Large Language Models–Driven Construction of a Spatial-Narrative Knowledge Graph for Beijing’s Central Axis

Abstract

The current state of cultural heritage data is characterized by fragmented resources and weak interconnectivity. Efficient integration, systematic organization, and in-depth interpretation of massive, multi-source, and heterogeneous data have become the core challenges in the digital protection and inheritance of cultural heritages. The "theme-juxtaposition" structure emphasized by spatial narrative theory is highly compatible with the discrete distribution characteristics of cultural heritage elements along the Beijing Central Axis. Based on this theoretical framework, this study constructs a Beijing Central Axis ontology model that integrates metadata space, Geo narrative space, historical narrative space, and cultural narrative space. In the knowledge graph construction phase, the category system and relationship design of the ontology model are used as few-shot prompts. The Qwen3 series of large language models are employed to systematically mine the metadata information and historical event associations of the Central Axis through four stages: data extraction, relationship definition, similarity relationship calculation, and relationship normalization. The experimental results show that in the information extraction task, the overall average precision and F1 score reached 0.75 and 0.52, respectively. However, when dealing with complex relationships of cultural heritages, especially in the extraction of directions and events, the average recall rate was relatively low at only 0.41, indicating that there is still room for optimization in the model performance.