Generating a Context-Aware Sentiment Lexicon for Aspect-Based Product Review Mining

A great share of current sentiment analysis techniques is based on special purpose lexicons providing information about the semantic orientation (e.g. positive, negative, neutral) of its entries. Due to the high labor costs of manually assembling such resources, recent work has focused on automatically inducing the polarity of given terms. We follow this line of work while focusing on the domain of user-generated product reviews, a main field of application for sentiment analysis. In this domain, a major observation is that the semantic orientation of terms is often context-dependent which poses an additional challenge to the automatic construction of such lexicons (e.g. positive: “longbattery life” vs. negative: “long shutter lag time”). We propose a novel unsupervised method to induce a context-aware sentiment lexicon by utilizing the semi-structuredness of user-generated product reviews.