The singular value decomposition-based anchor word selection method for separable nonnegative matrix factorization

Abstract

One of the recent methods for the topic modeling is separable nonnegative matrix factorization (SNMF). In general, SNMF consists of three main steps, which are, generating a word co-occurrence matrix, selecting anchor words, and recovering a topic matrix. The anchor words strongly influence the interpretability of extracted topics. In this paper, we propose a new method for selecting the anchor words by using singular value decomposition (SVD). We assume that the most dominant words in each latent semantics created by SVD are the potential candidates for the anchor words. Our simulations show that the SVD-based anchor word selection method can reach better interpretability scores of extracted topics than the common convex hull-based method on two of three datasets.