Improving Topic Quality with Interactive Beta-Liouville Mixture Allocation Model

One of the major tasks in natural language processing is to categorize texts into different categories. Topic models are an important set of tools for categorizing texts and so are mixture models since both models learn patterns from data in an unsupervised manner. The introduction of latent Dirichlet allocation (LDA) triggered a lot of research in this domain. Recent research investigates the use of distributions other than Dirichlet for the topic proportions in LDA especially generalized Dirichlet and Beta-Liouville distributions in addition to adding useful attributes specific to the task at hand. Improving the quality of topics extracted from these models is important for accurate inference and unsupervised language tasks. Owing to this cause, in this paper, we propose interactive Beta-Liouville mixture allocation (iBLMA) model which combines the clustering capabilities of mixture models with interactive learning which helps the user modify the topic weights of irrelevant words within the topic. We show the efficiency of our model with experiments on two different text datasets.