yuiseki/wikidata-gazetteer

Wikidata Gazetteer

Every geographic item in Wikidata: 12,205,328 places carrying 62,506,107 names in 552 languages, with coordinates, the administrative hierarchy, and a count of how many Wikipedias write about each one.

Built from the 20260831 JSON dump. CC0, like Wikidata itself.

Code, provenance and the build pipeline: https://github.com/yuiseki/wikidata-gazetteer

Why this exists

Wikidata holds the deepest multilingual place-name data of any openly licensed source. Measured against OpenStreetMap, which carries 4,200,089 localized names across its 7,678,203 named place=* objects, this is 5.1 names per place against 0.55.

The pieces were already published separately: full entity dumps, RDF triples, per-language label tables. None of them is a gazetteer. This joins the names to the positions and the hierarchy.

Two tables

places, one row per place:

names, one row per name, joined back on qid:

One row per name rather than a merged best name. Divergence between spellings is information, and a reader deciding which form to prefer needs to see them all.

Coverage

Names per place: 30.0% have one, 70.9% have three or fewer, 6.3% have more than ten.

What is actually in here, which is not only settlements

Selection is by P625 or P1566, that is, anything with a location. The twenty commonest classes:

Physical features and buildings outnumber settlements. That is a fact about granularity rather than about importance: there are half a million mountains and about two hundred countries, so any ranking by count puts mountains first and countries nowhere. Which of them matters depends on the text being matched.

Things in here that are not places at all

Selecting on P625 takes everything Wikidata gives a coordinate to, and it gives coordinates to more than places.

Languages, 729 of them. Wikidata places a language where it is spoken. Q150 French sits at 48.85, 2.35, which is Paris; Q1860 English at 51.0, 0.0; Q7737 Russian at 55.0, 38.0. 729 items out of 12.2 million is nothing by count and a tenth of the matches in running text, because a document writes "French" far more often than it names most towns. Measured against two corpora, removing them cut matches by 9.9% in an English travel guide and 10.7% in United Nations documents.

Organisations, about 36,000. nonprofit organization 13,617, business 11,635, organization 11,250, carrying the coordinates of their premises. These do less harm: their names are long and specific, like Boston Children's Museum, so they rarely collide with ordinary text.

Of the 840,857 names that survive a one-word prominence filter, 1,786 resolve to a language and 5,462 to an organisation.

Removing them

-- the language classes, found by taking the classes that co-occur with -- Q34770 language and keeping those whose label names a kind of language SELECT * FROM places WHERE NOT listhasany(instance_of, [ 'Q315','Q1036511','Q1097949','Q11499915','Q1149626','Q11820611','Q1208380', 'Q1288568','Q1322198','Q135295328','Q152559','Q17376908','Q1790577', 'Q20162172','Q20671152','Q21663239','Q215844','Q2315359','Q23492', 'Q250858','Q25295','Q2630831','Q2737212','Q2966838','Q3123468','Q33215', 'Q332','Q33289','Q33384','Q335214','Q33742','Q33831','Q33956','Q3329375', 'Q34228','Q34770','Q38058796','Q399495','Q4085712','Q436240','Q45762', 'Q455374','Q4536543','Q470775','Q61566','Q645304','Q778873','Q838296', 'Q839470','Q941501','Q951873' ]);

-- and the organisations SELECT * FROM places WHERE NOT listhasany(instance_of, ['Q163740','Q4830453','Q43229']);

Matching every name in this table against running text will still produce false positives, because street names, house names and hotel names are in it, and because a one-word name like Council names a town in Idaho. Two rules help, measured on 1.31 billion characters of UN documents: require a match to cover a whole run of capitalised words, and require a one-word name to clear a prominence bar. At 100 sitelinks for single words, the top twenty-five names matched in that corpus are all real.

The hierarchy runs in both directions, unevenly

located_in (P131) is the one to build on. 10,806,828 places carry it, and reversing it yields 395,046 distinct containers, of which 357,043 are rows in this table, so 90.4% of the hierarchy resolves without leaving the dataset.

contains (P150) states the same relation downward, and is far sparser: only 80,862 places name their children, across 1,093,138 edges. Where it is present it is well kept. Japan lists its 47 prefectures; Tokyo lists its 96 wards and municipalities. Treat it as a cross-check on the reversed P131, not as the hierarchy itself.

How much of this is linked to the other big gazetteers

Two thirds of these places, 64.6%, are linked to neither. That is worth knowing before treating agreement between gazetteers as evidence: this source is largely not a restatement of either of the other two.

Known limits

Japanese labels reach only 4.0% of places, and other non-European languages are thinner still. The language distribution follows Wikipedia's, so ceb has 3,459,736 labels from bot-generated articles while most languages have far fewer.

33.2% of places carry a GeoNames id, so two thirds have no link to GeoNames at all. That makes this a substantially independent source, which is useful when weighing agreement between gazetteers, but it also means the two do not simply overlap.

Wikidata is crowd-edited. Positions and hierarchies are as good as the last person to touch them.

Licence

CC0-1.0, inherited from Wikidata. The code that built it is Apache-2.0 at https://github.com/yuiseki/wikidata-gazetteer