Network Public Opinion Analysis And Its Application Research Based On Topic Model

With the deepening of the Internet plus strategy implementation,network public opinion has become a hot spot in the vanguard of the battlefield to grasp the current network public opinion.Based on the study on the development of network public opinion and topic model,we proposed two topic model:Automatic Label Selected Topic Model(ALS-TM)and Automatic Time Selected Dynamic Topic Model(ATS-DTM),respectively from the viewpoints of both static and dynamic.ALS-TM and ATS-DTM can capture the current direction of network public opinion hotspot and its evolution.The first part,we proposed ALS-TM to improve the average topic coherence.From a static perspective,ALS-TM combine the average topic coherence and K-Means algorithm,to find the optimal classification labels of documents.We show that ALS-TM strengthen the extraction efficiency of the topic.The experimental results show that:firstly,the adaptive label constraints will help decrease the perplexity.Under a number of different topics,the perplexity is significantly lower than LDA and CTM.With the increase of number of topics,perplexity of ALS-TM declines gradually.Secondly,the adaptive label constraints helps improving the average topic coherence.Topic words and semantic coherence matrix readability generated by ALS-TM is stronger than LDA,which shows that ALS-TM can better express the core topics.Thirdly,the adaptive label constraints helps promoting independence of topics.ALS-TM shows more average difference between topics,which indicate independence of topics in ALS-TM is higher than LDA and CTM.The second part,we proposed ATS-DTM which is powerful to analysis topic evolution.From a dynamic perspective,ATS-DTM combined with time windows similarity index and ALS-TM,divided the best observation time windows,analysis public opinion evolution trend and detect abrupt public opinion.The experimental results show that:firstly,the difference among the network public opinion in different time windows is obvious.Major of time windows show higher average topic coherence.Secondly,network public opinion usually continues 15?30 days.In early June the hot topic,such as "House prices","purchase a house","income",get more attention,and these topics decline in early July.And duration of these topics is about 30 days.Thirdly,the network public opinion about emergency can be effectively detected.At the beginning of August,”811 Exchange Rate Reform" event is accurately captured and reflected,which core word is "exchange rate""devaluation","RMB".