Research On Text Sentiment Mining In Social Network Based On LDA Model

The SNS such as microblogging and BBS is proving increasingly popular among netizen for the freedom and interactive.The SNS has produced lots of opinion data like emotions and users' preferences.As propelled by the rapid growth of SNS data,it is urgent to utilize automated tools to monitor the user relationship and topic trend.It is inadvisable to directly utilize the methods which are traditionally used on text opinion mining,for SNS data is short,non-normative and semi-structured.In this research a model is proposed to aim specially at sentiment analysis on social network concerning text content and user link structure.The research focuses on topic-level sentiment analysis in social networks.For each user,the text information and link structure are given,and the goal is to assign a sentiment label for a specific topic to the user.(1)Analyze the users' behavior characteristics in social networks.Construct a panorama for dataset xiciEdu and sinaWeibo,and compare the difference.Analyze the post number distribution and friends' number distribution,and summarize the behavior characteristics.Analyze whether it is feasible to introduce homophily into social network analysis:Firstly,are connected users more likely to share opinions?Secondly,among connected users,are there more users with same opinion than those with different opinions?(2)Analyze the topic-level user influence in social networks.Introduce the user influence into sentiment analysis not only to utilize the structural information eftectively,but also to make up the content scarcity.Extract the relational network,then use quantity methods to evaluate user A's influence over user B on topic T from two aspects:user A's authority on T and 'interaction value' in T.Construct a influential user list UB for user B and introduce it into sentiment categorization as supplementary feature.The experiment results show that the topic-level user influence model is effective.(3)Propose a topic-level opinion mining model in social networks which is called SUSTM(social-user-sentiment topic model).Firstly,add a sentiment level into the LDA topic model to extend it to a four-level generative model:document,topic,sentiment and word,which can mine the topic and sentiment simultaneously.Then based on sentiment distribution ?,find out three most influential users for every user i,modify the sentiment distribution,and assign a sentiment label for user i.The experimental results of SUSTM,KNN and Backpropagation Neural Network show that the model proposed is effective.