<p>Document-level relation extraction (DocRE) aims to predict relations between entities in a document containing multiple sentences. In recent years, with the continuous development of relation extraction technologies, DocRE has gained widespread attention. However, current methods for DocRE still face the long-tail and multi-label problems. Most previous methods have focused on enhancing the feature representation of entity pairs by integrating richer contextual information within the document, however, they often overlook some critical aspects: semantic information pertaining to relations and latent associations between entity types and the relations, both of which are vital for capturing nuanced relational patterns. Against this backdrop, to explore the intrinsic connections between these potential features and thereby alleviate both the long-tail problem and the multi-label problem, this paper proposes a semantic interaction model based on prior data labels and entity-type correspondence, called PSGE. Firstly, we construct a relation-entity type integrated directed graph by leveraging relation labels and entity types derived from prior knowledge, and model an adjacency matrix based on the connections between entity types and relations. Then, we use Graph Attention Network (GAT) to aggregate relation-entity type features, supplement the semantic information of entities and enhance the relatedness of relations. Finally, experiments conducted on two widely used DocRE datasets (DocRED and DWIE) demonstrate that the proposed relation extraction model PSGE outperforms the baseline model ATLOP, achieving F1 score improvements of 0.44/1.94 on DocRED and DWIE.</p>