In recent years, various recommender system algorithms have emerged one after another, such as traditional matrix factorization (MF) techniques and deep learning (DL) techniques, which have greatly improved the performance of recommender systems. For fully mine user preferences and further study the important role of review texts, this paper presents a dual SVD++ recommendation model based on attentional convolutional neural network, which uses the user's rating data for items to mine the implicit feedback of users and items, through the convolutional neural network (CNN) of the attention mechanism is used to represent the important comment text features of users and items, Using dual SVD++, the learned review text features are fused to generate recommendations. The model uses mean square error (MSE) and mean absolute error (MAE) as evaluation indicators. On the Amazon public datasets, it is compared with five typical recommendation models, and the feasibility of the model is illustrated.