Bayanat Labs Arabic-Latin Name Variants

Arabic name variants Arabic transliteration Arabic name matching romanisation of Arabic names Wikidata names

Key numbers

Counted by script from the released files. Each key number is read from meta/stats.json, meta/bench_results.json or meta/sample_check_summary.json of the Hugging Face dataset at revision ceb1906e957768dbd36f21bf53ec8230eaebaaf3 and pinned to its SHA-256; percentages are computed from those files and rounded to one decimal place. MANIFEST.json lists every file with its size and SHA-256.

Downloads and identifiers

Get started

Load a config with the datasets library. The code installs from GitHub at the v0.1.0 tag; the package is not on PyPI.

from datasets import load_dataset

tokens = load_dataset("Bayanat-Labs/arabic-latin-name-variants", "token_variants", split="train")
pairs = load_dataset("Bayanat-Labs/arabic-latin-name-variants", "entity_variants", split="train")
clusters = load_dataset("Bayanat-Labs/arabic-latin-name-variants", "clusters", split="train")
bench = load_dataset("Bayanat-Labs/arabic-latin-name-variants", "name_matching_bench", split="test")

pip install "arabic-latin-name-variants @ git+https://github.com/Bayanat-Labs/arabic-latin-name-variants@v0.1.0"
arabic-latin-name-variants lookup "محمد" --release path/to/downloaded/dataset

What is released

Attribution: Arabic-Latin Name Variants by Bayanat Labs. Derived from Wikidata (CC0 1.0). https://huggingface.co/datasets/Bayanat-Labs/arabic-latin-name-variants

Intended uses

• Expand a name in search, or match a name across scripts in name screening and record linkage, with the Latin spellings attested for each Arabic name word.
• Test an Arabic-Latin name matcher on the name_matching_bench pairs and compare it with the four baselines.
• Support entity linking and pronunciation lexicons with the labelled pairs and the word alignment.

Out of scope

• Treating a name match as proof that two records are the same person: many people share a name, and the tables hold spellings, not identities.
• Reading a name string as a fact about a person: name strings are written by Wikidata editors, and some that read as descriptions or statements may remain.
• Reading counts as how often a spelling is used in a population: they describe Wikidata.

Rights at a glance

Training, display, redistribution and synthetic-voice use are separate permissions.

Limitations

• Sample precision: this is a separate check by Bayanat Labs on random samples, and the judgements are Bayanat Labs' own reading, not a review by native speakers. We read 300 entity_variants rows drawn at random, half from each scope, and judged whether the Latin name renders the sounds of the Arabic name (transliteration), does so only for part of the name, with the rest translated, extra or missing (mixed), or is a translation, an exonym or another name (translation or exonym). Of the 123 rows labelled transliteration, 120 were transliterations (precision 97.6%); of the 121 labelled mixed, 103 were mixed (precision 85.1%; 15 of the others are translations or other names, 3 are full transliterations); of the 56 labelled translation_exonym, 47 had no transliteration relation (precision 83.9%). We also read 300 multi-word names drawn at random, half from each scope, whose alignment feeds the token tables: 290 were aligned correctly in every word pair (96.7%). The judged rows are in meta/sample_check.jsonl, and the summary in meta/sample_check_summary.json.
• The test-set pairs are built by rules from Wikidata, not labelled by people. Matching pairs are rows the relation filter labelled transliteration, and that filter uses the same consonant-skeleton similarity, which favours that baseline at one in a hundred. At one in a thousand the threshold rests on a few high-scoring non-matching pairs, so the skeleton's recall at that rate is unstable. The spelling-table lookup is built without the held-out entities but still benefits from common name words they share with the others.
• Wikidata over-represents public figures, and English and other Western European spellings dominate (English 18.7%, eight Western European language codes 88.6% of language labels on Latin name rows). Counts describe Wikidata, not how often a spelling is used in a population.
• An ala_lc value exists for 2.7% of Arabic names; most are Latin names in macron and underdot letters that may follow IJMES, BGN/PCGN or another system and are not verified as ALA-LC.
• People's names are personal data in many jurisdictions even when public. For people the release holds the QID, class, countries and name strings from Wikidata. Name strings are written by Wikidata editors; strings that read as descriptions or statements about a person are removed by rule, and some may remain. Report any to hello@bayanatlabs.com.
• For people who may be alive (latest possible birth year of nineteen hundred or later, or none known) a name pair is kept only when the two strings read as the same name; this also removes many real names, such as nicknames, translated names and full names with a middle name: 847,342 pairs, 20.4% of people's pairs.
• Columns on a person's row: qid, scope, entity_class, countries, ar_text_raw, ar_text_norm, ar_langs, ar_kinds, buckwalter, ala_lc, ala_lc_origin, din_31635, latin_text_raw, latin_fold, latin_langs, latin_kinds, method_qids, relation, skeleton_sim, token_alignment, alignment_score and lastrevid. lastrevid is the Wikidata revision number of the item in the dump, not data about the person.
• People who may be under eighteen are left out. The rule needs a birth date or a birth year in a description; 42,133 people have neither and could not be tested. Requests to remove an entry: hello@bayanatlabs.com.

Sources and licences

The names come only from the Wikidata JSON dump of 26 October 2026, read once as a stream from an official Wikimedia mirror (Academic Computer Club, Umeå University) with the Bayanat Labs open-data user agent, and never stored; the checksums computed during the read match the ones Wikimedia publishes for the file. Scope, string rules, the relation filter, the alignment and the benchmark are set out on the dataset card and in the repository's method section. Source details

How to cite

Cite as:

Bayanat Labs (2026). Arabic-Latin Name Variants, version 0.1.0 (Wikidata JSON dump of 2026-10-26). Bayanat Labs. https://huggingface.co/datasets/Bayanat-Labs/arabic-latin-name-variants

• CITATION.cff

No DOI. Cite the URL in the citation above and the date you read it.

@misc{bayanatlabs2026arabiclatinnamevariants,
  author       = {{Bayanat Labs}},
  title        = {Arabic-Latin Name Variants, version 0.1.0 (Wikidata JSON dump of 2026-10-26)},
  year         = {2026},
  month        = nov,
  howpublished = {\url{https://huggingface.co/datasets/Bayanat-Labs/arabic-latin-name-variants}}
}

How we handle citations

Changelog

Corrections and removal requests

If something in this release should not be there, or is wrong, email hello@bayanatlabs.com. We handle removal requests within 30 days, and platform requests within 24 hours. How removal requests work

Related product: Bayanat Search