Abstract: Automatically labeling data can reduce manual annotation costs in the process of distant supervised relation extraction generally, existing two problems, sentence label noise and long-tail relation distribution. To solve the problems, a relationship extraction method was proposed to fuse entity information from knowledge graphs and constraints between entities and relations. The proposed method was designed to encode the target entity, its neighboring entities' attributes, and to encode the neighboring graph formed by the target entity and its neighbors. Additionally, the constraints between entity types and relations were encoded, and all this information was integrated through a multi-source fusion attention module to construct a relationship extraction model. The AUC value of the method on the NYT-10 dataset is 0.524, with P@100 value of 94.8%. The long-tail metric Hits@K has improved compared to previous state-of-the-art models, emonstrating excellent performance and showcasing the effectiveness of the method's integration of entity information and constraint information to address the two main issues of DSRE.