Historic natural history collections, maintained by museums throughout the world, contain millions of unique specimen records, representing an invaluable resource for scientific research and public inquiry. Yet, for many of these collections a significant proportion of records remain un-georeferenced, limiting their discoverability and usability. However, un-georeferenced records often contain text-based locality information and therefore have the potential to be georeferenced. Multiple applications and initiatives for supporting human-based georeferencing exist but the process of automated georeferencing at scale has remained somewhat elusive. This issue represents a major data bottleneck in the natural history space.

Here we present a novel approach to mass georeferencing for natural history collections. Our approach is built on a framework distinguishing the deterministic and semantic components of a full georeferencing pipeline. A hybrid methods model is applied where we utilize traditional data engineering and computational geography methods for deterministic tasks such as calculating spatial relationships, along with LLM-based methods for semantic steps such as locality string parsing and characterizing complex spatial language.

A key finding from our initial work, which has motivated our hybrid design, is that LLMs are excellent at parsing complex, highly varied text strings even when they include multi-part spatial relationships with complex, non-sequential structure. This parsing step is specifically the area which is most difficult for a fully code-based approach. However, when used for the complete georeferencing process LLMs produce coordinate and uncertainty estimates that are highly variable, unreproducible, not consistently accurate, and often systematically biased. This latter finding, demonstrated across multiple frontier models (and reported by Xie et al. 2025), makes a strong case against a fully LLM-based georeferencing approach. Therefore our pipeline is developed based on a hybrid model where LLMs handle only semantic tasks and code-based approaches leveraging validated spatial libraries are used for geometry calculations and similar tasks.

The pipeline operates on deduplicated locality strings and proceeds through several stages, each designed to extract as much information as possible through deterministic methods before passing anything to the language model. Locality strings range from simple named places to complex multi-part descriptions involving cardinal offsets, survey references, and chained spatial relationships. Administrative components and coordinate strings are stripped from locality strings prior to any LLM call, restricting the semantic task to only the geographic elements of the string and improving parse quality and consistency while reducing the potential for error. The LLM then performs a highly structured extraction of text string components, tagging and returning named places, survey references such as PLSS section and USGS quadrangle references, habitat descriptions, and unresolved spatial language in a schema-constrained output. Named places are resolved against open gazetteers scoped to the appropriate country and administrative region, with LLM-assisted disambiguation for cases where multiple viable candidates exist. A second LLM pass then translates the remaining spatial language into one or more spatial operations which are passed, along with previously tagged location information, to open-source spatial libraries for computation. The final result is structured to Darwin Core data standards including footprintWKT and coordinateUncertaintyInMeters.

A key design principle underlying the pipeline is our attention to spatial uncertainty, especially to address the false precision that is characteristic of a fully LLM-based approach. Rather than producing a single point estimate, the pipeline produces, where possible, a geometry representing the full spatial extent within which a collection event could have occurred, along with a minimum bounding circle represented by a radius. Different locality string types produce different geometry types, each encoding a different kind of spatial uncertainty in a consistent and transparent way. Our uncertainty approach is grounded in established georeferencing best practices (Chapman and Wieczorek 2020).

The pipeline is implemented in Python using open-source geospatial libraries and is under active development at the Smithsonian National Museum of Natural History in direct collaboration with department data managers and collections staff. We are also in the process of exploring the use of a locally-hosted LLM to carry out semantic tasks which will allow our pipeline to be used without reliance on proprietary frontier LLM models. Current test results on specimen datasets show geometry computed for a significant proportion of unique localities on a fully automated first pass, with the remainder flagged for structured human review.