Jointly modeling multi-grain aspects and opinions for large-scale online review

To aggregate opinions on aspects of entities mentioned in large-scale online reviews, it is important to automatically extract aspects of different granularities, identify associated opinions, especially aspect-specific opinions, and classify sentiment polarity. Recently, various topic models are proposed to process some of these tasks, but there is little work available to do all simultaneously. In this paper, we propose a Joint Aspect-Based Sentiment Topic (JABST) model to jointly extracting multi-grain aspects and opinions, which addresses all the tasks mentioned above. JABST models aspect, opinion, sentiment polarity and granularity simultaneously. To better separate opinion and aspect words, we propose JABST-ME, in which a maximum entropy (ME) classifier is applied to extend JABST. We evaluated the models on reviews of electronic devices and restaurants qualitatively and quantitatively. The experimental results show that the proposed models outperform state-of-the-art baselines.