ILDA: An interactive latent Dirichlet allocation model to improve topic quality

Abstract

User-generated content has been an increasingly important data source for analysing user interests in both industries and academic research. Since the proposal of the basic latent Dirichlet allocation (LDA) model, plenty of LDA variants have been developed to learn knowledge from unstructured user-generated contents. An intractable limitation for LDA and its variants is that low-quality topics whose meanings are confusing may be generated. To handle this problem, this article proposes an interactive strategy to generate high-quality topics with clear meanings by integrating subjective knowledge derived from human experts and objective knowledge learned by LDA. The proposed interactive latent Dirichlet allocation (iLDA) model develops deterministic and stochastic approaches to obtain subjective topic-word distribution from human experts, combines the subjective and objective topic-word distributions by a linear weighted-sum method, and provides the inference process to draw topics and words from a comprehensive topic-word distribution. The proposed model is a significant effort to integrate human knowledge with LDA-based models by interactive strategy. The experiments on two real-world corpora show that the proposed iLDA model can draw high-quality topics with the assistance of subjective knowledge from human experts. It is robust under various conditions and offers fundamental supports for the applications of LDA-based topic modelling.