Rumor Recognition System Based On Long Short-Term Memory Network And Deep Features

As a crucial channel to spread information,Weibo not only realizes the instant sharing,dissemination and interaction of information,but also becomes a place where the rumors come out.As the old saying goes,"rumors are fiercer than tigers",therefore,identifying and instantly clarifying rumors circulated in Weibo is of great significance in terms of the safety and stability of the entire society.Given the importance of rumor recognition in Weibo,this paper proposes a new methodology to increase precision to recognize rumors,which utilizes the three deep features that have not been used by previous scholars: rumor style,polarization of rumor reviews,as well as portrait matching of users end.Significant results have been achieved on the issue of rumor recognition.Specifically,the main work of this paper includes the following four points:1.Based on the Long Short-Term Memory,Historical-Gate Long Short-Term Memory improve the internal structure,on the basis of the original three gates structure,a fourth gate is added: History gate.This model uses self attention mechanism to update history gate information.History gate can store history information,so that the model can better capture the previous and subsequent information of a long input sequence.In the Weibo rumor data,GRU,LSTM and H-LSTM are used for experiments.Experiments show that compared with GRU and LSTM,H-LSTM reduces the number of iterations required for convergence by 800 and 400 steps respectively,and improves the accuracy by 4.32% and 2.17% respectively.2.Use two existing models to construct two deep features.This thesis using doc2 vec model to calculate the sentence vector D of the latest Weibo content of the user and the sentence vector M of all the historical Weibo content of the user,taking M as the user portrait,using the vector distance formula to calculate the distance S between D and M,S is the matching degree of the user domain portrait of the user's latest Weibo,and the size of S determines whether the user's latest Weibo content is the content that he often focus on or excel at.Using Snownlp emotion analysis library,this thesis analyzed the emotion of each comment on Weibo,counted the number of comments with bipolar emotion,and calculated the polarization degree of comment emotion.3.Using the one-way analysis of variance,the three deep features obtained are calculated to determine whether they have a significant effect on rumor recognition.The results show that the P-value of the three deep features are less than 0.01,indicating that the three deep features are important for rumors.The significant difference between the two types of data and non-rumored data proves the validity of the feature.The experimental data set also contains five shallow features,which are the number of fans,the number of Weibo posts,the number of retweets,the number of followers,and the gender.The gradient descent tree is used to select the features of three deep features and five shallow features,and calculate the importance of the features.The results show that three deep features of the eight features are ranked in the top three,which proves the importance of deep features for rumor recognition.4.Taking the three deep features as feature sets,this composite rumor recognition model is constructed by SVM,which realizes rumor recognition in Weibo.The experimental data set contains a total of 1538 rumors and 1849 non-rumors.The findings showed that the method in this paper has a good effect on rumor recognition in Weibo,the recognition accuracy rate reaches 0.9853,and the F1 score reaches 0.9839.The accuracy rate and F1 value are higher than all existing literature on rumor recognition.