Generative Topic Model Using Variational Bayesian Inference

Abstract

In this thesis, we will propose a new model which is a continuous extension of the LDA and discuss about some applications using this model. The newly proposed model is called Continuous Semantic Topic Embedding Model(CSTEM) based on the Latent Dirichlet Allocation model with continuous assumption for the word-topic distribution. This assumption leads to the introduction of new parameter for the probabilistic model playing role as a global parameter which reflects how likely a certain word occurs in a document regardless of the topic variable. We will verify our model from the various points of view and insist that this model outperforms other topic models and is worth to use for some appropriate purposes. The verifications will be done via experiments using various corpora. And we will show that this model can be applied to other applications such as an analysis of the time-dependent topic model, which is helpful to analyze the trend of the topics over time intuitively.