Research On Topic Tracking Models Based On Bayesian Network

The rapid development of the Internet and the increase of Internet users lead to the exponential growth of network information.It is becoming more and more difficult to obtain news information that users really care about from complex information.Tracking the follow-up news information of users' concerns can effectively save users' query time,and at the same time,it can organize and classify news information according to topics,so topic tracking technology becomes particularly important.In the network environment with a huge amount of information,efficiently organizing news information is the biggest challenge for topic tracking.The existing probabilistic topic models mainly include Latent Dirichlet Allocation(LDA)topic models,language topic models,naive bayesian network topic models and belief network topic models.The above models have complex probabilistic derivation,and the storage of data in the derivation process taking up a lot of space during topic tracking.In order to solve this problem,referring to the idea of simple bayesian network retrieval model,two topic tracking models are proposed.The work of the thesis mainly includes two aspects:(1)For the problems of low tracking efficiency of the probabilistic topic model and large calculation amount of the derivation process,simple bayesian network retrieval model's related knowledge is used for topic tracking,and a Simple Bayesian Network Static Topic Tracking(SBNSTT)model is proposed,gives the model's topology,topic and story similarity calculation formula.The SBNSTT model is a directed graph with two levels of nodes including terms and topics.The arc in the figure indicates the index relationship between terms and topics.The similarity calculation of topic and story is converted into conditional probability deduction of topic and report.The algorithm of propagation+evaluation is used in probability deduction.Under the premise of ensuring accuracy,the inference process is simplified and the tracking efficiency is improved.(2)Considering that events are subclasses of topics,observing events is convenient for understanding various aspects of the topic,so a layer of event nodes is added to the SBNSTT model to build a Bayesian Network Static Topic Tracking(BNSTT)model.The BNSTT model is a directed graph with three layers of nodes including terms,events,and topics.The direction of the arc in the figure indicates the inclusion relationship of the three.Thesimilarity between the topic and the report is obtained by the similarity between the report and the event to be tested,and the similarity between the event and the topic.Verify the performance of the new model on the official TDT4 data set.The experimental results show that the DET curve of the SBNSTT model is below the vector space topic model,and the tracking performance is better.Compared with the SBNSTT model,the optimal detection error trade-off of the BNSTT model is reduced by 1.7%,and the tracking performance is further improved.