Extracting opinionated (sub)features from a stream of product reviews using accumulated novelty and internal re-organization

Opinion stream mining extends conventional opinion mining by monitoring a stream of reviews and detecting changes in the attitude of people toward products. However, next to the opinions of people on concrete products, product features-on which people also bestow their opinions-are equally important: such features appear on all products of a given brand and can deliver clues to product vendors on what improvements should be done in the next version of a product. In this study, we propose an opinion stream mining framework that discovers implicit product features and assesses their polarity, while it also monitors features and their polarity as the stream evolves. An earlier version of this framework has been presented in Zimmermann et al. (2013). The extended framework encompasses an additional mechanism that merges clusters representing similar product features. We report on extensive experiments for both the original framework and the extended one, using two opinionated streams.