Abstract: Aiming at realizing the rumor detection on microblog, this paper deeply excavates the semantic information of the body content of microblog, and emphasizes the emotional tendency reflected by users in microblog comments, so as to improve the effect of rumor identification. In order to improve the rumor detection accuracy, based on XLNet word embedding method, the Transformer’s Encoder model is used to extract the semantic features of microblog body content. Combined with the BiLSTM+Attention network, the emotional feature extraction of microblog comments is realized. Two kinds of feature vectors are spliced and fused to further enrich the input features of neural network. Then, the microblog event classification results are output, and the microblog rumors detection is achieved. The experimental results show that the accuracy of the model in rumor recognition reaches 94.8%.

Key words: rumor detection, sentiment analysis, XLNet, Transformer model, deep learning