Abstract:

As the existing topic model can not accurately reflect the periodic variation and spatial distribution of topics in spatiotemporal context,a spatiotemporal contextual topic model for topic tracking is proposed according to the fact that the Internet information often contains the publishing time and site. In the investigation,first,by associating the distribution of subtopics with spatiotemporal context,a model is established to describe the cycle and strength of topics. Then,the parameters of the proposed model are estimated through EM algorithm,and are employed to obtain the snapshot and cycle of topics. Finally,the time-based topic similarity is calculated to estimate the subsequent topic information,thus realizing the topic tracking. The tracking experiments of food safety events indicate that,as compared with the traditional topic tracking method only depending on the text features,the proposed method can obviously improve the tracking efficiency of the topic as well as the tracking accuracy of subtopics. It is thus concluded that the proposed method helps to achieve more accurate topic retrieval.

Key words: topic model, context, generation model, probability distributions, text processing