Graph-based Entity Linking with Heterogeneous Attention for Technical Documents

Abstract

The entity linking in technical documents has specific issues, with many specialized terms, complex domain relationships, and the knowledge base often being outdated. This study proposes a high-end graph-based framework that combines heterogeneous graph attention mechanisms with deep context modeling. Using an adaptive relational reasoning approach, we combine information from the knowledge base and text to construct a comprehensive representation for stable entity disambiguation. Many experiments using large-scale, multi-domain datasets have shown that matching performance has significantly improved; compared to previous methods, accuracy and F1-score have increased by more than 5%, often exceeding 90%. The core components supporting the good results are graph attention, domain knowledge integration, and context integration; hyperparameter experiments determined stable and scalable settings. The system is general-purpose, maintaining good accuracy and completeness even in the absence of a vocabulary or ontology. This approach provides a stable and scalable foundation for the automatic entity resolution of engineering, scientific, and industrial documents. It has various applications in intelligent document analysis within knowledge management, information retrieval, and highly specialized technical fields.