README.md

metadata

task_categories:
  - feature-extraction
language:
  - ar
size_categories:
  - 100K<n<1M
pretty_name: Moroccan Land-Registration Notices (Bulletin Officiel, 2020–2026)
tags:
  - information-extraction
  - structured-extraction
  - arabic
  - legal
  - land-administration
  - low-resource
  - gold-standard

Moroccan Land-Registration Notices (Bulletin Officiel, 2020–2026)

208,833 Arabic land-registration notices published in the Bulletin Officiel of the Kingdom of Morocco between 2020 and 2026, with structured fields extracted for each, plus a 198-notice evaluation set annotated by hand and independently twice.

Under Article 37 of the Dahir of 12 August 1913, as amended by Law 14-07, an application to register a parcel must be published in the Bulletin Officiel, and any third party has two months from that publication to file an objection. The published extract is therefore not an administrative summary but the legal instrument that opens the objection window. A century of this record exists only as Arabic text inside PDF files; none of it was queryable before this work.

⚠️ Read this before using the full config

The two configurations do not have the same quality and must never be pooled.

config	notices	fields produced by	use for
full	208,833	a deterministic parser	deriving, mining, pre-training, weak supervision
gold	198 × 2	human annotators, twice, independently	evaluation only

The full config is silver data. Measured against the gold config, its per-field F1 has a median of 96.4 % — but that median hides a wide spread. Four fields are exact and 24 of 29 exceed 90 %, while the administrative levels are much weaker:

field	F1 (%)	field	F1 (%)
bo_origine, bo_origine_date	100.0	situation.prefecture	87.7
indice, requisitions_concernees	100.0	droits_reels	87.0
concerne_requisition	99.3	zone_collective	85.7
demandeurs.nom	98.4	situation.quartier	79.1
date_depot	98.1	situation.commune	73.4

A dataset in which dates are right 98 % of the time but communes are right 73 % of the time supports some queries and quietly breaks others. Check the field you depend on before you depend on it.

Do not evaluate a system on the full config. Its labels come from the system this dataset was built to evaluate. Evaluating on them measures agreement with that parser, not correctness.

Quick start

from datasets import load_dataset

full = load_dataset("<org>/moroccan-land-registration-notices", "full", split="train")
gold = load_dataset("<org>/moroccan-land-registration-notices", "gold", split="adjudicated")
# gold has three splits: annotator_1, annotator_2 (both unadjudicated),
# and adjudicated — the one to score against.

print(full[0]["source_text"])      # texte arabe brut de l'annonce
print(full[0]["situation"])         # {"texte": ..., "prefecture": ..., "commune": ...}

The id column joins the two configs: all 198 gold ids are present in full, so the parser's output and the human annotation for the same notice can be compared directly.

gold_ids = set(gold["id"])
paired = full.filter(lambda r: r["id"] in gold_ids)   # 198 notices

What is in the corpus

336 bulletin issues, 2020–2026, 83 registry offices, two editorial sections.

notice type	count	section
مطلب التحفيظ — registration application	98,220	A
مطلب التحفيظ — demarcation completion	102,891	B
خلاصة إصلاحية — rectifying summary	6,314	A, B
إصلاح غلط — erratum	1,245	A, B
التحفيظ الجماعي — collective registration	163	B

Counts are by effective type. Note the trap in row one and two: the upstream header does not distinguish a registration application from a demarcation-completion notice — the section column does. Section A carries the initial extracts, section B the demarcation-completion notices. 2,226 notices are published under a header that contradicts their content and are reclassified by textual markers; notice_type keeps the header, and notice_type_effective the reclassification.

Source text is 127 MB, median 541 characters per notice. It comes from the PDF text layer via pdftotext; there is no OCR stage, so its defects are character-order and layout artefacts, not recognition errors.

Fields

Identification and provenance:

column	type	note
id	string	{bulletin}_{section}_{requisition}, unique
bulletin, year, publication_date	string, int16, string	issue of the gazette
section	string	A or B — see the trap above
registry_office	string	one of 83
requisition_number, notice_type, notice_type_effective	string
page_start, page_end	int32	page range in the PDF
source_text	string	the verbatim Arabic notice — the input of the task
annotation_source	string	parser in full, human in gold
annotator	string	non-null in gold only

Extracted fields — 29 in the evaluation schema, all normalised rather than span-based (dates are ISO, areas are square metres, references are canonical, shares are fractions), which is why value-level scoring is the only meaningful protocol here:

column	type
date_depot, date_bornage, date_bornage_prevue	string (ISO)
demandeurs	list of {nom, part}
nom_donne, nom_actuel, nom_propriete, nature	string
situation	struct {texte, prefecture, cercle, caidat, commune, douar, quartier}
superficie_m2, superficie_texte	float64, string
limites	struct {nord, est, sud, ouest}, each {texte, ref_titres[], ref_requisitions[]}
droits_reels	string
origine_propriete	list of {texte, type, date}
bo_origine, bo_origine_date, concerne_propriete, concerne_requisition	string
annule_annonce_bo, nouvel_avis	string, bool
zone_collective, indice, requisitions_concernees	string, string, list
parser_slices	string (JSON) — the parser's intermediate slices, full only

Empty values are null throughout. The parser writes "" and annotators sometimes wrote null; both mean "field absent" and are normalised to null here, because distinguishing them in the published data would invite false comparisons.

The cross-reference fields

limites.*.ref_titres and limites.*.ref_requisitions are the most valuable and the most difficult part of the dataset. Boundary descriptions mix literal text ("a river", "a public road", a neighbour's name) with references to other parcels, cited by registration or title number. Resolved across the corpus, they link parcels to one another — a parcel adjacency structure recovered from public text alone, in a country whose cadastre publishes no geometries.

References are normalised to a single canonical form from at least five observed spellings. Where a neighbour is cited without its registry-office code, the prefix is inherited from the citing requisition, which is sound because a parcel can only adjoin parcels administered by the same office. Ambiguous cases are discarded rather than resolved arbitrarily: fabricating an identifier would silently create an adjacency that does not exist, and a false edge is indistinguishable from a true one downstream.

The gold config

198 notices stratified by type and year, covering 149 bulletin issues and 70 registry offices: 116 registration applications, 50 rectifying summaries, 24 errata, 8 collective registrations.

Each notice was annotated twice and independently. The gold config ships three splits: annotator_1 and annotator_2 are the two independent annotations, unadjudicated, and adjudicated resolves the 178 values the two disagree on. Inter-annotator agreement — computed from the two unadjudicated splits — is a median per-field F1 of 92.3 % (mean 91.1 %). The fields that stay low are those where the convention is genuinely under-specified rather than merely unwritten: situation.quartier (66.7 %), situation.commune (80.2 %), origine_propriete.type (82.4 %).

How this gold standard was built matters, and we recommend reading the accompanying paper before trusting it. The first version was produced by having an annotator correct pre-filled parser output — the economical and widespread protocol. Against that version the parser scored 100 % precision on 22 of 29 fields. Re-annotating the same notices under an instruction to recompute rather than validate dropped the same fields to between 36 and 67 %. The published version is the re-annotated one. One caveat survives:

Shipping the two unadjudicated splits is deliberate: a user may adopt either annotator, take their intersection, or re-adjudicate under their own conventions, and the residual disagreement stays visible instead of being absorbed into a merge. The adjudicated split is the single reference for those who need one. It was resolved from the source text alone — never against the parser, whose labels a reference adjudicated against it would agree with for reasons no score would reveal — and it is the split the benchmark is scored against. If you report one number on this dataset, report it on adjudicated.

The annotation guide is included in the repository as ANNOTATION_GUIDE.md. It is what makes the agreement figures interpretable, and the incompleteness of its first version is the documented cause of part of the initial disagreement.

Personal and sensitive information

This dataset names people. It contains 265,031 mentions of natural persons — 163,356 distinct names — each associated with a declared share in an identified property at an identified location. Names appear both in the demandeurs field and, for 96 % of them, verbatim inside source_text.

Redacting the names is not an option that preserves the dataset: source_text is the input of the extraction task, and a censored input measures nothing. Pseudonymising only the demandeurs field would be worse than useless — it would suggest a protection that the raw text immediately defeats.

The publication in the gazette is constitutive, not incidental: it is what opens the objection window, and the record is legally required to be public. That is a real argument, and it is not a sufficient one. Publishing a gazette as PDF and publishing a structured, indexable, forkable corpus are different acts with different consequences, and we do not treat the first as authorising the second without conditions. Hence:

• access is gated, under the conditions stated on the access form;
• users undertake not to re-identify, enrich, or cross-reference individuals;
• records are withdrawn on request — open an issue or contact the maintainer;
• no derived aggregate over named individuals is published here.

If you are building on this dataset, the honest default is to work from the non-personal fields (location, area, dates, boundaries, references) and to touch demandeurs and source_text only when your task genuinely requires them.

Known limitations

• Party names are surface forms. Homonymy and orthographic variation are not resolved, and the same person may appear under several spellings. Person-level entity resolution is future work; do not treat a name as an identifier.
• The full labels are parser output with the per-field quality given above, not ground truth.
• The parser is partly tuned to the gold set. Ten corrections were guided by error analysis on it, so full's quality on unseen notices may be slightly below the reported figures.
• The evaluation set is 0.09 % of the corpus, stratified by type and year but not by registry office — 70 offices for 198 notices.
• Coverage is 2020–2026 only. The gazette has published these notices for over a century; the earlier record is not included.
• Bidirectional text. Identifiers mixing Arabic letters and digits are stored in logical order, which does not match how they render. Reading the rendering rather than the string reverses them. Both of our annotators made this error, in opposite directions. Compare strings, not screenshots.

Licensing

Two distinct layers, and they do not have the same status:

• The annotations, the extracted fields, the schema and the tooling are released by the dataset authors under CC BY 4.0.
• The source text is an official publication of the Kingdom of Morocco. It is reproduced here as short extracts of a legal instrument for research purposes. Its reuse is governed by the terms of the Bulletin Officiel and not by the licence above.

Use of the dataset is additionally subject to the access conditions.

Contact and withdrawal

To request the withdrawal of a record, or to report an error in an extracted field, open an issue on this repository or contact the maintainer. Withdrawal requests concerning personal data are honoured without justification.