Segmented Topic Model

Segmented Topic Model is a new form of topic model which can take into account the inner structures in documents. The basic ideas are:

• As in LDA, one document d has a multinomial distribution v(d) over latent topics
• In this document, each segment d,s (sentence or paragraph) also has a multinomial distribution over topics. This distribution is generated from a two-parameter Poisson-Dirichlet process r(d,s)~ Poisson-Dirichlet(v(d),a,b)