HDPsent: Incorporation of Latent Dirichlet Allocation for Aspect-Level Sentiment into Hierarchical Dirichlet Process-Based Topic Models

We address the problem of combining topic modeling with sentiment analysis within a generative model. While the Hierarchical Dirichlet Process (HDP) has seen recent widespread use for topic modeling alone, most current hybrid models for concurrent inference of sentiments and topics are not based on HDP. In this paper, we present HDPsent, a new model which incorporates Latent Dirichlet Allocation (LDA)-based sentiment learning into an HDP topic modeling framework. This model preserves the benefits of nonparametric Bayesian models for topic learning, while simultaneously learning latent sentiment aspects. It automatically generates different word distributions for each single sentiment polarity within each topic that has been learned. We present results using existing corpora consisting of multi-aspect hotel and restaurant reviews, and discuss ramifications and applications of such a model for product reviews that are intrinsically hierarchical.