Generalized Knowledge Distillation for Topic Models

Kohei Watanabe, Koji Eguchi

Abstract: Topic modeling is used in the analysis of textual data to estimate the underlying topics within the dataset. Knowledge distillation has been attracting attention as a means of transferring knowledge from a large teacher model to a small student model in the field of deep learning. Knowledge distillation can be categorized into three types depending on the type of knowledge to be distilled: response-based, feature-based, and relation-based. To the best of our knowledge, previous studies on knowledge distillation used in topic models have all focused on response and/or feature knowledge, but these methods cannot transfer the structural knowledge of the teacher model to the student model. To solve this problem, we propose a generalized knowledge-distillation method that combines all three types of knowledge distillation, including the relation-based knowledge distillation with contrastive learning, which had not been used for neural topic models. Our experiments show that our neural topic model, trained with the proposed method, improves topic coherence compared to baseline models without knowledge distillation.