A semantic-template mapping approach for intelligent thematic map generation

Large Language Models (LLMs) present significant potential for intelligent cartography by enabling the generation of thematic maps through natural language interaction. However, applying LLMs to professional cartography faces a logical mismatch: the open-ended probabilistic reasoning of models often conflicts with the rigorous constraints required for valid cartography. To address this gap, this study proposes a Semantic-Template Mapping approach. This method explicitly decouples the probabilistic semantic reasoning of the LLM from the deterministic cartographic execution, leveraging the model’s semantic intelligence while ensuring engineering robustness. Specifically, we first define a four-dimensional template representation for natural-language-driven thematic map generation, comprising map type, application scenario, thematic semantics, and visual density. We then implement a hybrid intention parsing mechanism that maps unstructured instructions to specific configurations of this template. Finally, a rule-embedded component-based renderer synthesizes visuals from cartographic specifications derived from the selected template configuration. We verify the effectiveness of the proposed method through both case studies and quantitative evaluation. Experimental results on benchmark datasets demonstrate a maximum intention parsing accuracy of 90.8%. Importantly, the system maintains a 100% cartographic validity rate. This study provides a technical pathway for intelligent cartography, effectively balancing interaction flexibility with engineering reliability.