Incorporating User Input with Topic Modeling

Topic models such as Latent Dirichlet Allocation (LDA) can discover topics from a large collection of documents in an unsupervised fashion and thus is one of the most popular text analysis tool currently in use. However, when using it in practice, the topics discovered by topic model don’t always make sense to end users. The poor quality topics will substantially undermine a topic model system’s usability. Due to the unsupervised nature of topic model, this is difficult to incorporate user’s domain knowledge or feedback to the topic model. In this paper, we introduce a novel constrained LDA model, named cLDA, that is capable of incorporating user inputs in the form of document pairwise constraints. Document pairwise constraints can be document must-links and document cannot-links which represent the semantic similarity of documents. The effectiveness of the proposed cLDA model is shown in several aspects on a benchmark dataset.