| With the development of social network,social media platform has become an indispensable part of people's daily life.The analysis of massive social behavior data created by users can provide new opportunities for studying human interaction and collective behavior.Often these behavioral data will be in the form of text.Social texts have the features of mass,short,multi-style,fast transmission and so on.The traditional method has gradually shown its drawbacks,we can not quickly find useful information accurately from the complex mass data.The topic detection and tracking technology is to identify the dynamic development of things from the topic level.It can be in the absence of human intervention to automatic process the text flow,identify the topic of the generation and changes,detection of popular hot topics.When using offline LDA model to process text information,each word in the document is used as a piece of data.For a large amount of text datas,the memory and time required for processing are very large.We use the online information processing to deal with larges of instant information by the online latent dirichlet distribution.The first time slice can be regarded as the offline latent dirichlet allocation model.The posterior probability of the previous is treated as the prior probabilistic probability of the current time slice.Simply load current time slice text processing information,need not be loaded on the entire data set.The factor graph and the belief propagation algorithm are combined to estimate the parameters,and an incremental algorithm is proposed for the online model,and the superiority of the method is proved by related experiments.The sliding window mechanism taking into account a text of a keyword in the vicinity of a large number of words into a sliding window.In this paper,a model suitable for short text processing is proposed by combining PTM and sliding window topic model.And the superiority of the model is proved by experiments. |
PDF Full Text Request