Research On Fake News Detection Based On Siamese Network

With the development of online media platforms,people are becoming more and more accustomed to browsing fragmented news information on the Internet.However,in addition to the state media,the existence of personal media players and various content farms on the Internet has created a mixed phenomenon of true news and fake news.Although online media platforms can rely on methods such as censorship teams or users' reports to limit false news,the large numbers of news sent all the time cannot be fully reviewed.Therefore,it is of great significance to study the fake news detection method which can reduce the burden of the platform,promote the purification of the network environment and do not rely heavily on human resources.Aiming at the problem of the large amount of fake news and the difficulty of rapid identification by manpower,this thesis proposes rapid detection based on headlines and fake news detection method based on Siamese networks,and conducts experimental verification.Aiming at the rapid detection of fake news,a fast detection method based on the headline is proposed.Due to the short length and sparse features of the headline,a single model is onesided for feature extraction of the headline.Therefore,after comparing the advantages and disadvantages of several models,this thesis combines the Long Short-Term Memory network and the one-dimensional convolutional neural network with the attention mechanism,taking into account the advantages of global and local feature extraction.Experiment results show that the headline-based model can detect headlines with typical features well.However,since fake news is not limited to this category,it still needs to be further analyzed in conjunction with the body.In order to solve the complicated and confusing problem of fake news,a fake news detection algorithm based on Siamese network is proposed.Among them,unlike the traditional detection algorithm,we make full use of the existing information after the encoding layer,and splicing the output vector after word embedding layer with the output vector of the encoding layer replaces the complex feature extraction structure in the traditional model.It can not only make the model more concise and effective,but also reduce the number of parameters and increase the speed of calculation.In addition,the weight sharing between the two subnetworks in the Siamese Network can not only reduce the amount of parameters that need to be processed,but also and reduce the complexity of the model.Moreover,vectors of different spatial dimensions can be mapped to the same dimension,so that the data distribution of the headline and the body is consistent.In the attentive pooling layer,we use the bidirectional attention mechanism to interact between the headline and the body to strengthen the more relevant parts of the body and the headline.In the prediction layer,we explored a variety of combinations of the headline and body to find the optimal combination.Experiments performed on FNC,the public dataset in the field of fake news detection,shows that the effectiveness of the fake news detection model based on Siamese Network,and it can balance speed and accuracy.