Extracting and Classifying Geospatial Expressions Using Large Language Models: A Case Study of Urban Planning Documents

ABSTRACT Generative Artificial Intelligence (GenAI) and Large Language Models (LLMs) present new opportunities to advance text‐based analyses, particularly for identifying and disambiguating geospatial information expressed as natural language text. This study explores the capabilities of an LLM to identify, classify, and spatially query geospatial expressions found in urban planning documents. Urban planning encodes regulatory information, domain expertise, and local knowledge, much of which is geographic in nature, in reports and plans. Interpreting this geographic information is challenging for professionals (e.g., planners and lawyers) and community members. This paper presents an automated pipeline developed to assess an LLM's ability to extract geospatial expressions, determine their relevant spatial scale (micro, meso, and macro), and to identify urban features referenced by a specific development proposal. Three types of planning documents were examined from a mid‐sized city (Kitchener, Ontario, and Canada), namely a City‐authored Official Plan (OP), private consultants' development proposal reports, and transcripts from public meetings. A hierarchical Retrieval‐Augmented Generation (RAG) model processed text to extract explicit and implicit geospatial expressions, and a multi‐agent architecture converted these expressions into SQL queries against municipal data layers. Performance varied systematically by expression type and document type. The model handled implicit expressions best in the public‐meeting transcripts but performed poorly on explicit expressions in property‐specific reports. On a paired comparison, the pipeline outperformed two established geoparsing tools (GeospaCy and Mordecai3) on locally‐grounded planning documents. While this work illustrates how LLMs could automate interpretation of urban planning documents, the approach could also be applied to other application contexts where geospatial references are encoded in text forms.