Research On Topic Detection And Tracking Of News

With the development of online news,people can obtain more extensive information,but they also face the difficulty of quickly identifying hot topics,so the Topic Detection and Tracking technology have emerged.The topic model can detect interpretable topics.However,most topic models lack optimization for news and have shortcomings such as only considering single word co-occurrence information and insufficient semantic integration.In view of the above problems,the topic model based on the Graph Auto-Encoder(Topic GAE)and the topic tracking method based on the topic inference results are designed for the News Topic Detection and Tracking System.Topic GAE is used in the Topic Detection Module,which explicitly introduces fine-grained co-occurring word pairs and semantic features into document modeling to obtain richer topic-related information.In order to combine a variety of information to model the topic,Topic GAE takes the graph constructed by corpus as input.The graph takes words and documents as nodes,builds edges based on the documents' bag-of-words and fine-grained co-occurrence word pairs,and represents node features through word embedding.In order to infer topics from the graph,Topic GAE introduces the idea of graph auto-encoding,uses the encoder to extract the latent distributions,uses the decoder to reconstruct the word co-occurrence information,and updates the results through backpropagation.The topic tracking method based on the topic inference results is used in the Topic Tracking Module.The topic vector is calculated by combining topic words and word embeddings,and topic similarity is measured according to the cosine distance of the topic vector.Single-pass clustering between new topics and past topics is carried out with topic as granularity to realize incremental tracking of topics.Experiments on news datasets show that Topic GAE has state-of-the-art topic coherence and text clustering performance,especially on datasets with few documents.