Finding Niche Topics using Semi-Supervised Topic Modeling via Word Embeddings

Topic modeling techniques generally focus on the discovery of the predominant thematic structures in text corpora. In contrast, a niche topic is made up of a small number of documents related to a common theme. Such a topic may have so few documents relative to the overall corpus size that it fails to be identified when using standard techniques. This paper proposes a new process, called Niche+, for finding these kinds of niche topics. It assumes interactions with a user who can provide a strictly limited level of supervision, which is subsequently employed in semi-supervised matrix factorization. Furthermore, word embeddings are used to provide additional weakly-labeled data. Experimental results show that documents in niche topics can be successfully identified using Niche+. These results are further supported via a use case that explores a real-world company email database.