Neural Dynamic Focused Topic Model

Topic models and all their variants analyse text by learning meaningful representations through word co-occurrences. As pointed out previous work, such implicitly assume that the probability of a topic to be active its proportion within each document are positively correlated. This correlation can strongly detrimental in case documents created over time, simply because recent likely better described new hence rare topics. In this work we leverage advances neural variational inference present an alternative approach dynamic Focused Model. Indeed, develop model for evolution which exploits sequences Bernoulli random variables order track appearances topics, thereby decoupling activities from proportions. We evaluate our on three different datasets (the UN general debates, collection NeurIPS papers, ACL Anthology dataset) show it (i) outperforms state-of-the-art generalization tasks (ii) performs comparably them prediction tasks, while employing roughly same number parameters, converging about two times faster.