Found in Translation: Using Large Language Models to Resolve Metes and Bounds Descriptions to Modern Addresses

Daniel Moulton, Federal Reserve Bank of Philadelphia
Robyn Smith, Federal Reserve Bank of Philadelphia
Larry Santucci, Federal Reserve Bank of Philadelphia

Historical property deeds frequently describe parcels using metes and bounds—written boundary descriptions referencing physical features, distances, and directions—rather than standardized addresses. This presents a significant challenge for researchers seeking to locate these properties. Using Philadelphia property deeds from 1910-1965, we present a pipeline leveraging large language models (LLMs) to translate metes and bounds descriptions into street addresses by matching them to publicly available parcel data. Our approach uses an LLM to structure metes and bounds descriptions into a format where we can find the corresponding parcel in modern GIS data. First, we prompt an LLM to extract structured location information from the deed: the fronting street, the intersecting street, distances from the starting point, etc. Next, we transform the City of Philadelphia’s Office of Property Assessment public dataset so we can locate each modern parcel in relation to the street intersections defining the city block in which the parcel resides. Finally, we use the structured information produced by the LLM to find the beginning point of the property—constrained to follow real-world intersections and streets. This allows us to find the parcel which best matches the metes and bounds description. Deeds employ a wide range of language to describe identical real-world beginning points which is compounded by inconsistent abbreviations, OCR errors, innate document errors, and errors in the public parcel data. Rules-based approaches struggle to balance recall and precision across linguistic diversity, while LLMs handle such variation more robustly. On a subset of Philadelphia deeds containing both a metes and bounds description and an address, our LLM-assisted method correctly maps ~98% of properties. We present a georeferenced set of nearly 8,000 Philadelphia property deeds with restrictions on the ownership, use, or occupancy as well as preliminary analysis into the geospatial distribution of these restrictions.

No extended abstract or paper available

 Presented in Session 207. Advancing Spatial Methods