<scp>LLM</scp> ‐ <scp>GeoTextCog</scp> : A Cognitive Enhancement Framework for Geospatial Scene Understanding in Geographic Texts Using Large Language Models

ABSTRACT Geographic texts (geo‐texts) play an indispensable role in preserving and transmitting geospatial knowledge. The spatial scene understanding capability of large language models (LLMs) in processing textual information is primarily demonstrated in two key steps: the accurate identification of geographical entities and spatial relationships, and the ability to perform relative position reasoning and spatial scene reconstruction. However, current LLMs are primarily limited to extracting surface‐level information and establishing shallow semantic connections from text, which restricts their capacity for deeper spatial cognition and reasoning. Inspired by the mechanisms underlying human mental map and spatial reasoning, this paper proposes a cognitive enhancement framework LLM‐GeoTextCog. It aims to achieve a transformation from descriptive text to a structured “spatial scene sketch”. The development of this framework involves two key methodological efforts: (1) The first effort focuses on enhancing LLMs' recognition capabilities for geographical entities and spatial relationships. To achieve this, two specialized datasets are constructed: GeoBase, a domain knowledge fine‐tuning dataset, and GeoRel, a scene cognition fine‐tuning dataset. Through a two‐stage LoRA fine‐tuning process, a dedicated geo‐texts cognition model, Qwen3‐8B‐GeoMine, is developed. (2) The second effort is focused on transforming spatial scene reconstruction into a practical 2D sketch generation capability for LLMs. This is realized through a novel paradigm of expert‐guided prompt engineering. The process follows three key steps: core entity anchoring, local coordinate system construction, and spatial relationship mapping, which together convert fragmented textual information into structured and interpretable geographical spatial scenes. Experimental results demonstrate that: (1) Through testing on the GeoRel dataset, Qwen3‐8B‐GeoMine achieves an F 1 score improvement of 9% to 13% in geographical entity recognition and spatial relationship extraction compared to the baseline model; (2) Through expert evaluation testing, the 2D scene sketch generation evaluation is conducted on the sample dataset. The results show improvements of 46.0% in core entity recognition accuracy and over 50% in spatial relationship reasoning accuracy. It enables both large‐scale regional spatial association reconstruction across cities and small‐scale restoration of entity spatial layouts within cities, demonstrating substantial potential for advancing geospatial intelligence in LLMs.