Abstract: Topic detection and tracking(TDT) is widely used. As one of research tasks for TDT, new event detection can provide prior knowledge to TDT, so it is of great theoretical research significance in the field of TDT. Because LDA model can not automatically identify new events, and the number of LDA topic is determined by the artificial, or by repeated experiments, it is of low efficiency. This paper presents a new method based on LDA and correlation of topic terms, which considers the correlation of subject terms and report time, it can dynamically adapt updated topics and then detect the new event. Experiment results demonstrate that this method is of some advantages and the sensitivity of new events detection is increased.

Key words: latent Dirichlet allocation(LDA), topic detection, new event detection, correlation of the topic terms