The model uses the BERT to extract the semantic feature representation of the short text, inputs the semantic features into the Bi-GRU and extracts the semantic information with contextual timing features. The model feeds the features into the Maxpooling layer to filter the optimal features and classify them to get the category of the short text. A correction algorithm is added to mitigate the performance degradation for the momentum bias generated by the Adam algorithm in the fitting. The Adam algorithm is improved by comparing the corrected momentum values at two consecutive time steps and selecting the maximum value of momentum in the two time steps to substitute into the gradient calculation. The improved Adam algorithm adds an adaptive adjustment factor to the learning rate and uses the gradient value of the previous iteration to achieve adaptive adjustment of the learning rate and improve the classification accuracy. Experiments show that the classification accuracy of DTSCF-Net is 94.86%, which is 2.07% and 1.71% higher than that of the benchmark model BERT and BERT-Bi-GRU respectively in the same experimental environment. The results demonstrate that the proposed method in this paper has certain performance improvement.