A Rumor Events Detection Method Based on Deep Bidirectional GRU Neural Network

Traditional rumors detection methods often rely on statistical analysis to manually select features to construct classifiers. Not only is the message feature selection difficult, but the gap, between the representation space where the shallow statistical features of information exist and the representation space where the highly abstract features including semantics and emotion of information exist, is very big. Thus, the result of traditional classifiers based on the shallow or middle features is not so good. Due to this problem, a rumors deteciton method based on Deep Bidirectional Gated Recurrent Unit (D-Bi-GRU) is presented. To capture the evolution of group response information of microblog events over time, we consider the forward and backward sequences of microblog flow of group response information along time line simultaneously. The evolution representations of deep latent space including semantic and emotion learned by stack multi layers Bi-GRUs to rumor detection. Experimental results on a real world data set showed that rumor events detection by considering bidirectional sequence of group response information simultaneously can obtain a better performance, and stack multi-layers Bi-GRUs can better detect rumor events in microblog.