Research On Chinese Rumor Detection Model Based On Graph Convolution

With the growing popularity of social media,people have changed the way they access information.The Internet has become a bridge for people to communicate with the outside world.However,the authenticity of many information on the Internet is dubious,and some news even intentionally mislead others.On the one hand,the spread of these false rumors will greatly reduce the credibility of the media,and social platforms full of false remarks will gradually lose the trust of the audience;on the other hand,the spread of false rumors may also cause very serious damage to people’s political and economic life.Therefore,the detection of false rumors is of great significance to individuals and society.Traditionally,this verification of authenticity has been carried out by experts and journalists by comparing published information with established facts and cross-checking with trusted alternative sources.However,on various social media platforms,rumors spread quickly and the categories are very diverse,which makes it difficult to achieve effective results in the detection of rumors on social media.Therefore,this thesis extracts text features,sentiment features,user features,propagation features and other auxiliary features from the rumor data from Weibo platform,and proposes a rumor detection model based on graph convolutional neural network.The main research results of this thesis are as follows.1.Visualized analysis of rumor and non-rumor data.This thesis builds a hierarchical attention model from words,Weibo posts or Weibo comments to rumors.The hierarchical attention model has certain interpretability and can extract and display the typical features of rumors in terms of emotional expression.This thesis analyzes the visualization results and finds that fake rumors and non-rumor data have obvious differences in sentiment,stance and semantics.This has important guiding significance for feature engineering and designing neural network models that are more suitable for rumor detection tasks.2.Feature engineering.By combining multiple sentiment dictionaries,this thesis obtains the sentiment category,sentiment polarity and sentiment score features of Weibo posts and Weibo comments.Other extracted features based on text content include statistical features such as emoticons,punctuation,personal pronouns,etc.User-based features include the user’s gender,number of followers,etc.Through experimental verification,these manually extracted statistical features improve the performance of the model to a certain extent.3.A rumor detection model is constructed to predict the rumor data.The graph convolution model can extract the rumor propagation features from the rumor data,and get a better representation.In order to obtain a richer data representation,this thesis not only constructs a text-based graph convolution model but also constructs a graph convolution model based on sentiment features to extract the propagation features of the rumor text itself and the propagation features of rumor sentiment,respectively.Through experimental verification,the extraction of rumor sentiment propagation features improves the performance of the model to a certain extent.