Study And Implementation On Opinion Leader Discovery And Typical Opinion Extraction

With the development of information technology, lots of online commnunication platforms have emerged. Forum has already been a main communication way, and an increasing number of netizens express opinions by it. During the dissemination of information, opinion leader, as an author or comment, plays an important role. Positive opinion leader group represents a group of netizens, whose opinions resonate well with followers, and make them express similar sentiment. Moreover, for each event, typical opinion owns rich emotions and plays a key role in analyzing public opinion. Based on above facts, this thesis focuses on detecting opinion leader from forum and mining typical opinion from comments.Firstly, as current opinion network model build researches pay more attention to explicit links but neglect implicit links, this thesis puts forword a new method for detecting explicit and implicit links. And these detected links can further be divided into positive and negative links. Then an opinion network and user network model can be built based on those links, and these models will be a solid basis for detecting opinion leader and extracting typical opinion.Secondly, based on above methods, opinion and user network can be built from web comment information, and then opinion leader can be detected by Dynamic Opinion Rank algorithm from Chinese news comments. After that, the most influential comment and user can be found by the methods of clustering user nodes, calculating scores and ranking.Thirdly, considering the importance of positive opinion leader group from multi-themes, this thesis builds a multi-themes network based on dynamic opinion leader detection. After traversing all comments, every user’s socre can be got, and top-k users with highest scores will be selected as positive opinion leader group.Fourthly, because of the importance of typical opinion in web public opinion analysis, this thesis divides comment set into opinion communities by sentiment word clustering, building sentiment phrase trees, and then finding typical opinion from society by proposed Longest Sequencial Sentiment Phrase Extraction algorithm.Experiments show the proposed methods in this thesis can detect opinion leader, positive opinon leader group from forum perfectly, and extract typical opinion of each society accurately.