Do users rate or review?

Abstract

Current approaches for contextual sentiment lexicon construction in phrase-level sentiment analysis assume that the numerical star rating of a review represents the overall sentiment orientation of the review text. Although widely adopted, we find through user rating analysis that this is not necessarily true. In this paper, we attempt to bridge the gap between phrase-level and review/document-level sentiment analysis by leveraging the results given by review-level sentiment classification to boost phrase-level sentiment polarity labeling in contextual sentiment lexicon construction tasks, using a novel constrained convex optimization framework. Experimental results on both English and Chinese reviews show that our framework improves the precision of sentiment polarity labeling by up to 5.6%, which is a significant improvement from current approaches.