In order to solve problems for the disappearance of gradients, the deficiency of text feature and the mismatch of extracting phrase features in attention mechanism during the training of neural network in text classification, a new method base on dense-pool connection and phrase attention mechanism is proposed. Firstly, the method is used to extracting features while alleviating the gradient disappearance problem through the residual network and reuse important features through dense pooling connection. Then, the phrase attention mechanism is used to solve the problem of phrase dimension mismatch in the traditional attention mechanism. Finally, the results show that the accuracy of the model can achieve 92.7% in the AG news dataset for all variants. In addition, the convergence and classification accuracy of three comparison models are analyzed in different hyperparameters. It is concluded that the improved model can effectively alleviate the disappearance of gradients and solve the problem of phrase feature extraction, thereby improving the classification accuracy.