How ONYRI Sanitize Finds Names and Company Names in a Document
How ONYRI Sanitize finds person and company names: rules and dictionaries, legal-form matching, an on-device model, and what still needs a manual mask.
ONYRI Sanitize finds names two ways. Rules and country dictionaries catch known first names, last names, and company names that carry a legal form. On a computer, a name-recognition model that runs on your device fills in what the rules missed.
How does it find personal names?
The first layer is rules and dictionaries. ONYRI Sanitize checks each word against first-name and last-name lists built for the document's country. A title right before a word, like Mr. or Dr., raises its score. So does a form label, such as Name: followed by a value, as in Name: Jane Example.
- French names, for France, Belgium, and Luxembourg
- English names, for the United States and the United Kingdom
- German names, for Germany, Austria, and Switzerland
- Spanish names, for Spain
- Italian names, for Italy
- Dutch names, for the Netherlands
A single common word rarely triggers a mask by itself. An ambiguous name is also an everyday word, but written with a capital letter, it is usually enough on its own. In lowercase, it needs a second clue nearby: a title, a form label, or another name next to it, before ONYRI Sanitize treats it as a person's name.
How does it find company names?
The second layer looks for a legal form right after a capitalized name, and sometimes right before it, as in SARL Nordval Renovation. Acme Holdings Ltd, Cascade Trust Bank LLC, and Müller & Fischer GmbH are all caught the same way. ONYRI Sanitize checks the legal forms used across the countries it covers.
What does the on-device name model add?
A name-recognition model completes the picture, on a computer with a mouse or a trackpad. It skips a phone, a tablet, and a slow or data-saving connection. It runs on your device. Its files come from ONYRI's own domain, the same place the OCR files come from, and they load quietly in the background.
While it works, the app shows “looking for more names…”. The export waits for it, up to 90 seconds. After that, or if the model cannot run here, the download goes ahead with rules and dictionaries alone. Nothing is blocked for good.
| Layer | What it catches | Where it runs |
|---|---|---|
| Rules and dictionaries | First and last names on file for the document's country, helped by a title or a form label | Every device, every plan |
| Legal-form matching | A company name followed by SAS, GmbH, Ltd, or another legal form, with a few French forms allowed before the name too | Every device, every plan |
| On-device name model | Names and company names the first two layers missed | Computer only, files from ONYRI's domain |
What happens once a name is found?
On a computer, once the on-device model recognizes Jane Example anywhere in the document, it masks that exact spelling everywhere else too, in every mode. The match respects capitalization, so JANE EXAMPLE printed in capitals on a stamp counts as a different spelling and may need a manual mask.
What does it miss, and how do you fix it?
No dictionary and no model is perfect. Detection is based on patterns and known names, not on understanding the document, so review the preview before you download. Once in a while, an ordinary word that matches a name pattern gets masked too. Uncheck it in the Detected data list to keep it visible.
- A nickname, or a name spelled in an unusual way
- A name written only in a photo, a logo, or a stamp
- Handwriting, which no layer reads
- A company name with no legal form, off a computer or if the model could not run
- A rare name in a country ONYRI Sanitize does not yet cover
To catch what is left, open the Masking tool above the pages. Search the document finds a word and, for three hits, offers “Mask all 3 results.” Select text lets you highlight a name by hand and choose “Mask all 3 occurrences” or “Only this one.” A one-off word can also go under Also mask, so every occurrence is caught from then on.
Frequently asked questions
Does the name-recognition model send my document to a server?
No. The model runs on your device, inside your browser. ONYRI Sanitize never uploads your file, its name, its text, or any value it finds. Only counters and statistics without content, like the number of pages and masks, reach the server, to apply your plan's limits.
Does the name-recognition model work on a phone or a tablet?
No. It needs a computer with a mouse or a trackpad, and it skips a slow or data-saving connection. On those devices, ONYRI Sanitize still masks names through rules, dictionaries, and legal-form matching, and you can finish by hand with the Masking tool.
What if a name is not in any dictionary?
On a computer, the name-recognition model often catches it anyway. If it still slips through, use Select text to highlight it and mask every occurrence, or use Search the document to find it everywhere. You can also add it under Also mask for next time.
Why is a company name sometimes left unmasked?
Most often, no legal form sits next to it, so the legal-form layer skips it. Acme Holdings alone is not enough. Acme Holdings Ltd is. On a computer, the name model often adds the bare name anyway. Otherwise, add it under Also mask.
Does this make a document anonymous under the GDPR?
Not on its own. Under Article 4 of the GDPR, a name is a direct identifier of a natural person. Token mode only replaces it with a label. Under Recital 26, the data can stay personal data if someone can link the label back. Masking reduces your exposure, but it is not a legal guarantee of compliance. Review the preview, and ask your data protection officer for a real compliance question.
Sources & references
- Article 4 GDPR — DefinitionsIntersoft Consulting (gdpr-info.eu)
- Recital 26 GDPR — Pseudonymised dataIntersoft Consulting (gdpr-info.eu)
- What is personal data?European Commission