Detecting Rumor Based on Graph Convolution Network and Attention Mechanism
[Objective]This paper proposes a rumor detection method based on a graph convolutional network and attention mechanism,which utilizes comment forwarding and text semantic features.[Methods]Firstly,we analyzed the forwarding and replying relationship among comments and constructed a comment relationship feature map to explore the comments'correlations.Then,we used the BERT model to generate the vector representation of sentences based on their text semantic similarity.We also built the semantic feature map of comments by calculating the cosine similarity and fully extracting their semantic relevance.Third,we completed information dissemination among nodes based on a Graph Convolutional Network(GCN).We also used the attention mechanism to distinguish the impact of original and other comments on rumor detection.Finally,we obtained an accurate representation of the comment nodes.[Results]Our model's accuracy on the Twitter15 and Twitterl6 public datasets reached 0.860 and 0.870,with F1 mean values of 0.858 and 0.866.Compared with the BiGCN method,our model's accuracy improved by 5.1％and 1.5％on the Twitterl5 and Twitterl6 datasets,and the F1 mean improved by 5.0％and 1.9％,respectively.[Limitation]We only used texts for rumor detection.Future research will combine images,user attributes,and time attributes to improve the model's accuracy.[Conclusion]The proposed method can effectively improve the performance of rumor detection,providing valuable references for related tasks.
Graph Convolutional NetworksAttention MechanismRumor DetectionBERT Model
NETL
NSTL
万方数据
