Abstract This paper presents a text feature extraction model based on stacked variational autoencoder (SVAE). A noise reduction mechanism is designed for variational autoencoder in input layer of text feature extraction to reduce noise interference and improve robustness and feature discrimination of the model. Three kinds of deep SVAE network architectures are constructed to improve ability of representing learning to mine feature intension in depth. Experiments are carried out in several aspects, including comparative analysis of text feature extraction model, sparse performance, parameter selection and stacking. Results show that text feature extraction model of SVAE has good performance and effect. The highest accuracy of SVAE models of Fudan and Reuters datasets is 13.50% and 8.96% higher than that of PCA, respectively.