Topic Modeling for Segment-based Documents

Statistical topic models have traditionally assumed that a document is an indivisible unit for the generative process, which may not be appropriate to handle documents that are relatively long and show an explicit multi-topic structure. In this paper we describe a generative model that exploits a given decomposition of documents in smaller, topically cohesive text units, or segments. The key-idea is to introduce a new variable in the generative process to model the document segments in order to relate the word generation not only to the topics but also to the segments. Moreover, the topic latent variable is directly associated to the segments, rather than to the document as a whole. Experimental results have shown the significance of the proposed model and its better support for the document clustering task compared to other existing generative models.