Abstract
Dynamic Stacked Topic Model (DSTM) proposed here is a topic model, for analyzing the hierarchical structure and the time evolution of topics in document collections. Such document collections as news articles and scientific papers are framed hierarchical. In newspaper, for instance, an article related to the soccer is published in the sports section and that related to the election in the politics section. Furthermore, both topics and sections naturally evolve with a certain timescale. In the proposed model, to capture correlations between topics and the time sequence of topics in sections, a section is modeled as a multinomial distribution over topics based on the previous topic distribution as well as a topic assumed to be generated based on the word distribution of previous epoch. The inference and parameter estimation processes can be achieved by a stochastic EM algorithm, in which the maximum a posteriori estimation of hyperparameters and the collapsed Gibbs sampling of latent topics and sections are alternately executed. Exploring real documents also described demonstrates the effectiveness of the proposed model.
