NoeFlandre/osm-polygon-wikidata-sentence-relevance

OSM Polygon Wikidata Sentence Relevance

This dataset contains normalised sentences extracted from OpenStreetMap (OSM) polygons and their linked Wikipedia / Wikivoyage pages. Each row is a deduplicated sentence occurrence scoped to a polygon, language, and content hash.

Dataset scope

Current release: Afghanistan only (afghanistan shard).

Dataset summary

• Total sentence rows: 54462
• Input sentence occurrences: 56310
• Duplicates removed: 1848
• Cross-source duplicate groups: 5
• High-confidence residual boundary violations: 0
• Unique polygons: 161
• Unique Wikidata entities: 154
• Unique documents: 1680
• Languages: 115
• Rows with coordinates: 54462 / 54462
• Sentence length (chars): min 1, mean 99.65, max 2967
• Coordinate extent: 30.4864 → 38.4539, 60.8661 → 74.0156
• Input dataset revision: 9bf844b11231fafd2c0cf17d2e40cc8eff23336b
• Pipeline version: 0.1.0
• Exported Parquet SHA-256: 50495f15688ad8b32e1e0052b4d560c96a5c13b87ec04d72089728cadc315ea0

Sources

Key	Count
wikipedia	54152
wikivoyage	310

Source dataset and recorded input revision

The sentences were extracted from the input dataset NoeFlandre/osm-polygon-wikidata-only at the recorded input revision 9bf844b11231fafd2c0cf17d2e40cc8eff23336b. This value is a recorded immutable commit identifier; the card does not assume which acquisition mechanism produced it.

Geographic coverage

One dot represents each unique polygon with coordinates; the map extent is derived from the data.

Language coverage

Full language breakdown (115 languages)

Language	Rows
en	6072
fa	3871
de	3586
fr	2997
ps	2473
es	2424
uz	2203
bn	1762
ru	1617
sl	1607
ar	1522
ca	1430
it	1195
ceb	1163
id	1127
zh	927
ur	860
nl	857
fi	796
tr	795
ta	759
sv	694
tg	685
uk	614
hu	586
sr	584
pl	572
el	560
ja	541
pnb	535
he	518
no	450
az	439
ml	439
pt	433
cs	378
hi	330
arz	304
vi	289
ms	281
ko	263
th	249
sd	247
pa	227
ro	204
mk	188
be	176
ka	175
eo	170
et	167
bg	158
ky	158
hy	152
tt	147
nn	137
dag	128
eu	128
te	125
lt	116
mt	111
da	98
lv	96
ckb	95
hr	94
ks	91
kk	82
sk	82
mos	81
ast	80
sh	80
fat	79
gl	79
af	63
sat	63
si	47
bs	44
as	43
azb	43
jv	37
mr	37
mn	34
la	30
gur	28
ff	22
sq	20
tw	18
oc	15
skr	15
bar	12
frr	12
my	12
mai	11
cy	9
ha	9
hsb	9
os	9
tl	9
mzn	8
fy	7
new	7
ce	5
diq	5
ne	5
cv	4
gd	4
tum	4
wuu	4
ba	3
kab	3
mad	3
sco	3
xmf	3
ug	2
war	2
rw	1

Processing method

The exact implementation is available in the GitHub source at the producing commit. Each row is produced by the following fixed sequence:

1. Section input. One Wikipedia or Wikivoyage section linked to an OSM polygon is processed at a time. Its language is supplied to the segmenter, and section order is retained.

2. Model segmentation. Sentence boundaries are inferred with segment-any-text/sat-12l-sm at revision d70c72a9331b2d5a9e82baad00c64964a23a09bb. It receives the full section text and its language code, predicts boundary probabilities, and returns slices of the original text at the selected offsets. The model is not asked to rewrite text: it does not generate or paraphrase sentences.

3. Boundary repair. This conservative residual-boundary repair is a deterministic second pass that looks only inside each model-produced slice. The marks ?, !, ؟, 。, ！, ？, ।, or ॥ are boundaries when the next non-space character is a letter. A period is split only when both sides contain at least 12 characters, the preceding letter-token is longer than two characters, and the next letter is not a lowercase continuation. Arabic-tagged period boundaries must additionally continue in Arabic script; terminal punctuation stays with the preceding sentence, together with any immediately closing quote or bracket.

   For example, Delegation arrived. They left. becomes two rows. The guards preserve short abbreviations (Dr. Smith), numeric values (10.5), lowercase continuations (approx. five), and URL query strings (https://example.org/?q=x).

4. Normalization. Each segment is trimmed, normalized to Unicode NFC, stripped of configured zero-width characters (U+200B, U+2060, U+FEFF) and leading MediaWiki edit markers, then has control characters replaced with spaces and whitespace collapsed. Case, punctuation, accents, ZWNJ, and ZWJ are preserved.

5. Context and identity. Sentences are indexed in section order; previous_sentence and next_sentence are assigned after splitting. sentence_id derives from polygon_id, language, and the SHA-256 of the normalized text.

6. Polygon-scoped deduplication. Exact normalized duplicates are collapsed only within the same polygon and language. Cross-source collisions retain all contributors in duplicate_sources, with Wikipedia as the canonical row.

7. Publication audit. Publication scans every normalized sentence with the same deterministic rules and counts any boundary candidate still found inside a row. This audit does not run the model again: it verifies that the model-plus-repair output contains no boundary the repair itself considers unambiguous. A publishable release must therefore record exactly 0 residual candidates in both the manifest and card. This is a regression guard; it does not claim that every ambiguous boundary is linguistically perfect. The validator also checks the schema, Parquet hash, generated statistics, example row, and plot hashes before release.

Real example row (from the export)

Show one row from the canonical-sorted Parquet

{
  "sentence_id": "06f1a382b4c9d124a42db84688701b6df9abff18ccaa8b2de15a57a21a000f80",
  "polygon_id": "afghanistan-latest:way:1008689574",
  "wikidata": "Q15040964",
  "document_id": "Q15040964:wikipedia:ar:8074118:75435549",
  "article_id": "Q15040964:ar:8074118:75435549",
  "source": "wikipedia",
  "language": "ar",
  "site": "arwiki",
  "page_title": "مطار نيلي",
  "section_id": "016fc0bca703007bed7068ab5f28ed3420404cdd5f4ec962353759b6e1bbc300",
  "section_index": 3,
  "section_path": [
    "روابط خارجية"
  ],
  "sentence_index": 2,
  "sentence_text_raw": "حالة الطقس في مطار نيلي .",
  "sentence_text_normalized": "حالة الطقس في مطار نيلي .",
  "previous_sentence": "سجل المطار لمطار نيلي في Landings.com.",
  "next_sentence": "بوابة طيران بوابة أفغانستان",
  "url": "https://ar.wikipedia.org/wiki/%D9%85%D8%B7%D8%A7%D8%B1_%D9%86%D9%8A%D9%84%D9%8A",
  "page_id": 8074118,
  "revision_id": 75435549,
  "revision_timestamp": "2026-06-13T13:19:25Z",
  "document_content_hash": "52a99eed9884247242099dcb41b5793027614f347fef9b06d3e7d1a1d05747dd",
  "section_content_hash": "a4414ef098f11ffd5fc9aecbea544c085a238dba016e15feee7844a477e4915a",
  "sentence_content_hash": "92504074d860df7bf7a8c631eb7e962657770e0391d98ee9635c3218c674e518",
  "duplicate_occurrence_count": 1,
  "duplicate_sources": [
    "wikipedia"
  ],
  "polygon_name": "فرودگاه نیلی",
  "osm_primary_tag": "aeroway=aerodrome",
  "osm_tags": [
    {
      "key": "aeroway",
      "value": "aerodrome"
    },
    {
      "key": "name",
      "value": "فرودگاه نیلی"
    },
    {
      "key": "name:ar",
      "value": "مطار نيلي"
    },
    {
      "key": "name:en",
      "value": "Nili Airport"
    },
    {
      "key": "name:fa",
      "value": "فرودگاه نیلی"
    },
    {
      "key": "name:zh",
      "value": "尼力机场/尼力機場"
    },
    {
      "key": "name:zh-Hans",
      "value": "尼力机场"
    },
    {
      "key": "name:zh-Hant",
      "value": "尼力機場"
    },
    {
      "key": "source",
      "value": "ourairports.com"
    }
  ],
  "region": "afghanistan",
  "lat": 33.73966181362143,
  "lon": 66.15739853509896,
  "input_dataset_revision": "9bf844b11231fafd2c0cf17d2e40cc8eff23336b",
  "pipeline_version": "0.1.0"
}

Schema

The export is a single Parquet table. osm_tags is represented as a Viewer-compatible list of {key, value} structs.

Show all fields

Field	Type	Nullable	Description
sentence_id	string	no	Deterministic ID from polygon, language, and sentence content hash.
polygon_id	string	no	OSM polygon identifier.
wikidata	string	no	Wikidata entity QID for the polygon/page.
document_id	string	no	Document/page identifier within its source/site/language.
article_id	string	yes	Article identifier where available.
source	string	no	wikipedia or wikivoyage.
language	string	no	Language code of the sentence.
site	string	no	Source site (e.g. en.wikipedia.org).
page_title	string	no	Page title.
section_id	string	no	Section identifier.
section_index	int64	no	Section ordinal within the document.
section_path	list<string>	no	Section breadcrumb path.
sentence_index	int64	no	Sentence ordinal within the section.
sentence_text_raw	string	no	Segment text after surrounding-whitespace trimming.
sentence_text_normalized	string	no	Normalised sentence text used as the dedup key.
previous_sentence	string	yes	Prior sentence in the section (context).
next_sentence	string	yes	Next sentence in the section (context).
url	string	no	Source URL of the document.
page_id	int64	no	Source page ID.
revision_id	int64	no	Source revision ID.
revision_timestamp	string	no	Source revision timestamp.
document_content_hash	string	no	Hash of the source document.
section_content_hash	string	no	Hash of the source section.
sentence_content_hash	string	no	Hash of the normalised sentence (dedup key component).
duplicate_occurrence_count	int64	no	Number of source occurrences collapsed into this row.
duplicate_sources	list<string>	no	Distinct sources among the collapsed occurrences.
polygon_name	string	yes	Human-readable polygon name.
osm_primary_tag	string	yes	Primary OSM tag of the polygon.
osm_tags	list<struct<key: string, value: string>>	no	OSM tags of the polygon, encoded as a list of {key, value} structs so the Hugging Face Viewer can ingest the export.
region	string	no	Input region/extract name.
lat	double	yes	Latitude of the polygon centroid, if known.
lon	double	yes	Longitude of the polygon centroid, if known.
input_dataset_revision	string	no	Exact input revision recorded for reproducibility.
pipeline_version	string	no	Pipeline version recorded for reproducibility.

Provenance and reproducibility

The Parquet metadata and manifest.json record the input dataset, input revision, pipeline version, model revision, producing source commit and artifact hashes. Identical pinned inputs and locked dependencies produce deterministic artifacts.

Uses and limitations

Suitable uses include multilingual corpus analysis and research on how places are described across Wikipedia and Wikivoyage. This is not a labelled relevance, similarity or classification dataset.

• Extraction depends on upstream OSM / Wikimedia availability, coverage and language balance.
• Sentence segmentation uses an automatic multilingual model. Publication rejects high-confidence residual boundaries, but ambiguous short or mixed-script text can still be imperfect.
• Deduplication is exact, not semantic, and does not imply relevance.

Licensing

The code is covered by the repository licence. Dataset content retains its upstream terms, including ODbL for OpenStreetMap and CC BY-SA or project-specific Wikimedia terms. Source URLs and revision identifiers are retained for attribution; no single SPDX identifier covers the combined content, so the card uses license: other.