive feature reviews, the strategies perform reasonably well up to 0.765 on average, in terms of f-measure.