Unifying points of interest taxonomies: mapping OpenStreetMap tags to the Foursquare category system

TL;DR

This work tackles the interoperability challenge posed by heterogeneous POI taxonomies across OpenStreetMap (OSM) and Foursquare (FS). It introduces a three‑part framework: a manually curated benchmark mapping OSM tags to FS categories, an embedding‑based semantic alignment stage, and an LLM‑based refinement stage to robustly select the best matches, all supported by a scalable update pipeline. The authors release cleaned taxonomies, the oracle benchmark, enriched FS descriptions, embedding results, and end‑to‑end notebooks to enable reproducible evaluation and long‑term maintenance. The results show that embedding retrieval combined with LLM refinement substantially improves alignment accuracy, achieving around 85% coverage at the top FS level and over 72% at deeper levels, with meaningful gains over baseline methods. Overall, the work delivers an openly available benchmark and toolchain that enable reproducible, scalable unification of heterogeneous POI taxonomies for urban analytics and smart city applications.

Abstract

The heterogeneity of Point of Interest (POI) taxonomies is a persistent challenge for the integration of urban datasets and the development of location-based services. OpenStreetMap (OSM) adopts a flexible, community-driven tagging system, while Foursquare (FS) relies on a curated hierarchical structure. Here we present an openly available benchmark and mapping framework that aligns OSM tags with the FS taxonomy. This resource integrates the richness of community-driven OSM data with the hierarchical structure of FS, enabling reproducible and interoperable urban analytics. The dataset is complemented by an evaluation of embedding and LLM-based alignment strategies and a pipeline that supports scalable updates as OSM evolves. Together, these elements provide both a robust reference resource and a practical tool for the community. Our approach is structured around three components: the construction of a manually curated benchmark as a gold standard, the evaluation of pretrained text embedding models for semantic alignment between OSM tags and FS categories, and an LLM-based refinement stage that enhances robustness and adaptability. The proposed methodology provides a scalable and reproducible solution for taxonomy unification, with direct applications to urban analytics, mobility studies, and smart city services.

Figures (6)

• Figure 1: Integration and semantic mapping between OSM and FS data sources. (a) Examples of corresponding POIs from OSM and FS in the city of Bologna (Italy), showing differences in tagging structure and coverage. (b) Overview of the OSM-FS mapping pipeline, including exact lexical matching, semantic validation, and embedding-based alignment refined with LLM assistance. The process produces FS-tagged OSM POIs and an extended dataset combining both taxonomies.
• Figure 2: Absolute and relative frequencies of OSM–FS correspondences in the manual benchmark mapping, by FS main category and match type (lexical, semantic, main-category only).
• Figure 3: ROC curves for embedding-based retrieval with and without FS category descriptions. Adding descriptions improves the discriminative ability of similarity scores, even when top-1 accuracy slightly decreases.
• Figure 4: Comparison of match percentages for OSM and FS categories at $k=20$ across the four prompting strategies. On the left, the radar chart shows the match percentages across OSM tags, while on the right, the radar chart displays the corresponding match percentages for FS categories.
• Figure 5: Top-1 accuracy (%) by hierarchy depth for the all-MiniLM-L6-v2 model. Each bar shows the baseline performance without refinement (blue) and the additional improvement achieved with the Fallback – No example prompt at $k=20$ (green). Values are annotated at the top of each bar.