Aiming at the problem that the feature information of review text is not fully captured and the aspect information extraction ability is poor in the traditional aspect-level sentiment classification model, an XLNet-CNN-GRU dual-channel aspect-level review text sentiment classification method is proposed. The static and dynamic word vectors generated by Glove and the XLNet model are stacked and embedded. It can carry more semantic information and effectively solve the problem of polysemy. In order to extract global and local features of the review text, the word vectors generated by stacking are input to the GRU and CNN dual-channel of the fused attention mechanism. The experimental results show that XLNET-CNN-GRU on the Restaurant, Laptop and ACL 14 Twitter datasets of SemEVAL-2014, compared with TD-LSTM, AE-LSTM, ATAE-LSTM, ATBL-MHMN, and ON-LSTM-SA, the accuracy of the model is improved by 2.04%, 2.3%, and 2.33%, the F1 value is improved by 2.18%, 1.94%, and 2.77%. Higher accuracy and F1 value are obtained, and the effect of aspect-level sentiment classification is improved.