Research On Micro-blog Rumors Detection Based On Comments Sentiment And Ensemble Learning

Micro-blog,as one of the most popular social network applications,brings convenience to people,but there are also full of network rumors in microblog platform.According to the characteristic of micro-blog,such as the wealth of microblogging information,freedom and fasting of spread,micro-blog rumors spread indiscriminately on the platform,causing serious harm to individuals and the society.As the premise of research,monitoring and governance of the social internet rumors,automatic detection of rumors has caused more and more concern in the community and the relevant researchers.This thesis takes micro-blog rumors as the research object,analyzes the relevant text information and user information of micro-blogs,and extracts the deep and hidden features of micro-blogs.Combined with the idea of ensemble learning,a classification model was built based on the improved stacking ensemble method to complete the automatic detection of rumors.The main works of this thesis includes:(1)Deep features are extracted,which are based on rumor micro-blogs' text information and user information.The features of micro-blog rumors proposed in previous researches are relatively simple and superficial,lacking further analysis of relevant text and user information.This thesis analyzes the differences between rumor micro-blogs and normal micro-blogs in comments and user information,and extracts the deep recessive features,such as proportion of negative sentiment comments,user's credit value and whether there are authoritative users in the rumors dissemination process,and then proposes the quantitative methods of new features.(2)In order to calculate the value of micro-blogs' comments polarity feature more accurately,this thesis proposes a new sentiment classification method,which is based on machine learning and semantic rules.Firstly,it collects and organizes five lexicon resources to construct a more comprehensive polar emotion lexicon.Then,the whole text is divided hierarchically,and through analyzing and summarizing the expression form and sentence structure of micro-blog text,the semantic rules and calculation methods are defined to calculate the emotional polarity of text more precisely.Finally,a new sentiment classification method is proposed,which is combined the method of sentiment rules and machine learning.The semantic sentiment information,extracted by sentiment rules method,is used to expand the feature set of the machine learning classification method after being extended to semantic sentiment features.The new method solves the problems that the machine learning classification method ignores contextual semantic relations and the sentiment lexicon method includes few unregistered words to some extent.(3)The thesis designs a construction method of classification model for rumors detection based on CE-Stacking ensemble method.After analyzing the classification models used in previous researches,it is found that most of classification models constructed by classification algorithms are single,and has poor generalization performance,and the research on the construction of strong classification model is lacking.This thesis combines the thinking of ensemble learning,and improves the stacking ensemble method with feature of comments polarity,and constructs the ensemble classification model to improve the accuracy of rumors detection.Finally,we extract Sina micro-blog data for experimental verification.The results show that compared with previous research methods and models,the proposed sentiment classification method and rumors detection model have a certain improvement.