Sentiment-based topic model for mining usability issues and failures with user products

This paper describes an approach to problem phrase extraction from texts that contain user experience with products. During the last decades, consumer products have grown in complexity, and consumer dissatisfaction is increasingly caused by usability problems, in addition to problems with technical failures. Moreover, user reviews from online resources, that describe actual difficulties with products experienced by users, affect on other people's purpose decisions. In this paper, we present two probabilistic graphical models which aim to extract problems with products. We modify Latent Dirichlet Allocation (LDA) to incorporate information about problem phrases with words’ sentiment polarities (negative, neutral or positive). The proposed models learn a distribution over words, associated with topics, both sentiment and problem labels. Topic models were evaluated on reviews of different domains collected from online consumer review platforms. The algorithms achieve a better performance in comparison to several state-of-the-art models in terms of the likelihood of a held-out test and in terms of an accuracy of classification results. Qualitative analysis of the topics discovered using the proposed models indicates the utility of considering sentiment information in users’ critical feedback. Our contribution is that incorporating sentiment and problem information about words with reviews’ topics by the model's asymmetric priors gives an improvement for problem phrase extraction from user reviews published in English and Russian.