A deep neural network model based on pre-training model (Roberta) and deep feature word vector was proposed to deal with the problem of Chinese text classification, in order to solve the problem that the word vector represented by the traditional model has unclear sequence, context, grammar, semantics and deep information representation. The sentence vector containing context semantics and grammar information and the word vector containing sentence structure features were generated by Roberta model. The word vector was extracted and fused by DPCNN model and revised gate recurrent unit（RGRU）to obtain the feature word vector containing deep structure and local information. The sentence vector and feature word vector were fused together to obtain a new vector. Finally, after the new vector passed through the softmax activation layer, the result was output. In the experimental results, F1 value, accuracy and recall were chosen as the evaluation criteria, they reached 98.41%, 98.44% and 98.41% in the long text of THUCNews. At the same time, the model had also achieved good results in short text classification.