Traditional approach for building text classifiers requires a large number of labeled documents for training a good text classifier. For reallife text classification applications, it is difficult to obtain a large number of labeled documents, so how to get a better result with these labeled and unlabeled documents has become a hot research topic. This paper proposes a new method to expand the set of labeled documents, first we extract a set of representative features from the labeled documents, then according to these representative features we choose the similar samples and add them to the labeled documents. The experiments prove that the method can effectively improve the classification results.