NoeFlandre/geoname-ambiguity-fr

GeoName Ambiguity FR

GeoName Ambiguity FR is a CC0 Wikidata-derived benchmark for French geographic entity disambiguation, focused on ambiguous place names and deterministic administrative context.

The dataset is built only from Wikidata structured data. It does not scrape or derive examples from Wikipedia prose.

Dataset Scope

Version 0.1 includes:

• French communes
• former French communes
• French departments
• French regions

Hamlets, villages, localities, and natural places are intentionally deferred to later versions because their Wikidata coverage is noisier.

Dataset Statistics

Metric	Value
Unique normalized mentions	6,126
Candidate Wikidata entities	19,891
Rows	71,836
Train rows	57,368
Validation rows	7,012
Test rows	7,456

Difficulty distribution:

Difficulty	Rows
easy	7,724
medium	8,881
hard	55,231

Each context type has 17,959 rows: admin_template, feature_type_template, coordinate_template, and minimal.

Top ambiguous mentions include Saint-Pierre, Neuville, Saint-Julien, Saint-Denis, Saint-Aubin, La Chapelle, Saint-Laurent, Fontaine, Beaumont, and Saint-Martin.

Quality Diagnostics

Generated diagnostics are committed in data/quality_report.json.

Diagnostic	Value
Candidate count min / median / max	2 / 4 / 89
Rows with fuzzy distractors	63,388
Candidate department coverage	99.03%
Candidate region coverage	98.75%
Candidate population coverage	88.19%

Candidate relation distribution:

Relation	Count
exact	177,480
alias_exact	278,892
fuzzy_distractor	121,460

Coordinate bounds include metropolitan and overseas French administrative places represented in Wikidata.

Schema

Each row contains:

• id: stable row id within a split
• mention: surface place-name mention
• normalized_mention: lowercase accent-stripped grouping key
• context: deterministic generated context
• context_type: admin_template, feature_type_template, coordinate_template, or minimal
• candidates: candidate Wikidata entities with labels, aliases, coordinates, admin metadata, population, and relation to the mention
• answer_id: correct Wikidata QID
• difficulty: easy, medium, or hard
• source: always wikidata
• source_license: always CC0-1.0
• retrieved_at: Wikidata retrieval date
• wikidata_query_hash: SHA256 hash of the rendered query configuration
• candidate_sources: Wikidata URLs for all candidates

Candidate relation values are exact, alias_exact, and fuzzy_distractor, so exact name disambiguation can be separated from fuzzy matching distractors.

Baselines

Baseline	Overall accuracy
random candidate	0.2196
most populous candidate	0.2101
same-department exact match	0.4391
string similarity only	0.3411

The same-department baseline reaches 0.7304 accuracy on admin_template contexts and drops to 0.3411 on minimal/coordinate contexts.

Reproducibility

The repository commits the SPARQL query template and generated metadata in data/reproducibility.json, including endpoint URLs, retrieval date, row counts before and after filtering, raw response hash, query hash, seed, page size, and the exact v1 Wikidata instance classes.

The full raw Wikidata/API cache is not committed by default. Build caching lives under .cache/wikidata/ and is ignored by git.

Citation

@dataset{flandre_2026_geoname_ambiguity_fr,
  author = {Flandre, Noé},
  title = {GeoName Ambiguity FR},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/NoeFlandre/geoname-ambiguity-fr}
}

Licenses

Code in this repository is released under the MIT license. Dataset files are released as CC0-1.0 because they are derived only from Wikidata structured data, which is available under CC0. Wikidata attribution is included as good scholarly practice.

Source data: Wikidata, retrieved on 2026-05-14. Wikidata structured data is available under CC0-1.0.

Build

uv sync
uv run python scripts/build_dataset.py
uv run python scripts/validate_dataset.py
uv run python scripts/run_baselines.py
uv run python scripts/report_quality.py

Local loading:

from datasets import load_dataset

dataset = load_dataset("data")

Docker:

docker build -t geoname-ambiguity-fr .
docker run --rm geoname-ambiguity-fr