Spanish-language and multilingual records: why PII tools miss names

Why do PII tools miss names in Spanish and bilingual records?#

PII tools miss names in Spanish and bilingual records because name detection usually comes from a model trained mostly on English text, and the supporting rules assume English wording. A model that has learned to expect a surname after Mr. has little to work with when a customer writes habló la señora Pérez, or texts a name in lowercase.

Patterned identifiers hold up better. Email addresses, card numbers and US phone numbers look the same in any language, so regular expressions still catch them. Coverage drops on names, street addresses and free-text descriptions of people, which are exactly what fill technician notes and customer messages in a home services company.

Open-source tools make the design easy to see. Presidio, the open-source toolkit now run as a community project under the Data Privacy Stack organization, combines a named-entity model with pattern rules, checksums and context words. For names, the model does most of the work, so its training language matters most. Context words, the nearby terms that raise confidence in a match, are usually configured per language as well, so an English list adds nothing on a Spanish note. The project itself warns that automated detection cannot guarantee it finds all sensitive information, and that warning weighs heavier outside the language a pipeline was tuned for.

Where names hide in a contractor's job records#

Spanish shows up across the whole job record, not only in a separate Spanish queue. In a contractor running ServiceTitan, Housecall Pro or Jobber, bilingual customer service reps type booking notes, technicians add job notes from a phone, and customers text back in whichever language is easier for them.

The table maps the usual record types to the specific way names slip through. Use it to decide where your bilingual reviewers spend their time.

Five name patterns that English-tuned tools get wrong#

Five name patterns come up again and again when a bilingual reviewer checks an English-tuned scanner's output. Each one has a specific fix, so test for them by name rather than hoping a general model improves.

• 1Given names that are also everyday words, such as Rosa, Luz, Blanca, Paz, Dolores and Ángel. A word-list rule over-redacts la luz de la cocina, the kitchen light, while a weak model misses Luz llamó otra vez, Luz called again.
• 2Double surnames and particles, as in María Fernanda García López or Juan de la Cruz. Tools often mask the first surname and leave López or de la Cruz, which can still point to one household in a small service area.
• 3Missing accents and lowercase. Exports and texting gateways often drop accents, and customers type jose or maria. Rules that depend on capitalization or exact spelling lose them.
• 4Titles and kinship words: don Ramón, doña Elena, la señora Ruiz, su esposo Beto. These are strong signals in Spanish, but only when the context word list includes them.
• 5Code-switching inside one sentence, such as Customer said que su suegro Arturo will be home after lunch. A detector run on the whole note may label it English and send it to the English model, which skips the Spanish clause.

A checklist for mixed-language records#

The fix for mixed-language records is a pipeline that treats language as a property of each piece of text, not of the dataset. The steps assume records are exported with a stable record ID, so every finding can be traced back to its source.

One step does more than the rest: use your own customer table as a dictionary. The customer master in your field service system already holds the names most likely to appear in notes, in every spelling customers gave you. Route matches on names that are also everyday words to a reviewer instead of masking them automatically, or the kitchen light disappears along with Luz.

• 1Split records into units small enough to carry one language: a message, a note, a field, or a sentence in long notes.
• 2Detect the language of each unit, store a confidence value, and send low-confidence or mixed units through both pipelines.
• 3Run a name and location model trained for Spanish on Spanish units and the English model on English units.
• 4Add Spanish context words and titles to your rules: me llamo, mi nombre es, se llama, señor, señora, don, doña, teléfono, dirección, calle and avenida.
• 5Extend surname rules so a match expands across particles and a second surname.
• 6Match against a copy of the text with accents and case stripped, then map each finding back to the original.
• 7Run known customer and employee names from the field service customer list, CRM and payroll as an exact-match deny list against every unit, matching on the accent-stripped copy.
• 8Draw a separate human review sample from each language slice and record misses per slice.

How to review the Spanish slice#

The Spanish slice needs its own reviewers and its own pass mark. A blended result can look acceptable while many Spanish notes still carry a name, simply because English notes outnumber them.

Review samples should be larger relative to the slice for languages your pipeline handles less well. You are measuring a process you trust less, so you need more evidence before you sign off. Fix the release threshold before anyone reads a sample, so a weak Spanish result cannot be talked up after the fact.

Illustrative: a bilingual HVAC and plumbing contractor#

Illustrative: a fictional HVAC and plumbing contractor in the Southwest runs ServiceTitan, with a bilingual call center and technicians who write notes in both languages. Before a planned export, its IT lead ran an English-tuned scanner over booking notes, job notes and customer texts.

The English sample looked clean. A bilingual reviewer then read a separate sample of Spanish and mixed notes and found names such as Rosa and Ángel left in place, second surnames remaining after the first was masked, and relatives named after su hijo or su suegra. The team added per-note language detection, a Spanish model, Spanish context words, the customer master as a deny list and surname expansion. A fresh Spanish sample met the same release rule as the English one, and the company kept its Spanish notes in scope instead of dropping them.

How SourceX reviews language slices#

SourceX reviews mixed-language records one language slice at a time, starting before any file moves. At the fit check stage the only question is descriptive: which systems hold Spanish or mixed-language text, and roughly what share of the records it covers.

If the records proceed, Preparation reports review results for each language slice, so a gap in one language is visible rather than averaged away. The privacy record in the SourceX Evidence Packet states which languages were reviewed, by whom and against what release rule, and the supplier approves that result before anything moves to Delivery in the SourceX five-step transaction.