💬 NLP

Coordinates from Context: Using LLMs to Ground Complex Location References

This paper proposes and evaluates an LLM-based strategy for geocoding complex, compositional location references, demonstrating that a relatively small fine-tuned model can achieve performance comparable to much larger off-the-shelf models by effectively leveraging geospatial reasoning.

Original authors: Tessa Masis, Brendan O'Connor

Published 2026-01-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tessa Masis, Brendan O'Connor

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are reading a travel diary that says, "The campsite is about halfway between the old oak tree and the red barn."

If you tried to find this spot on a standard GPS or Google Maps, you'd be stuck. Those tools are great at finding named places like "Central Park" or "Eiffel Tower," but they don't have a database entry for "the spot between the oak and the barn." This is the problem the paper tackles: Geocoding compositional locations. It's the art of turning a description of a relationship between places into actual coordinates on a map.

The authors, Tessa Masis and Brendan O'Connor, wanted to see if modern Large Language Models (LLMs)—the AI brains behind chatbots—could solve this puzzle. They discovered something surprising about how these AIs think, and they built a new system to fix the AI's weaknesses.

Here is the breakdown of their journey, using some everyday analogies:

1. The Two Skills: The Encyclopedia vs. The Detective

The researchers realized that solving this puzzle requires two different skills, and they tested the AI on them separately:

Skill A: The Encyclopedia (Geospatial Knowledge)
- The Test: "What are the coordinates of Paris, France?"
- The Result: The AI is okay at this, but not perfect. It's like a student who has read a lot of books but sometimes forgets the exact page number. When asked to guess the size of a city (a "bounding box") just from its name, the AI often guessed wildly off, sometimes by hundreds of kilometers.
Skill B: The Detective (Geospatial Reasoning)
- The Test: "Here are the coordinates for the Oak Tree and the Barn. The campsite is halfway between them. Where is it?"
- The Result: The AI is actually a great detective. When given the facts (the coordinates of the reference points), it can do the math and logic to figure out the new location very accurately.

The Big Insight: The AI is better at reasoning with facts provided to it than it is at recalling facts from its own memory.

2. The Solution: The "Tool-Augmented" Team

Since the AI is a great detective but a forgetful encyclopedia, the authors decided to give it a helper. They created a two-step team:

The Librarian (The Geoparser): This is a traditional, reliable tool (like a standard GPS API) that knows exactly where named places are. It looks up "Oak Tree" and "Red Barn" and hands the AI their exact coordinates.
The Detective (The LLM): The AI takes those coordinates and the description ("halfway between") and calculates the final location.

The Analogy: Imagine you are trying to find a hidden treasure.

The "Direct" approach is asking the AI to guess the location based only on a vague story. It often fails.
The "Geoparser-augmented" approach is giving the AI a map with the starting points marked, then asking it to draw the path to the treasure. This works much better.

3. The Secret Weapon: Bounding Boxes

Most GPS systems use a single dot (a center point) to mark a location. But what if the location is a whole field or a vague area like "the middle of the forest"? A single dot is too precise and often wrong.

The authors used Bounding Boxes instead. Think of this as drawing a rectangle around the area on a map.

Instead of saying "The treasure is at exactly 40.7° N," they say "The treasure is somewhere inside this rectangle."
This is a much more honest and accurate way to describe places that don't have a specific address.

4. The Results: Small AI vs. Big AI

The paper tested huge, expensive AI models against smaller, cheaper ones.

The Surprise: A relatively small, fine-tuned AI (trained specifically on this task) performed just as well as the massive, billion-dollar models.
The Takeaway: You don't need the biggest, most powerful computer to solve this. You just need the right strategy (the Librarian + Detective team) and a model that has been taught how to use it.

5. A Quirk in the System

The researchers noticed something funny: When they gave the AI the exact coordinates from the Librarian, the AI sometimes got too lazy. Instead of doing the math to find the "halfway" point, it would just grab the coordinates of the two reference points and draw a giant box around them.

Analogy: It's like asking a student, "What is 5 plus 5?" and the student just writes down "5 and 5" instead of "10."
They found that training the AI (fine-tuning) helped it stop doing this and actually start doing the math.

Summary

The paper argues that to find locations described in text (like "between X and Y"), we shouldn't just ask an AI to "know" the answer. Instead, we should:

Use a reliable tool to find the known places.
Let the AI use its reasoning skills to figure out the unknown place based on those known spots.
Represent the answer as a "box" on a map rather than a single dot.

This approach allows us to map places that have never been named before, using a mix of traditional tools and smart, reasoning AI.

Technical Summary: Coordinates from Context

Problem Definition

The paper addresses the challenge of geocoding compositional location references. While traditional geocoding links explicit place names (e.g., "Gaza City") to geographic coordinates, it often fails when text describes locations relationally without a specific name (e.g., "about midway between Winsford and Crewe"). Such descriptions are common in domains like environmental impact statements and food security outlooks. Existing tools struggle with these inputs because they rely on linking to pre-existing geographic databases (like GeoNames or OpenStreetMap), which rarely contain these composite entities. Furthermore, standard tools typically output point coordinates, which are insufficient for representing larger or irregular geographic areas.

Methodology

The authors propose a strategy that disentangles geospatial knowledge (retrieving facts about known locations) from geospatial reasoning (inferring new locations based on relational descriptions).

1. Task Formulation

The geocoding task is defined as a two-step process:

The Recaller ( $g$ ): Maps mentioned locations ( $m_i$ ) within a description to their geographical information ( $q_i$ ), such as center coordinates.
The Reasoner ( $f$ ): Takes the natural language description and the retrieved coordinates of mentioned locations to predict the bounding box (defined by $\{lon_{min}, lat_{min}, lon_{max}, lat_{max}\}$ ) of the target location.

2. Proposed Approach: Geoparser-Augmented

The authors propose an end-to-end strategy that leverages the strengths of both traditional tools and Large Language Models (LLMs):

Recaller: A traditional geoparsing tool (simulated as an oracle using OpenStreetMap data, or realistically using the Google Maps API) retrieves accurate coordinates for the locations mentioned in the text.
Reasoner: An LLM uses the text description and the retrieved coordinates to reason about the target location's position and generate a bounding box.

3. Baselines and Comparisons

The study evaluates several configurations:

Direct: An LLM attempts to generate the bounding box using only the text description, without external coordinate retrieval.
End-to-end LLM: The LLM acts as both the recaller (generating coordinates for mentioned locations) and the reasoner.
GBSP: A prior work using a grammar-based semantic parser that relies on polygon data for mentioned locations.

4. Evaluation Metrics

The authors utilize five metrics:

Coverage: Percentage of examples where a prediction is made.
Distance Error: Distance between predicted and ground-truth centroids.
Area Precision, Recall, and F1-score: Metrics measuring the overlap between the predicted and ground-truth bounding boxes.

Key Findings and Results

Geospatial Knowledge vs. Reasoning

Through isolated evaluations, the authors found that off-the-shelf LLMs struggle significantly more with generating accurate geospatial knowledge (predicting coordinates for known names) than with reasoning over provided geospatial data.

When asked to predict coordinates for known locations, models showed large distance errors (often >200km).
When provided with coordinates and asked to reason about a relational description, models performed much better, with distance errors dropping to <35km for top models.
Increasing model size improved knowledge retrieval more than reasoning capabilities.

Performance of Proposed Methods

Geoparser-Augmented vs. Direct: The Geoparser-augmented approach generally outperformed the Direct approach, particularly in Area Recall and Distance Error. However, the Direct approach sometimes yielded higher Area Precision, suggesting that without external data, LLMs generate smaller, more conservative (but potentially incomplete) areas.
Model Size and Fine-tuning: A relatively small, fine-tuned LLM (e.g., Qwen 14B or Llama 8B) achieved performance comparable to much larger off-the-shelf models (e.g., Llama 405B, Qwen 72B).
State-of-the-Art: The best results were achieved by the Geoparser-augmented approach using Qwen models. The fine-tuned Qwen 14B achieved an Area F1-score of 0.266 (with Google Maps geoparser) and 0.248 (with oracle), outperforming the previous best method (GBSP, F1=0.240) despite having access to less complex geometric data (center coordinates vs. polygons).

Qualitative Analysis

Error analysis revealed that models often flip latitude/longitude signs (e.g., positive vs. negative) due to ambiguity in pretraining data formats. Additionally, in the Geoparser-augmented setting, models sometimes exhibited "over-reliance" on provided coordinates, simply taking the min/max of the mentioned locations' coordinates to form the bounding box, rather than reasoning about the specific spatial relationship described in the text.

Significance and Contributions

The paper claims the following contributions:

Disentanglement of Capabilities: It provides empirical evidence that for geocoding compositional references, LLMs' reasoning skills are superior to their internal knowledge bases, suggesting that external knowledge augmentation is critical.
Novel Strategy: It proposes the first end-to-end strategy for geocoding compositional location references, combining traditional geoparsers for knowledge retrieval with LLMs for reasoning.
Bounding Box Representation: The authors introduce the use of bounding boxes to ground locations not present in geographic databases. This offers a simple yet informative way to represent area-based entities, bridging the gap between point coordinates and complex polygons.
Efficiency: The work demonstrates that small, fine-tuned models can match the performance of massive, off-the-shelf models, making the task more accessible and computationally efficient.

The authors conclude that their methods achieve state-of-the-art performance for this specific task and offer practical recommendations: use a Geoparser-augmented approach with a smaller fine-tuned model for best overall performance and recall, or a Direct approach with a large model if prioritizing precision and lacking access to external geoparsers.