The inter-sentence relation in a document is characterized by complex contextual information, large span of correlation and many kinds of relations, leading to the poor effect of sentence-level relation extraction models when addressing inter-sentence relations. Graph networks have been widely used in the research of document-level relation extraction due to their advantages in modeling local structural features and long-distance context dependencies. However, most previous studies modeled document in a coarse-grained manner, which ignores the richness and otherness of hierarchical features in a document. Consequently, contextual information modeling is not sufficient and fails to participate in deep reasoning efficiently. In this paper, we propose a document-level relation extraction model based on the Hierarchical Dependency Tree and Bridge Path (HDT-BP). The model uses sentence as a unit to independently extract the fine-grained features of each hierarchy and reconstructs the chain-structured document based on multiple dependent relationships into a hierarchical dependency tree. Moreover, the relational bridge entity is introduced during relation extraction to improve the model performance by modeling the bridge path feature. Experimental results demonstrate that our model exhibits superior performance on the DocRED dataset and achieves a significant improvement in extracting relational facts that never appeared in the training set. Extensive additional experiments further verify the effectiveness of our model.