Heterogeneous Graph Path Reasoning Based Document-Level Relation Extraction for Equipment Domains

Document-level relation extraction is one of the foundational tasks for constructing an equipment domain knowledge graph. However, existing document-level relation extraction models in the equipment domain face the following challenges: (1) Most employ implicit logic reasoning mechanisms with poor interpretability, lacking explicit modeling of reasoning processes, which results in suboptimal performance for extracting long-distance and complex relations. (2) They do not consider the issue of long-tailed relations caused by imbalanced label distributions. To address these challenges, this paper proposes a document-level relation extraction model based on heterogeneous graph path reasoning. For the long-distance and complex relations present in documents, the model constructs three types of reasoning paths: continuous paths, multi-hop paths, and default paths, to explicitly model the reasoning phenomena within documents. By integrating entity embeddings with the reasoning path features, the model enhances performance in extracting longdistance and complex relations. The adaptive focal loss function is employed to balance positive and negative relations, focusing on low-confidence samples to optimize the extraction performance of long-tail relations. Experimental results on both the public dataset DocRED and the self-constructed DLDED dataset demonstrate that the proposed model outperforms common baseline models such as TDGAT.