osm-tag-corpus

Frozen source material for embedding OpenStreetMap tags: every tag the OSM wiki documents, with its description in each language it is documented in, its related terms, and how often it is used worldwide.

No embeddings, and no fixed passages. The format in which a tag becomes a sentence changes how well it retrieves, so this ships the material and leaves the wording to the caller.

A record

{
  "tag": "tourism=hotel",
  "key": "tourism",
  "value": "hotel",
  "lang": "en",
  "title": "Tag:tourism=hotel",
  "description": "Hotel – an establishment that provides paid lodging.",
  "lead_sentences": [
    "A hotel is an establishment that provides paid lodging, usually on a short-term basis.",
    "..."
  ],
  "related_terms": [
    "hotel",
    "resort",
    "spa"
  ],
  "on": [
    "node",
    "area"
  ],
  "implies": "",
  "status": "approved",
  "count_all": 442848
}

field	meaning
tag	key=value, or just the key for a key page
lang	the wiki language this record's text comes from
description	the one-line summary from the wiki infobox
lead_sentences	the page's opening paragraph, split into sentences
related_terms	terms listed on the page. Sometimes brand names
on	node, way, area, relation
implies	tags the wiki says this one implies. Often empty
status	approved, proposed, rejected_or_proposed, unknown
count_all	uses worldwide, from taginfo

A tag documented in several languages appears several times, one record per language. Fold to one score per tag at retrieval time and a hit in any language counts for that tag.

related_terms carries mappings no description supplies: amenity=fast_food lists Burger King, Döner, Kebab, McDonald's and Subway.

Size

31,913 records, 9,803 distinct tags, 70 languages, 48,852 body sentences. Built from taginfo dumps of 2026-01-30. See provenance.yaml for which tables were read and which were deliberately skipped.

The wiki documents tags unevenly across languages, and the twenty listed in the metadata are those with at least 85 records. The distribution has a long tail:

en 9467	de 3301	ru 2901	es 2709	fr 2047	ja 2033
cs 1985	pl 1816	pt 948	uk 907	sk 760	it 672
ko 340	el 299	zh-hans 255	nl 156	fi 142	ca 136
da 109	zh-hant 89	fa 85	+ 49 more

lang holds the wiki's own code, so zh-hans and zh-hant are separate there even though the Hub metadata collapses them to zh.

Assembling a sentence

How a tag is worded into a sentence changes how well it retrieves, so no formatted text is shipped here and the wording is left to whatever uses this.

yuiseki/osm-tag-embedding builds passages from these fields in several patterns, embeds them with several models, and scores the combinations.

Source

GitHub: yuiseki/osm-tag-corpus

Extracted from the taginfo database dumps.

License

ODbL-1.0. (The Hub's license identifier for this is odbl.)

This dataset is a Derivative Database of OpenStreetMap:

(c) OpenStreetMap contributors, available under the Open Database License.
https://www.openstreetmap.org/copyright

Sentences assembled from these fields, and any index or embedding built from them, are derived from OSM and carry ODbL too. The full notice is in LICENSE beside this file.