A novel classification method based on pseudo-relevance feedback (PFR) was proposed in order to solve the sparseness problems in short text classification. The short texts were expanded using the web pages which are similar to them in semantic level. The feature vector generation algorithm was modified to extract both the local features and the global features. The method can alleviate the sparseness problem of the final feature matrix, which is common in short text classification because of the limited length of the texts. The experimental results on an open dataset show that the method can significantly improve the short text classification effect compared with state-of-the-art methods.