Abstract: In order to enhance the ability to extract entity pair relationship in dialogues, this paper proposes a DEP-GAT model by introducing dependency relation into a heterogeneous graph attention network. Initially, the basic characteristics of each word are obtained through the preprocessing layer. Subsequently, in the discourse coding layer, context features are extracted and dependency information is added to further understand the speech structure. Eventually, a heterogeneous graph is constructed by utilizing the features, and an effective message passing mechanism is designed to enable the updated dialogue entity pairs to contain all the context information and grammatical features of the entire dialogue, thereby further enhancing the ability of the model to extract entity relations. The experimental results reveal that, on the DialogRE data set, the DEP-GAT model performs better than the baseline model does, with an increased F1 value of 2.9% in the development set and 1.8% in the test set respectively.