ReaGeo: LLM-Based End-to-End Geocoding Framework
ReaGeo is an end-to-end geocoding framework that leverages large language models to replace traditional multi-stage approaches reliant on geographic databases. The method converts geographic coordinates into geohash sequences, framing coordinate prediction as text generation. A Chain-of-Thought mechanism enhances reasoning over spatial relationships, while reinforcement learning with a distance-deviation-based reward optimizes accuracy. Experiments show ReaGeo handles explicit address queries and vague relative location queries effectively. The model demonstrates strong predictive capabilities, overcoming workflow complexity, error propagation, and dependence on structured knowledge bases.
Key facts
- ReaGeo is an end-to-end geocoding framework based on large language models.
- It converts geographic coordinates into geohash sequences.
- The coordinate prediction task is reformulated as a text generation problem.
- A Chain-of-Thought mechanism enhances reasoning over spatial relationships.
- Reinforcement learning with distance-deviation-based reward optimizes generation accuracy.
- Experiments show accurate handling of explicit address queries.
- The model effectively resolves vague relative location queries.
- It overcomes limitations of traditional multi-stage approaches.
Entities
—