Abstract
The automated construction of high-quality, domain-specific Geographic Knowledge Graphs (GeoKGs) is pivotal for intelligent applications, such as river network modelling, digital watershed management, and cultural heritage conservation in river basins. However, this task faces a core challenge of inaccurate and incomplete knowledge extraction. While Large Language Models (LLMs) offer a new paradigm, their direct application is often undermined by inherent hallucinations and a limited grasp of complex geospatial semantics. To address these challenges, this study proposes and validates a novel framework, LLM-Reflex-GeoKG, which integrates a Reflexion-style self-reflection loop with a collaborative dual-LLM generator-critic scheme and a multi-stage extraction strategy for GeoKG construction.Experiments on data from the Yangtze River Delta river network demonstrate that our framework achieves F1 scores of 0.898 and 0.823 on geographic entity recognition and relation extraction tasks, significantly outperforming baseline systems. This research confirms that our framework improves both the accuracy and completeness of automated knowledge acquisition while reducing reliance on extensive manual annotation, offering a practical paradigm for building reliable, domain-specific GeoKGs.