Collaborative filtering algorithm widely used in the recommendation systems has the problems of sparseness and cold start. For this reason, a recommendation model based on deep neural networks and dynamic collaborative filtering is proposed in this study. The model combines a pre-trained BERT model with bidirectional GRU to extract hidden feature vectors from users and commodity reviews. Furthermore, coupled CNN is used to construct the score prediction matrix and the temporal changes in user interests are incorporated through dynamic collaborative filtering. Finally, the experiments on an Amazon’s data set show that the proposed model increases the accuracy of commodity score prediction.