<p>The data sparsity problem constrains traditional recommendation methods, and the recommendation methods based on review texts effectively alleviate the sparsity of rating data by capturing user profiles and item attibutes from review texts. However, many of them ignore the fact that there exists a wealth of correlation information between items. This correlation information may help the recommender system understand the similarity between items and accordingly recommend to users items they may be interested in, which helps to improve the accuracy and diversity of recommendations. Based on the idea raised above, we propose a deep learning model based on semantic information and correlation between items (SMCI), which extracts rich semantic information from review texts and converts it into vector representation. This representation not only incorporates sentiment and topic information in reviews but also captures the correlation between users and items. In order to fully incorporate the correlation features between items, SMCI maps items’ similar representations in vector space based on users’ purchase sequences and predicts final ratings using the scaled dot product’s multi-head self-attention mechanism. Comparative experiments are conducted on seven public datasets. The experimental results show that the proposed model outperforms the baselines in terms of performance, which verifies the validity of our model.</p>