Dynamically Modeling Semantic Dependencies in Web Forum Threads

The huge amount of knowledge in web forums has motivated great research interests in recent years. However, tracking semantic dependencies in each thread in web forums has posed a challenging problem for researchers. In this paper, we explore an unsupervised topic model to burst through this issue by simultaneously modeling the semantics and the reply relationship in a thread. The proposed model is a dynamic extension of Latent Dirichlet Allocation (LDA) for the structure of web forum threads, where each post is considered as a mixture of topics that vary along the asynchronous conversation. The experimental results on two different forum data sets show encouraging performance of our proposed PPM in ranking the influence of posts.