Construction of a reservoir flood control scheduling knowledge graph through co-creation between large language models and domain experts

Documents related to reservoir flood control scheduling are typically unstructured, posing significant challenges to effective management. Moreover, differences in understanding among practitioners hinder the accurate extraction of complex semantic relationships, thereby reducing scheduling efficiency. Knowledge graphs (KGs), as structured bases, can capture these relationships and support decision-making. Existing KG construction methods fall into two categories: schema-driven methods, which achieve high accuracy but suffer from limited completeness due to restricted schema coverage; and schema-free methods, which expand coverage by leveraging large language models (LLMs), yet tend to introduce noise owing to the lack of domain-specific constraints. To balance the accuracy and completeness of KG construction, this paper proposes the LLMs–Expert Collaborative Co-Creation through Exploration–Refinement Iteration (LECERI) method. The approach combines LLM autonomy with expert guidance, iteratively exploring and refining the schema. Candidate elements are automatically mined and then adjusted by experts for semantic consistency through iterative refinement. Experimental results show that the proposed LECERI method improves the accuracy and completeness of the resulting KG, achieving precision, recall, and F1-score of 0.52, 0.61, and 0.56, respectively. These results indicate the effectiveness of the collaborative strategy between LLMs and domain experts.