Research On Multi-Modal Network Rumor Detection Model Based On Meta-learning

Rapid development in the Internet era has made posting and obtaining information easier,leading to a sharp increase in rumor numbers.Compared with the traditional text rumors,rumors with images are more deceptive,making sources and authenticity hard to verify.Therefore,these multi-modal Internet rumors are even more harmful.Detection of multi-modal rumors has become a new challenge.However,most existing methods are difficult to solve this problem and only adopt the standard concatenation for achieving feature fusion among different modes.Not only that,if there is a sudden rumor event,it is necessary to obtain new background knowledge from the event and add it to the existing rumor detection model.However,adding background knowledge from emergent events requires building a new model from scratch or continuing to fine-tune the model,which can be challenging for real-world environments and requires a lot of cost and human effort.To address the above issues,this thesis proposes two new multi-modal rumor detection models,BPDANN and Meta＿BPDANN.The BPDANN model uses bilinear pooling methods to fuse text features and image features.The model also adds an event classification module,which aims to remove event-specific features and maintain shared features between events.At the same time,on the basis of the BPDANN model,this thesis adopts the learning strategy of the multi-modal rumor detection task based on meta-learning,and trains a new multi-modal rumor detection model Meta＿BPDANN,so that when the model faces a new rumor detection task,it can converges after a small number of iterations,which improves the generalization ability of the model.This thesis conducts relevant experiments on two commonly used multi-modal rumor datasets,Weibo and Twitter.The experimental results exhibited that the two models in this thesis outperformed current the state-of-the-art methods.