We propose a novel document generation process based on hierarchical latent tree models (HLTMs) learned from data. An HLTM has a layer of observed word variables at the bottom and multiple layers of latent variables on top. For each document, the generative process rst samples values for the latent variables layer by layer via logic sampling, then draws relative frequencies for the words conditioned on the values of the latent variables, and nally generates words for the document using the relative word frequencies. The motivation for this work is to take word counts into consideration with HLTMs. In comparison with LDA-based hierarchical document generation processes, the new process achieves drastically better model t with much fewer parameters. It also yields more meaningful topics and topic hierarchies. It is the new state- of-the-art for the hierarchical topic detection.
展开▼