LLM-Reflex-GeoKG: A reflexion-enhanced LLM framework for automated geographic knowledge graph construction

Area of Science:

• Geoinformatics
• Artificial Intelligence
• Data Science

Background:

• Automated construction of domain-specific Geographic Knowledge Graphs (GeoKGs) is crucial for intelligent applications.
• Challenges include inaccurate knowledge extraction and Large Language Models' (LLMs) limitations with geospatial semantics.

Purpose of the Study:

• To propose and validate a novel framework, LLM-Reflex-GeoKG, for accurate and complete GeoKG construction.
• To address LLM limitations in geospatial knowledge extraction.

Main Methods:

• Integration of a Reflexion-style self-reflection loop.
• Utilizing a collaborative dual-LLM generator-critic scheme.
• Employing a multi-stage knowledge extraction strategy.

Main Results:

• Achieved F1 scores of 0.898 for geographic entity recognition and 0.823 for relation extraction on Yangtze River Delta data.
• Significantly outperformed baseline systems in accuracy and completeness.

Conclusions:

• The LLM-Reflex-GeoKG framework enhances automated knowledge acquisition for GeoKGs.
• Reduces reliance on manual annotation, offering a practical solution for reliable GeoKG development.