Research On False Information Detection Based On Multimodal Event Memory Networ

With the rapid development of mobile Internet,information technology and social platforms,social media is being used more and more widely.While providing convenience to users,social media also provides fertile soil for the breeding and spread of false information,resulting in the spread of false information faster and more far-reaching than ever before.At present,the detection of false information on social media is mainly regarded as a classification task,and the detection methods are mainly divided into two types: one is single-mode method,and the other is multi-mode method.Single mode method mainly extracts single feature from text or image.The multi-modal approach mainly studies the relationship between multiple modes,such as text features,image features and transmission structure,and combines multiple features from different angles to form the final multi-modal representation.Existing single-mode false information detection methods focus on mining features in text content and user profile,and do not make full use of the global semantic relationship of text content.On the other hand,the existing multi-modal methods focus on the simple concatenation of text and image features,but do not consider the interaction between different modes and the rich semantic information behind the text.In addition,most false information detection models are not effective in predicting emerging events.In view of the above two problems,this thesis mainly puts forward two different solutions.(1)To solve the problem of single mode method,a false information detection model based on metapath graph attention network is proposed.The text content and source posts based on false information are constructed into posting-word-user graph structure,and then the graph structure is decomposed into posting-word subgraph and posting-user subgraph based on posting-word and posting-user path,and the subgraph attention network is used to learn the representation of nodes.Finally,an attention mechanism is introduced to fuse the final representation of nodes in the false information detection subgraph and make the corresponding prediction.(2)In order to solve the problem of multi-modal method,a false information detection model based on multi-modal event memory network is proposed,which is mainly composed of multi-modal feature extractor,event memory network and false information detector.The multi-modal feature extractor extracts the corresponding Text and image features through text-CNN and VGG-19,captures the relationship between different modes by using the attention mechanism,and then obtains the rich deep semantic knowledge in the text through BERT to form the final multi-modal joint representation.The event memory network mainly extracts the invariant features of each event to enhance the detection ability of new events in the model.Finally,a combination of multi-modal features and shared features between events is input into the false information detector to predict the truth and falsity of each event.In this paper,a large number of experiments are carried out on the public data set.The experimental results show that the accuracy of the proposed method is improved compared with other single-mode and multi-mode baseline methods,and the generalization ability of the model is improved.