A Novel Hierarchical Topic Model for Horizontal Topic Expansion With Observed Label Information

Abstract

Hierarchical topic models, such as hierarchical Latent Dirichlet Allocation (hLDA)and its variations, can organize topics into a hierarchy automatically. On the other hand, there are lots of documents associated with hierarchical label information. Incorporating these information into the topic modeling process can help users to obtain a more reasonable hierarchical structure. However, after analyzing various real-world datasets, we find that these hierarchical labels are ambiguous and conflicting in some levels, which introduces error and restriction to the latent topic and the hierarchical structure exploration process. We call it the horizontal topic expansion problem. To address this problem, in this paper, we propose a novel hierarchical topic model named horizontal and vertical hierarchical topic model (HV-HTM), which aims to incorporate the observed hierarchical label information into the topic generation process, while keeping the flexibility of horizontal and vertical expansion of the hierarchical structure in the modeling process. We conduct experiments on BBC news and Yahoo! Answers datasets and evaluate the effectiveness of HV-HTM on three evaluation metrics. The experimental results show that HV-HTM has a significant improvement on topic modeling, compared to the state-of-the-art models, and it can also obtain a more interpretable hierarchical structure.