Initialization in Gibbs Sampling Implementation of LDA

Latent Dirichlet Allocation (LDA) is a text mining technique used for automatic extraction of topics addressed in document collections. Sampling-based or variational inference algorithms are used to approximate the posterior in LDA. In this study, the effect of random initialization in Gibbs Sampling implementation of LDA on model coherence is investigated. For this purpose, topic models using the same parameter values are initialized with random and fixed values, and model coherence scores obtained are compared with each other.