Coordinates from Context: Using LLMs to Ground Complex Location References

TL;DR

This paper tackles grounding compositional location references by ground-truthing bounding boxes through a two-part system: a recaller that maps mentioned places to coordinates and a reasoner that predicts the final bounding box $b_l$ from a description $d_l$. It demonstrates that LLMs possess stronger geospatial reasoning than knowledge, and introduces a geoparsed-augmented end-to-end approach that leverages external coordinates to improve grounding, achieving state-of-the-art results on the GeoCoDe dataset with relatively small fine-tuned models. Bounding boxes provide a practical grounding for locations not present in geographic databases, enabling meaningful area-overlap metrics and more robust grounding of large-scale regions. Limitations include evaluation on a single, English-language dataset and reliance on third-party geoparsers, underscoring the need for broader, multilingual benchmarks and privacy-preserving tooling.

Abstract

Geocoding is the task of linking a location reference to an actual geographic location and is essential for many downstream analyses of unstructured text. In this paper, we explore the challenging setting of geocoding compositional location references. Building on recent work demonstrating LLMs' abilities to reason over geospatial data, we evaluate LLMs' geospatial knowledge versus reasoning skills relevant to our task. Based on these insights, we propose an LLM-based strategy for geocoding compositional location references. We show that our approach improves performance for the task and that a relatively small fine-tuned LLM can achieve comparable performance with much larger off-the-shelf models.

Figures (2)

• Figure 1: An illustrative example of our task. Given a compositional location description, we expect the model to predict the location's bounding box (defined by two pairs of latitude-longitude coordinates).
• Figure 2: Illustrative examples of our LLM geospatial knowledge and reasoning baselines; all examples refer to the same location. Knowledge baseline: the LLM is given only the location's name. Reasoning baseline: the LLM is given only a description of the location and mentioned locations' coordinates (mentioned location names in red, for emphasis). Point refers to a center coordinate prediction (written as $\{ lat, lon \}$); box refers to a bounding box prediction (written as $\{ lon_{min}, lat_{min}, lon_{max}, lat_{max} \}$).