arabnamer 0.1.3

Offline Arabic name transliteration and similarity — no LLM, no API calls, no network. Convert English names to Arabic and match Arabic name variants (hamza, tashkeel, taa-marbuta) 100% on your own machine. Built for KYC, compliance, and any workflow where names must never leave your infrastructure.

arabnamer — Arabic name transliteration & similarity

Offline Arabic name transliteration and similarity — no LLM, no API calls, no network.

from arabnamer import translit, similarity

translit("Mohammed Ali").arabic        # → 'محمد علي'
translit("Layla Al Saleh").arabic      # → 'ليلى الصالح'

similarity("أحمد حسن", "احمد حسن")      # → (True, 100)

Why offline matters

Names are personal data. Shipping them to Google Translate, OpenAI, Claude, or any cloud API means exposing PII to a third party — a hard compliance problem for finance, legal, healthcare, government, and MENA-region institutions bound by data-residency laws.

arabnamer solves both major Arabic-name problems on your own machine:

1. Transliteration — Mohammed Ali → محمد علي (the tricky ones: hamza variants, compound articles like El/Al/Abd, silent vowels, dialect spellings)
2. Matching — أحمد حسن ≡ احمد حسن ≡ أحمد حسن (scoring insensitive to hamza, tashkeel, taa-marbuta, and alef-maksura variants)

No model server, no internet, no API key, no request logs. Install once via pip, run anywhere — including air-gapped environments. The 38 MB pruned XGBoost model and 22,798-pair dictionary are bundled inside the wheel.

	Cloud APIs (Google, OpenAI, Claude)	arabnamer
Network required	✅ yes	❌ no — 100% offline
Names leave your infrastructure	✅ yes	❌ no
Per-call cost	metered	zero
Works in air-gapped / on-prem	❌ no	✅ yes
Model audit / replacement	❌ opaque	✅ open weights + retrainable
Accuracy (25-name MENA benchmark)	varies by model	98.0 avg, 23/25 pass ≥ 90

Built for: KYC / sanctions screening, compliance-gated entity resolution, library & archive cataloguing, Arabic NLP preprocessing, on-premise search relevance — any workflow where names must never leave your infrastructure.

---

Install

pip install arabnamer

Works offline after install — the model is bundled (gzipped, ~38 MB). No external services, no API keys.

---

Quick start

1. Transliterate English → Arabic

from arabnamer import translit, translit_batch

result = translit("Mohammed Ali")
print(result.arabic)    # 'محمد علي'
print(result.score)     # 0.0 (no reference supplied)
print(result.engine)    # 'xgboost'

# batch
results = translit_batch(["Ahmad Hassan", "Fatima Mansour", "Layla Al Saleh"])
for r in results:
    print(f"{r.input:<25} -> {r.arabic}")

2. Score with a reference (lenient, normalized)

from arabnamer import translit

r = translit("Ahmad Hassan", reference="أحمد حسن")
print(r.score)       # 100.0
print(r.accepted)    # True  (>= default threshold 85)

3. Arabic ↔ Arabic similarity (tashkeel / hamza / taa-marbuta insensitive)

from arabnamer import similarity

similarity("أحمد حسن", "احمد حسن")          # (True, 100)   hamza variant
similarity("مروة فرج", "مروه فرج")           # (True, 100)   taa-marbuta vs haa
similarity("محمد", "أحمد", threshold=75)     # (True, 75) — different names

4. Configurable: threshold + engine

from arabnamer import Transliterator

t = Transliterator(engine="model", threshold=90)   # stricter pass bar
t.translit("Mohammed Ali", reference="محمد علي")

# Rule-based engine (deterministic, dict-first)
t_rules = Transliterator(engine="rules")
t_rules.translit("Mohammed Ali")

# Hybrid: dict -> model -> rules fallback
t_hybrid = Transliterator(engine="hybrid")

---

How accurate is it?

Benchmarked on 25 generic Arab-name pairs (common first + last combinations covering compound articles, hamza variants, feminine endings):

Metric	Score
Average lenient similarity	98.0
Pass rate (≥ 70)	25 / 25
Pass rate (≥ 90)	23 / 25
Exact match (= 100)	20 / 25

Model: XGBoost, 386 boosting rounds × 335 output classes, 34 input features per character (char IDs + position + phonetic class + bigram/trigram IDs). See benchmarks/REPORT.md for the full breakdown.

---

What's inside this repo

arabnamer/
├── src/arabnamer/           # pip-installable library
│   ├── core/                # Transliterator orchestrator + Result
│   ├── prediction/          # XGBoost model loader + featurize
│   ├── rules/               # deterministic rule-based walker
│   ├── scoring/             # fuzzy match + Arabic normalizer
│   └── utils/               # tokenizer + dict lookup
│
├── model/                   # gzipped model + labels (ships in wheel)
├── dataset/                 # 22,798 EN -> AR name pairs (dict_FINAL.json)
├── training/                # scripts to retrain from scratch
├── tests/                   # smoke + benchmark eval
├── benchmarks/              # eval results + REPORT.md
└── docs/                    # data sources + architecture + API details

---

Data sources

Training data comes from three open sources plus a manual audit pass. Full provenance in docs/data_sources.md. In summary:

Source	License	Contribution
JRC-Names (European Commission, multilingual name gazetteer)	EU open-data	primary EN/AR name pairs
Google Translate (via deep_translator)	generated text	fill-in for names absent from JRC
Claude (Anthropic) LLM fill	generated text	supplementary fill for DI-thesis-specific names
Manual audit + rule-based cleanup	—	phonetic-compatibility filter, hamza repair, outlier removal

Model and dataset are released under CC-BY-4.0 (attribution required). Library code is MIT.

---

Reproducing the model

git clone https://github.com/sayedyousef/arabnamer
cd arabnamer
pip install -e ".[dev]"
cd training
python train_xgboost.py      # ~5 min on CPU, produces ~285 MB UBJ
python prune_trees.py        # ~1 min, saves pruned variants
python find_min_k.py         # ~30 sec, finds the smallest-identical model

XGBoost with multi-threaded histogram training is approximately (not bit-perfectly) deterministic. See training/README.md for the full note.

---

Why this library exists

Arabic name handling is a real gap in open NLP tooling. Most Arabic libraries (PyArabic, arabic-reshaper, AraBERT, CAMeL Tools) focus on text — rendering, tokenisation, stemming. Proper-noun transliteration and fuzzy-matching of name variants (أحمد vs احمد vs Ahmad) is underserved.

arabnamer is the first open library specifically for:

• Arabic ↔ English name transliteration at the word level
• Arabic name similarity that's insensitive to hamza, tashkeel, taa-marbuta, and alef-maksura variants
• Offline, pip-installable — no API keys, no network, no LLM dependency

Typical use cases:

• KYC & sanctions screening — match Arabic names against English watchlists
• Library & archive cataloguing — normalize author names across MENA languages
• Search relevance — expand queries to cover Arabic-script variants
• Data integration — reconcile person entities across EN/AR CRMs

---

Citation

If you use arabnamer in a paper, product, or research tool, please cite:

@software{yousef_arabnamer_2026,
  author  = {Yousef, Elsayed},
  title   = {arabnamer: Arabic name transliteration and similarity},
  year    = {2026},
  version = {0.1.0},
  url     = {https://github.com/sayedyousef/arabnamer},
}

A CITATION.cff file is included so GitHub's "Cite this repository" button works.

---

Commercial support

For production deployments, custom models trained on your domain corpus, on-premise integration, or paid support:

[e-mail]

---

License

• Source code (src/arabnamer/, training/): MIT License
• Dataset, model weights, test data: CC-BY-4.0 (attribution required)

---

Acknowledgements

• JRC-Names — European Commission Joint Research Centre name gazetteer (primary training data)
• XGBoost — gradient-boosted tree library
• RapidFuzz — fast fuzzy string matching
• The thesis work this library was extracted from (Doha Institute for Graduate Studies)