Rumor Detection And Intent Classification In Social Media

With the rapid development of Internet technology,social media has gradually become an inseparable and important part of people’s daily work,life,and study.However,while bringing convenience to people’s information exchange,these platforms also provide a breeding ground for online rumors.Facing the large number of online rumors,traditional manual monitoring and detecting approaches are becoming more and more insufficient,while they not only consume a lot of manpower and time,but also receive the results that may not be satisfactory.In addition,with the types and contents of rumors being diverse,it is necessary to bring forward a more effective and accurate system for detecting rumors which may have bad influences on the society.This research takes microblog data as an entry point,draws on the classification rules of rumors intent in journalism and communication,and divides rumors’ intents into three categories,namely malicious,well-intentioned and no purpose,designs and implements social media-oriented rumor detection and intent classification algorithms based on data mining,data enhancement and deep neural network methods,and finally build a rumor detection and intent classification system.The main work of this paper is listed as follows:(1)Construction and pattern analysis of Weibo Rumor data sets.Based on rumors crawled from the Weibo Untruthful Information Disclosure Platform,we first annotated the intent category and user category,then integrated nonrumor data from multiple public Weibo data sets to construct the rumor detection and intent classification data set.After that,relevant pattern analysis was performed on the data sets,including production of the word cloud,text length analysis,user gender analysis,user category analysis and etc.(2)Research of data augmentation methods on the rumor data set.According to the pattern analysis of the original data set.We found that the number of well-intentioned rumors data and purposeless rumors data are far less than malicious ones.In order to prevent over-fitting problems caused by unbalanced data distribution,this paper applied back translation and Easy Data Augmentation algorithm to augment the original data set and obtain multiple augmented data sets.Afterwards,some comparison experiments on several machine learning based methods and deep learning based methods were conducted to verify the effectiveness of our proposed data augmentation methods.(3)Propose a rumor detection and intent classification model and conduct relevant experiments.In the process of feature extraction and analysis,we find that there is a clear dependancy between the category of the user who posted the rumor(individual user or self-media user)and the intent category of the rumor.Therefore,based on Word2Vec word vectors and bi-directional LSTM model,combining with user category information,we build a rumor detection and intent classification model.The experiment results showed that the detection accuracy of the proposed model is more than 94.9%.(4)Establish rumor detection and intent recognition system.Based on the above work,this paper completed the construction of a rumor detection and intent classification system,and carried out some tests on system functionalities and performances.The whole system can be divided into data processing module,prediction module and application module.In the prediction module,RDIC with or without user category features are deployed.Users can access the system through the browser,and conduct autonomous detection.