With the rapid development of the network, text categorization has become a key technology in processing and organizing large text. There are already many different types of neural networks applied to text categorization and achieve good results. However, most of the document models use only a small amount of features as input which don't take into account the sufficient amount of information. Considering enough characteristics, the dimensions of the disaster will occur, and resulting in a substantial increase training time, difficulty in training the model. This paper considers using deep belief networks to extract features from the text, then using softmax regression classifier for classification. The deep belief networks not only has a high learning ability, but also can extract highly distinguishable and low-dimensional features from the original high-dimensional features. So taking advantage of the deep belief networks to classify the text, can take into account enough information document amount. And the result shows that the use of deep belief networks achieve good performance for text categorization .