A hybrid neural network based on XLNet for rumor detection

Abstract

When neural network is used to judge the authenticity of events, the original text and multiple user comments are important basis for detecting events. An event can be discribed by the word level features, sentence level features and long-term features in all user comments at the same time. A hybrid neural network model for rumor detection based on XLNet is proposed by integrating those features. Firstly, using XLNet as the language model can accurately describe the text information, save the long-term characteristics of events, and represent the same vocabulary according to the context. Secondly, the hybrid neural network can capture the temporal text features, retain the bidirectional text information according to the context, and save local features and global features at the same time. The attention mechanism in the network can increase the proportion of important features. Finally, the classification function is used to obtain the classification results. Relevant experiments were carried out on Weibo data set. The accuracy of hybrid neural network rumor detection model based on XLNet reached 93.56%, which proves the efficiency of using XLNet and hybrid network to detect rumors.