Abstract: Aimed at the issue of knowledge islands between water conservancy project data texts, this paper presents a dam safety knowledge graph construction method based on the UIE framework and the SE-RE-Joint model. We use a seven-step method to construct a multi-layer domain ontology library of main structure, affiliated buildings, and comprehensive evaluation. After data preprocessing, fine-tuning of small samples is used to realize preliminary entity recognition. We complete entity recognition and relationship extraction through semantic enhanced coding, relying on the bidirectional GRU to capture sequence dependence and CRF to optimize label sequence consistency. And, a Neo4j graph database is used to store knowledge. Test results show that the UIE+SE-RE-Joint model extracts 25,000 entities and 31,000 relational pairs, and achieves a prediction accuracy higher than 84%. The four types of errors－entity conflicts, boundary ambiguities, type confusions, and relationship misjudgments－all are lowered significantly relative to the comparison model. The map supports knowledge visualization and data retrieval; Through Python-based implementation, it completes knowledge dynamic updating, knowledge reasoning, and risk positioning, assisting the intelligent diagnosis in dam safety management.

Key words: dam safety, knowledge graph, engineering data text, semantic enhancement, deep learning model, knowledge extraction