While distant supervision enables automatic labeling of large-scale data, mislabeling is inevitable. Previous work has focused more on the relationship reduction (False Positive, FP) of mislabeling, but little work has explored the missing relationship (False Negative, FN) caused by an incomplete knowledge base. It is statistically proven that the dataset composed by distant supervision contains a large number of FN cases. One approach proposes a two-stage Pipeline model of “relationship first, entity later”, which can effectively alleviate the FN problem. However, this model cannot effectively utilize the intrinsic connections and dependencies between the two subtasks. Utilizing the relationship information may reduce the extraction of semantically irrelevant entities in the later entity extraction module, thus alleviating the extraction of redundant triples and improving the performance of the model. In order to solve these problems and utilize good relationship information, we present a joint relation extraction model that uses the message passing mechanism to feature fusion. First, the subject-object entity correspondence module generates a two-dimensional matrix containing the correspondence scores between all entities. At the same time, the model predicts all relationships that may exist in the sentence. Then, entity extraction is performed on these relationships, with the special feature that the input vectors to the entity extraction module are sentence vectors that have been fused with relationship information features through a message passing mechanism. This allows the model to focus on selecting entities with high relevance to this relationship in the entity extraction module. Finally, the correctness of the extraction triples is verified by the results of the subject-object entity correspondence module. Experiments show that our model achieves a state-of-the-art (SOTA) F1 score on the distantly supervised dataset NYT10-HRL and provides comparable results on dataset NYT11-HRL.