Collective Extraction for Opinion Targets and Opinion Words from Online Reviews

Abstract

Online reviews are very important for lots of Web applications. Extracting opinion targets and opinion words from online reviews is one of the core works for review analysis and mining. The traditional extraction methods mainly include two categories: the pipeline-based methods and the propagation-based ones. The former extracts opinion targets and opinion words separately, which ignores the opinion relations between them. The latter extracts opinion targets and opinion words iteratively by exploiting the nearest-neighbor rules or syntactic patterns, which would probably lead to poor results due to the limitations on predefined window size and the propagating errors of dependency relation parsing. Due to such shortcomings of traditional methods, we propose a collective extraction method for opinion targets and opinion words based on the word alignment model, which especially adopts the concept of Classification to simultaneously extract opinion targets and corresponding opinion words. In order to tackle the time-consuming and error-prone problem of manual annotation, we further devise a semi-supervised extraction method based on active learning. Finally, we carry out a series of experiments on real-world datasets to validate the effectiveness of the proposed methods.