Social Media Rumor Detection With Knowledge Enhancement And Bi-Directional Graph Convolutional Networks

Social media has greatly improved the timeliness and convenience of people's access to new information.However,the unrestricted and lack of supervision of social media has also led to the emergence and spread of social media rumors.The problem of rumors brought by social media is becoming more and more serious,and finding rumors from such a large amount of information on social media has become a daunting challenge.Traditional social media rumor detection models use manual features to train a supervised classifier,but the manual features cannot accurately represent the rumor;Deep learning models can automatically extract features of the rumor,but these methods only extract features from the text content,dating that lacks external knowledge.And traditional deep learning models cannot effectively pay attention to the both deep propagation and wide dispersion of social media rumors.Based on the shortcomings of existing social media rumor detection methods,this paper proposes a social media rumor detection model based on knowledge-enhanced representation and bidirectional graph convolutional network.The main contributions of this paper include:1).In order to better represent the text of the social media rumors propagation node,this paper proposes a knowledge-enhanced text representation,which introduces knowledge graph as additional prior knowledge to improve text representation.First,the method uses the pretrained knowledge graph embedding model to embed entities in the text.Then,for each propagation node,integrate the entity representation and traditional text representation as to perform knowledge-enhanced representation on the text.2).Organize the social media rumor spreading network into two forms: top-down graph and bottom-up graph.This paper designs the top-down graph convolutional network to extract the propagation feature of rumor events,as well as the bottom-up graph convolutional network to extract the dispersion feature.The result of fusing the bidirectional graph convolution network better detects social media rumors.3).The upstream nodes in the rumors propagation network contain richer information,especially the root node contains the original text content.In this paper,the node enhancement method is proposed,which captures the hidden relationship between the root node and the other nodes through concatenating their features.And the method set different weights to different nodes,so that the model can effectively use the rich information contained in the upstream nodes.The comparison experiment results show that our proposed method outperforms several state-of-the-art models.And the results of detailed ablation experiments also prove that knowledge-enhanced representation,bidirectional graph convolutional network,and node enhancement all have the effect of improving rumor detection tasks.Finally,the early rumor detection task also shows that our proposed model can effectively detect rumors early.