Incorporating Semantic Metadata into Nonnegative Matrix Factorization Based Topic Modeling and its Application to Main Topic Extraction

Abstract

Topic modeling is a type of a statistical model that has been proven successful for tasks including discovering topics and their trends over time. In many applications, documents may be accompanied by metadata that is manually created by their authors to describe the semantic content of documents, e.g. titles and tags. A proper way of incorporating this metadata into topic modeling should improve its performance. In this paper, we examine the use of semantic metadata for nonnegative matrix factorization based topic modeling and its application to main topic extraction. Besides a simple onelevel learning hierarchy, we adapt a two-level learning hierarchy method for this task. Our experiments show that these methods improve interpretability or coherence scores of extracted topics. Moreover, the two-level learning hierarchy approach can achieve higher interpretability scores than the one-level learning hierarchy method.