Marawanelbalal/pii-phi-corpus-final

Multilingual PII/PHI De-identification Corpus (Final)

This is the finalized, training-ready corpus for a multilingual PII/PHI de-identification (token classification) model scoped to en-GB, nl-BE, fr-FR. It merges two sources:

1. Our own synthetic clinical/administrative corpus (Marawanelbalal/synthetic-pii-phi-v2), unchanged.
2. A normalized subset of ai4privacy/pii-masking-openpii-1.5m (English/French/Dutch rows only), added as generic-register training text.

development/test are untouched — both splits are exactly our own corpus's leak-safe development/test sets, so evaluation numbers stay comparable across releases. All OpenPII rows are folded into train only.

Attribution (OpenPII)

ai4privacy/pii-masking-openpii-1.5m is licensed CC-BY-4.0 by Ai4Privacy / Ai Suisse SA. This release uses upstream revision a785eb528e28be2693c3718a27e066970de5dadb. Per that license's attribution requirement:

Ai4Privacy. OpenPII 1.5M: Multilingual PII Masking Dataset with Asia Pacific Extension. 2026. https://huggingface.co/datasets/ai4privacy/pii-masking-openpii-1.5m Licensed under CC-BY-4.0.

Changes made to the subset of OpenPII included here (English/French/Dutch rows only, out of its full 30-language, ~1.6M-row release) are documented in full below, per the license's requirement to indicate changes.

Why OpenPII needs normalization

OpenPII has a different annotation vocabulary. This dataset uses one fixed label taxonomy. The normalization keeps only labels that map to that taxonomy. It does not rewrite OpenPII text to imitate a target locale.

What was actually done

1. Name/title normalization (all rows, all languages). OpenPII tags given names, surnames, and honorifics as three separate labels (GIVENNAME, SURNAME, TITLE); this taxonomy has one NAME label and no title label at all. Adjacent runs of these three labels (gap <=2 characters — just whitespace, a period, or a hyphen between them) were merged into a single NAME span covering only the non-title parts. TITLE was dropped entirely rather than folded in, since an honorific alone isn't identifying. A lone GIVENNAME or SURNAME with no adjacent partner still became its own NAME span. Result: 55,878+ merges across the corpus, 37,386+ titles dropped, verified against hand-read samples (compound surnames, 3-part names, and title-stripped contexts all read correctly).

2. Native scheme-cue relabeling. A generic OpenPII identifier is mapped to NHS_NUMBER, INSZ, or NIR only when the unchanged source text names that scheme. The normalizer selects the nearest unclaimed identifier span. It does not replace Social Security, SSN, National Insurance, Burgerservicenummer, BSN, or any other source term. There is no NATIONAL_INSURANCE_NUMBER label in this release.

3. Canonical label mapping. Every remaining OpenPII raw label was mapped onto this taxonomy's exact label names (TELEPHONENUM -> PHONE_NUMBER, ORGANISATION -> ORGANIZATION_NAME, etc.), or dropped if it has no canonical equivalent (CREDITCARDNUMBER, DRIVERLICENSENUM, PASSPORTNUM, TIME, USERNAME, COUNTRY, AMOUNT, CURRENCY, BANKNAME, ACCOUNTNUM, and unsupported generic identifier spans). The source text stays unchanged when a span is removed. This prevents labels outside the taxonomy from entering the model label space. No clinical labels (DIAGNOSIS, PROCEDURE, SYMPTOM, MEDICATION, DOSAGE, lab/ vitals labels) exist in OpenPII at all — its source text is generic multi-domain narrative (business, travel, admin correspondence), not clinical content, so this dataset adds PII-detection genre diversity, not clinical vocabulary.

4. Annotation-boundary correction. Six inherited STREET spans in the own-corpus train split overlapped a following SYMPTOM or ZIPCODE span. Each was shortened to the street-name boundary before this corpus was rebuilt. No text, document, label, or other span changed.

What was deliberately NOT done

• OpenPII text was not rewritten. Addresses, number formats, spelling, and locale-specific terms stay as generated by OpenPII.
• No new taxonomy label was added for UK National Insurance Numbers. A generic identifier that does not map to an existing label stays in the text without an annotation.
• Identifier values were not regenerated to satisfy national checksum rules.

Splits

split	rows
train	(see manifest.json — own corpus train + all normalized OpenPII rows)
development	(unchanged from the base release)
test	(unchanged from the base release)

See manifest.json's composition field for the exact own-corpus-vs-OpenPII row breakdown within train.

Schema

• source_text (str): the full document text.
• privacy_mask (list[{start, end, label, value}]): character-offset spans with their entity label and the substring value.
• language (str): locale code (en-GB, nl-BE, fr-FR).
• source_dataset (str): provenance tag — either Marawanelbalal/synthetic-pii-phi-v2 or ai4privacy/pii-masking-openpii-1.5m, so the two sources stay distinguishable in train.
• dataset_split (str): train / development / test.
• multilingual_taxonomy.json: the 32-label, 65-class BIO training profile for this corpus.

Sources

• Marawanelbalal/synthetic-pii-phi-v2 — see that dataset's own README for its full generation methodology.
• ai4privacy/pii-masking-openpii-1.5m (CC-BY-4.0) — English/French/Dutch rows only, normalized as described above.