LATENT DIRICHLET LEARNING FOR HIERARCHICAL SEGMENTATION

Abstract

Topic model can be established by using Dirichlet distributions as the prior model to characterize latent topics in natural language. However, topics in real-world stream data are non-stationary. Training a reliable topic model is a challenging study. Further, the usage of words in different paragraphs within a document is varied due to different composition styles. This study presents a hierarchical segmentation model by compensating the heterogeneous topics in stream level and the heterogeneous words in document level. The topic similarity between sentences is calculated to form a beta prior for stream-level segmentation. This segmentation prior is adopted to group topic-coherent sentences into a document. For each pseudo-document, we incorporate a Markov chain to detect stylistic segments within a document. The words in a segment are generated by identical composition style. This new model is inferred by a variational Bayesian EM procedure. Experimental results show benefits by using the proposed model in terms of perplexity and F measure.