Incremental Topic Modeling for Scientific Trend Topics Extraction

Anonymous

Abstract: Caused by the exponential growth of scientific research, the number of scientific publications and reports, one of the most urgent and challenging tasks now is the early detection of trending topics. In this paper, we investigate recent topic modeling approaches to accurately extract trending topics at an early stage. The incremental training technique is suggested so that the model can operate on data in real-time. For validation, we propose a novel dataset that contains a collection of early-stage articles and a set of key collocations for each trend. The proposed metric estimates the delay in days when determining the trend, and the developed matching method suffices to calculate it automatically. The conducted experiments demonstrate that the topic model with regularization, namely ARTM, is superior to the base PLSA model. Apart from that, the best ARTM-based model is able to extract most of the labeled trends during the first year of their evolution.