TLDR: This research paper introduces new methods for geocoding compositional location references, such as “midway between Winsford and Crewe,” which traditional tools struggle with. The authors found that Large Language Models (LLMs) are better at geospatial reasoning over provided coordinates than generating accurate coordinates from memory. They propose a “Geoparser-augmented” strategy where a traditional geoparser provides coordinates for mentioned locations, and an LLM then reasons to predict a bounding box for the complex reference. This approach achieves state-of-the-art performance, with smaller fine-tuned LLMs matching larger models, offering practical solutions for grounding locations not found in standard databases.
Geocoding, the process of linking a location reference in text to an actual geographic location, is crucial for many analyses of unstructured data, from disaster response to disease surveillance. While traditional geocoding tools excel at identifying explicit location names like “Gaza City,” they often struggle with more complex, compositional descriptions—phrases like “about midway between Winsford and Crewe.” These types of references are common in fields that lack formal names for specific geographic areas, such as environmental impact statements.
A recent research paper, “Coordinates from Context: Using LLMs to Ground Complex Location References,” delves into this challenging area. Authored by Tessa Masis and Brendan O’Connor from the University of Massachusetts Amherst, the paper explores how Large Language Models (LLMs) can be effectively used to understand and pinpoint these intricate location descriptions. You can read the full paper here: Coordinates from Context: Using LLMs to Ground Complex Location References.
The core challenge lies in the fact that compositional locations often don’t exist in standard geographic databases. Existing tools typically link to point coordinates, which are insufficient for representing larger or more complex areas. This paper introduces a novel approach using “bounding boxes”—rectangular areas defined by two latitude and two longitude coordinates—to ground these locations more precisely. Bounding boxes offer a balance between simple point coordinates and complex polygons, making them suitable for LLM processing.
LLMs: Knowledge vs. Reasoning
A key insight from the research is the distinction between an LLM’s geospatial knowledge and its geospatial reasoning abilities. The authors conducted isolated evaluations and found that off-the-shelf LLMs, while possessing some geospatial knowledge, struggle more with generating accurate geographic coordinates from memory than with reasoning over provided geospatial data. For instance, LLMs performed significantly better when asked to reason about a location’s bounding box given the coordinates of nearby mentioned locations, compared to generating coordinates for a named location from scratch. This suggests that LLMs are better at processing and manipulating spatial information when it’s explicitly given, rather than recalling precise coordinates from their internal knowledge base.
The Geoparser-augmented Approach
Based on these findings, Masis and O’Connor propose an “Geoparser-augmented” strategy. This method combines a traditional geoparsing tool with an LLM. Here’s how it works:
- First, a traditional geoparser acts as a “recaller,” identifying explicit locations mentioned within a compositional description (e.g., “Winsford” and “Crewe”) and providing their precise center coordinates.
- Then, an LLM functions as a “reasoner,” taking the original compositional description and the coordinates of the mentioned locations to generate the bounding box for the described, unnamed location (e.g., “midway between Winsford and Crewe”).
This approach leverages the strengths of both components: the geoparser provides accurate, structured geospatial knowledge, while the LLM applies its reasoning capabilities to interpret the complex relationships described in the text. The researchers evaluated this method using both off-the-shelf and fine-tuned LLMs, including models from the Llama and Qwen families.
Also Read:
- Bridging Vision and Text for Better Geometric Reasoning in AI
- Unlocking Entity Understanding in Large Language Models
Performance and Practical Insights
The Geoparser-augmented approach generally outperformed a “Direct” approach, where the LLM had to rely solely on the natural language description without external coordinate input. It also showed significant improvements over prior work, including methods that used grammar-based semantic parsers. Notably, a relatively small fine-tuned LLM (Qwen 14B) achieved comparable performance to much larger off-the-shelf models, demonstrating the efficiency of their strategy.
Interestingly, the study observed that the Direct approach often resulted in higher area precision (smaller, more precise predicted areas), while the Geoparser-augmented approach yielded higher area recall (larger, less precise predicted areas). This was attributed to LLMs in the augmented setting sometimes over-relying on the provided coordinates of mentioned locations, using them directly as bounding box limits even when they extended beyond the actual described area.
For practical applications, the authors suggest using the Geoparser-augmented approach with a smaller fine-tuned LLM or a medium off-the-shelf model for the best overall performance, especially if prioritizing area recall or minimizing distance errors. If a geoparser isn’t available or if area precision is the main goal, they recommend the Direct approach with a large off-the-shelf model, as this relies more on the LLM’s inherent parametric knowledge.
This research marks a significant step forward in geocoding complex location references, offering robust, end-to-end methods that leverage the unique strengths of LLMs in geospatial reasoning, particularly when augmented with accurate external knowledge.
