Social media generates a large amount of information,only a small portion of which can be labeled by professionals as true or false rumors.To make full use of the vast amount of unlabeled data and detect false rumors in a timely manner,proposes a model called MSGCN based on multi-level structure and semi supervised learning.This model constructs a multi-level detection module based on graph convolutional neural net work to train limited labeled samples to extract multi-level propagation structure features,diffusion structure features,and global structure features.By perturbing the random model and integrating the dynamic output of unlabeled data for consistent prediction,the complementary pseudo label method is used to label the high confidence unlabeled data calculated by the model and add it to the training set to expand the sample.Under supervised cross-entropy loss and unsupervised consistency loss constraints,the model shows excellent performance.The experimental results on public Twitterl5,Twitterl6,and Weibo datasets show that the proposed model achieves accuracy of 88.3%,90.1%and 95.5%under 30%labeled samples,can achieve excellent performance with a small number of labeled samples.