Abstract: With the explosive growth of the submitted paper number, it is imperative to improve the satisfaction of recommendation system for paper review. To achieve the best match of reviewer and the paper is the key to obtain high quality papers. This paper proposes a BERT-GAT algorithm to solve the matching problem of reviewers and papers. First, based on the published papers of reviewers, keywords are extracted to form the research direction of experts and construct a bipartite graph. Then, the semantic information of the abstract or title of the paper is extracted by Bert, and the graph attention network (GAT) model is constructed based on the bipartite graph. Finally, the semantic information of the paper is fused with the graph attention network (GAT) to obtain the BERT-GAT model. Compared with other recommended algorithms, the BERT-GAT algorithm has achieved good results in each evaluation index on dataset of Power System Technology, which shows the effectiveness of the algorithm.