Dynamic non-parametric joint sentiment topic mixture model

Abstract

The reviews in social media are produced continuously by a large and uncontrolled number of users. To capture the mixture of sentiment and topics simultaneously in reviews is still a challenging task. In this paper, we present a novel probabilistic model framework based on the non-parametric hierarchical Dirichlet process (HDP) topic model, called non-parametric joint sentiment topic mixture model (NJST), which adds a sentiment level to the HDP topic model and detects sentiment and topics simultaneously from reviews. Then considered the dynamic nature of social media data, we propose dynamic NJST (dNJST) which adds time decay dependencies of historical epochs to the current epochs. Compared with the existing sentiment topic mixture models which are based on latent Dirichlet allocation (LDA), the biggest difference of NJST and dNJST is that they can determine topic number automatically. We implement NJST and dNJST with online variational inference algorithms, and incorporate the sentiment priors of words into NJST and dNJST with HowNet lexicon. The experiment results in some Chinese social media dataset show that dNJST can effectively detect and track dynamic sentiment and topics.