Wish an instant map!

The aim of this work is to develop a proof-of-concept (POC) for an open-source QGIS plugin capable of translating natural language queries into Overpass-QL to perform spatial, temporal and attributive filtering of OpenStreetMap (OSM) data. The plugin visualises query results directly in QGIS. For example, a query like "All cafés within 50 metres of a fountain in St. Gallen" is processed and rendered as a map. Unlike existing solutions that either require knowledge of Overpass-QL (e.g. QuickOSM) or are proprietary, this plugin aims to make such functionality accessible to non-experts.

The query processing consists of three main steps: (1) geoname recognition and geocoding (e.g. resolving "St. Gallen"), (2) recognition and semantic alignment of spatial entity sets (SES) with OSM attributes (e.g. mapping "café" to `amenity=cafe`), and (3) generation of Overpass-QL queries. Early attempts to implement the plugin using open source models such as LLaMA yielded suboptimal results. Consequently, a fine-tuned OpenAI GPT-4o model was used, resulting in significant improvements in query generation. Geonames were resolved using Photon, an OSM-based geocoder, in addition to OpenAI's assistant, and SES were mapped using semantic similarity analysis with pre-embedded OSM tags.

The finetuned LLMs were evaluated using 100 natural language queries, with the best fine-tuned GPT-4o model achieving a BLEU score of 0.67, significantly outperforming base models and open source alternatives. The exact match rate was 0.09, indicating room for improvement in the generation of perfectly accurate queries. Most of the generated Overpass-QL queries were functional within QGIS, with a high validity rate, although still lacking in semantic precision.

The resulting QGIS plugin, called Wish an Instant Map! (WAIM), was implemented in Python using the two preprocessing steps, together with the finetuned OpenAI LLM. The graphical user interface includes text input for queries, support for current map extents, and an expert mode for editing OverpassQL. While the system has demonstrated feasibility through black-box testing with English language queries, challenges remain, including reliability of generated queries and reliance on proprietary LLMs.
Future improvements can include the development of an OSM thesaurus to improve semantic matching, the integration of structured output for query validation, and the use of larger datasets or larger LLMs to fine-tune open-source models. Despite its limitations, WAIM illustrates the potential for combining AI with geospatial systems.