Topic Modeling and Word Embeddings

Abstract

This paper discusses approaches for static topic modeling, in particular an improved method based on topic parametrization from a continuous distribution over the space of word embeddings. Word embeddings corpora proves to reflect semantic interdependences. Thus, we incorporate vectorized word representations trained with Word2Vec neural network in a generative process of topic modeling. The alternative approach with beta approximation of mutual information distribution over embeddings is proposed and compared with vanilla LDA and Gaussian LDA methods. Copyright ©2018 for the individual papers by the papers' authors.