A hierarchical pachinko allocation model for social sentiment mining

Abstract

Existing topic models for mining sentiments from articles often ignores the fact that intra-topic correlations are common and useful to uncover a large number of fine-grained and tightly-coherent topics. This paper is concerned with the problem of social sentiment mining by modeling topic correlations. We aim to not only discover the connections between sentiments and topics, but also reveal the deeper relationship among topics where some topics may co-occur more frequently than others in articles. More specifically, we join sentiment mining with hierarchical pachinko allocation model to represent topic correlations by a hierarchy. In our model, the hierarchical pachinko allocation is employed to generate the latent hierarchical topic variables and sentiment variables. Experimental results on a collected news corpus show that our model can effectively identify latent topics in a hierarchical structure, and outperforms competing sentiment-topic models such as Latent Dirichlet Allocation based model in sentiment prediction.