Constrained-hLDA for Topic Discovery in Chinese Microblogs

Abstract

Since microblog service became information provider on web scale, research on microblog has begun to focus more on its content mining. Most research on microblog context is often based on topic models, such as: Latent Dirichlet Allocation(LDA) and its variations. However,there are some challenges in previous research. On one hand, the number of topics is fixed as a priori, but in real world, it is input by the users. On the other hand, it ignores the hierarchical information of topics and cannot grow structurally as more data are observed. In this paper, we propose a semi-supervised hierarchical topic model, which aims to explore more reasonable topics in the data space by incorporating some constraints into the modeling process that are extracted automatically. The new method is denoted as constrained hierarchical Latent Dirichlet Allocation (constrained-hLDA). We conduct experiments on Sina microblog, and evaluate the performance in terms of clustering and empirical likelihood. The experimental results show that constrained-hLDA has a significant improvement on the interpretability, and its predictive ability is also better than that of hLDA.