Adaptive Topic Tracking Research Based on Title Semantic Domain andDouble-state Model

Aiming at problems of sparse training corpora and topic excursion existing in topic detection and tracking, this paper examines twenty one most recent references and patents, proposes an adaptive topic tracking strategy based on title semantic domain topic model and double-state model. Title semantic domain topic model can enhance the title-centric semantic domain cohesion of reports, and reduce dimensions of reports' feature space effectively. The double-state strat- egy is a tracking technology based on the combination of static model and dynamic model: static model uses a given number of training reports to construct topic model which is the basis of topic tracking; dynamic model uses the sliding text window mechanism to capture new contents of a topic, remove outdated ones and reflect the changes of topic's focus in a timely manner. Experimental results show that the combination of double-state model tracking strategy and title se- mantic domain topic model can improve the performance of adaptive topic tracking system.