Temporal Topic Modeling in News Flow

This work is devoted to the task of Topic Detection and Tracking. The problem is principal for all kinds of news aggregators that display news in the timeline. The purpose of my research is to build Online topic model working with comparable quality to offline model. The objectives of the study included the review of existing approaches in this field, proposal of the more complex and better Online topic model and quality analysis of it. The paper presents a model that applies additive regularization of Probabilistic latent semantic analysis without additional assumptions on the dictionary. The algorithm is demonstrated on the Reuters corpus and corpus of the Russian texts assembled from the news sources and it works with comparable quality to offline model on both datasets. Also, tracking quality assessment methodology was proposed in the paper and applied to datasets.