Sentiment Analysis Of Social Media Based On Probability Graph Model

With the rapid development of Web2.0,Social Media has developed rapidly.Unlike the traditional media,Social media is highly dynamic and complex.As an important information carrier in social media,text has an exponential growth trend and is highly dynamic and complex.How to efficiently detect its hidden information,(i.e.Sentiment Analysis on text),nowadays has become a hot research topic.Topic sentiment mixed models recently have attracted much attention due to their good performance and synchronization of analyzing topics and the sentiment in a document.Through investigation into literatures,we find that 1.existing models do not consider the time information hidden in the text and ignore the characteristic of the short text,which makes it impossible to realize the sentiment evolution analysis for the short text,such as microblog short text;2.Few models consider the user information on the social medial text or ignore the social r-elationship between users(especially the multiple social ralationship),which makes it impossible to detect the individual community and extract the topic and sentiment for the community.In order to solve the above problems,two new models are constructed:1.Dynamic Joint Sentiment and Topic model(DJST).This model combines the characteristics of the short text and incorporates the time information,(i.e.the next model's parameters depend on the current model's parameters or outputs)into the JST model to dynamically analyze the sentiment for the microblog text.The experiment on the real Microblog datasets shows the effectiveness of this model in the term of detecting the sentiment evolution.2.Multiple Social Relation Dynamic Topic Sentiment Model(MSRDTS).This Model integrates multiple social relations and time information into the Reverse-JST model,which can effectively extract individual community,analyzes the activities of each community and find the people with high participation.Furthermore,it can capture the topic and sentiment for each community.The experiment on the real Enron e-mail datasets verifies the modeling ability of this model.