Anonymizing a document in another language: names, dates and numbers that change

In another language, the method stays the same: list what identifies people, search for it, mask it, then check the result. The forms change, though: a name can change its ending, order or script, and a date can use another calendar or other digits. A search for one exact form misses the others, so search for every form.

Why does a name search miss names in another language?

A name is personal data in every language and script. Article 4(1) of the GDPR, the EU data protection regulation, lists a name and an identification number as examples of identifiers. The patterns below follow the W3C's article on personal names, which warns that real names are often more complicated.

How do endings and gender change a surname?

Imagine Jane Example hiding names in a Polish contract. She searches for Kowalski and finds the signature line. She misses the clauses that say Kowalskiego or Kowalskiemu. Polish and Czech names change with their role in the sentence, which is called declension. A study in Polonica, a journal of the Institute of the Polish Language, says surnames ending in ski, cki or dzki decline like adjectives.

Gender adds a second change. In Russian, the W3C explains, the endings of the patronymic and family name show whether the person is a man or a woman. In Czech, a woman's surname normally takes a feminine form, as Nováková does for Novák. The Czech Ministry of the Interior says that since 1 January 2022 a woman may ask the civil register for the unchanged male form. She needs no special reason, so search both forms.

Search the stem, the part that does not change. Kowalsk finds Kowalski, Kowalskiego and Kowalska. Novák finds Nováka and Nováková. Read every hit, because a short stem also finds other words. Search the base form too: the study's author notes that legal texts and official letters most often use it. A stem search is a net, not a guarantee.

What if one person is spelled two ways?

A passport adds a second spelling. ICAO, the International Civil Aviation Organization, sets the passport standard in its Doc 9303. In the machine-readable zone (the lines scanners read), names use only the letters A to Z, and ICAO recommends several choices for one letter. So Jörg Müllerhoff can read MUELLERHOFF and JOERG, or MULLERHOFF and JORG. For Russian, ICAO turns Ж into ZH and Щ into SHCH.

The spelling also depends on who writes it. ICAO's appendix counts 32 ways to write the single name Muhammad in Latin letters. It also notes that English writes Omar Khayyam where German writes Omar Chajjam. In Japanese, the W3C adds, four different surnames can all be written Shōji in Latin letters, so the Latin form hides which one is meant.

How do dates, numbers and addresses change?

The European Patent Office (EPO) writes that imperial years appear everywhere in Japan, especially on official documents. Thai patents, it adds, show their dates in Buddhist years only. Here is 8 October 2026 in five common forms.

Do the sum before you search: the EPO gives Reiwa plus 2018, or a Buddhist year minus 543, for the Western year. Unicode's locale data lists Arabic-Indic and Persian digits as separate characters, which is why a plain search for 2026 finds neither.

Arabic documents may also use the Hijri calendar, which follows the moon. The EPO says a Hijri year cannot be converted directly.

Addresses can run backwards. The Universal Postal Union (UPU) says a Japanese address is generally written from the largest area to the smallest, and street names are not used. A Chinese address written in Chinese starts with the province, while the Latin-letter version runs the other way. Mask the whole address block, postcode included.

How do you check a scanned document in another language?

A plain scan holds no text until OCR (optical character recognition) reads it, and OCR needs the right language. The Tesseract documentation says you choose one language or several joined by a plus sign, and that English is assumed if you choose none.

Right-to-left text adds a trap. The W3C explains that text is stored in logical order but displayed in visual order. Numbers inside Arabic text still run left to right. So test before you trust: search for a name or number you can see. If nothing comes up, do not trust the search on that page. Pages that mix scripts need one test per script.

ONYRI Sanitize runs in your browser. It finds ID numbers, phone numbers and dates from 47 countries, and postcodes and street addresses from nearly all of them (there is no street-address detector for the UK or Belgium). It also reads Arabic, Persian and other digit systems. Its on-device OCR reads English, French and German, plus your browser's language if it is one of about thirty others. Limit: in Polish, Czech, Russian, Japanese or Korean, a bare name is masked only after a title or label, or in a signature.

What is a step-by-step method for a document you cannot read?

1. Find a reader who knows the language.
2. List every person and company, with every form: original script, Latin letters, passport spelling, short forms, gender endings.
3. Search each form, then the stem. Read every hit in its sentence.
4. Search numbers in every digit system. Read dates in the document's calendar.
5. Mask each address as one block. For scans, search a word you can see, in each script.
6. Mask by hand what no search finds: signatures, stamps, handwriting, names without a title.
7. Check the exported file, not the screen. Look at every page for each form, and search it too if the text is still selectable.

Frequently asked questions

Sources & references