Companion context dependent topic modeling

Contexts such as time, location and companion play an important role to determine topics of documents. We focus on companion context (friends, wife, husband etc.) as one of the most important contexts to determine topics. We propose companion context dependent topic model by the extension of basic graphical model: LDA (Latent Dirichlet allocation). We propose three kinds of LDA extensions. Firstly, we incorporate context class to extract context dominant topics (CTM). Secondly, in addition to CTM, we incorporate predefined contextual words to make context dominant topics more discriminative (eCTM). Thirdly, in addition to eCTM, we incorporate switch variables (sCTM), which discriminate background words from contextual and topical words to associate context related words to the topics more precisely. We conduct experiments on two data sets, and they show that the proposed model (eCTM and sCTM) can predict the word distribution of test documents with high probability and generate discriminative sets of topics than baseline CTM from the viewpoint of perplexity.