Why

Presidio ships recognizers for US, UK, Spanish, Italian, Indian, Singaporean and Australian identifiers. It ships none for the Maghreb. If you run Presidio over Moroccan documents today, a national ID card number, a bank account or a company identifier passes straight through.

This package adds them, following Presidio's own conventions so nothing else in your pipeline has to change.

Installation

pip install presidio-mena

Usage

from presidio_analyzer import AnalyzerEngine
from presidio_mena import register_mena_recognizers

analyzer = AnalyzerEngine()
register_mena_recognizers(analyzer)

text = "Client BE12345, tel 0612345678, RIB 012345678901234567890104"
results = analyzer.analyze(text, language="en")

for result in sorted(results, key=lambda r: r.start):
    print(f"{result.entity_type:<18} {result.score}")
# MA_CIN             0.5
# US_DRIVER_LICENSE  0.3     <- built-in, outranked
# MA_PHONE_NUMBER    0.85    <- raised by the context word "tel"
# PHONE_NUMBER       0.4     <- built-in, outranked
# US_BANK_NUMBER     0.05
# MA_RIB             1.0     <- check key confirmed

Presidio's own recognizers still fire on the same spans — that is normal, and the highest score wins. Because the MENA recognizers are country-specific and checksum-validated, they outrank the generic ones on Moroccan data.

It composes with the rest of Presidio exactly as you would expect:

from presidio_anonymizer import AnonymizerEngine

anonymized = AnonymizerEngine().anonymize(text=text, analyzer_results=results)
print(anonymized.text)
# Client <MA_CIN>, tel <MA_PHONE_NUMBER>, RIB <MA_RIB>

Registering for several languages, or into a hand-built registry:

from presidio_analyzer import RecognizerRegistry
from presidio_mena import register_mena_recognizers, mena_recognizers

registry = RecognizerRegistry()
registry.load_predefined_recognizers()
register_mena_recognizers(registry, languages=["en", "fr"])

# or take the instances and wire them yourself
for recognizer in mena_recognizers(supported_language="fr"):
    registry.add_recognizer(recognizer)

Individual recognizers can be used on their own, and every parameter Presidio exposes — patterns, context words, entity name — can be overridden:

from presidio_mena import MoroccanRIBRecognizer

recognizer = MoroccanRIBRecognizer(supported_entity="BANK_ACCOUNT_MA")

The check digit is the point

A pattern alone cannot distinguish a bank account from an asset tag: both are just digits. So the RIB and IBAN recognizers score their pattern low, then Presidio calls validate_result(), which runs the real checksum:

• a confirmed value is raised to full confidence;
• a value that fails the checksum is discarded entirely.

In practice, of 200 arbitrary 24-digit numbers, at least 190 are correctly rejected — only about 1 in 97 can carry a valid key. Your inventory references survive an anonymization pass; the bank details do not.

graph LR
    TXT["Document<br/><i>RIB 0123...0104</i>"]

    subgraph PRESIDIO["Presidio AnalyzerEngine"]
        BUILTIN["Built-in recognizers<br/>EMAIL · CREDIT_CARD · US_SSN ..."]

        subgraph MENA["presidio-mena"]
            PAT["Patterns<br/>MA_CIN · MA_PHONE_NUMBER<br/>MA_RIB · MA_IBAN · MA_ICE"]
            VAL["validate_result()<br/>mod-97 · MOD-97-10<br/><b>drops look-alikes</b>"]
            CTX["Context words<br/>FR · AR · EN"]
        end
    end

    OUT["RecognizerResult[]<br/>entity · span · score"]
    ANON["AnonymizerEngine<br/><i>RIB &lt;MA_RIB&gt;</i>"]

    TXT --> BUILTIN --> OUT
    TXT --> PAT --> VAL --> CTX --> OUT
    OUT --> ANON

    style MENA fill:#e8f0fe,stroke:#1a73e8,stroke-width:2px
    style VAL fill:#fce8e6,stroke:#d93025,stroke-width:2px
    style PRESIDIO fill:#f8f9fa,stroke:#5f6368
    style ANON fill:#e6f4ea,stroke:#1e8e3e,stroke-width:2px

Multilingual context

Moroccan documents mix French, Arabic and English in the same paragraph, so the context words that boost a score do too — rib, compte, banque, حساب, account. A CIN found next to carte nationale scores higher than the same string found alone.

Presidio's context enhancer compares single token lemmas, so every context entry is one word; multi-word phrases would never match.

Limitations

• CIN and ICE have no publicly specified check digit, so they rely on shape and context. Expect lower precision on those two than on RIB or IBAN — review a sample before trusting an unattended run.
• Only Moroccan identifiers ship today. Algerian, Tunisian and Gulf identifiers are the obvious next step and contributions are very welcome — the package is named mena, not morocco, on purpose.
• Names of people and prose addresses are Presidio's NER job, not this package's. Pair these recognizers with a suitable NLP engine.

Compatibility

Tested end to end against a real AnalyzerEngine and AnonymizerEngine, not just in isolation.

Contributing

New MENA recognizers are the most useful contribution — see CONTRIBUTING.md.

git clone https://github.com/novagiosAI/presidio-mena.git
cd presidio-mena
python -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -e ".[dev]"
pytest

License

Apache License 2.0 — free for commercial use, same as Presidio.