Beyond English: Semantic Matching Across Languages

01The failure is invisible, which is what makes it expensive

A recruiter who receives no German applicants for a Berlin role does not conclude that their screening is language-bound. They conclude there are no German applicants. The absence of a signal looks identical to the absence of candidates, and nothing in a conventional pipeline distinguishes them. This is why the problem persists in organisations that would fix it immediately if they could see it.

It compounds in the direction you would least want. The candidates most likely to write a CV in their own language are the ones least embedded in an English-speaking professional network — which is to say, precisely the people a company opening a new market is trying to reach. The screening artefact filters hardest against the population the expansion exists to hire.

02What reading across languages actually requires

Two things have to be true, and they are independent. The system must be able to *read* a document in one language against a requirement written in another, and it must *report* what it found in the language the reader works in. Systems that translate a CV into English before scoring it satisfy neither cleanly: translation loses the evidence trail, and a recruiter reading translated text cannot tell which words the candidate actually used.

Expertini's screening layer reads both documents in their original languages and extracts evidence per competency, citing what the source document said. A German CV assessed against an English job description yields English evidence that references the German original. The candidate is never asked to write in the employer's language; the employer is never asked to read in the candidate's.

The direction reverses cleanly too. An English CV against a German job description produces German evidence for the German-speaking hiring team, with the same competencies recognised. This is not translation in either direction — it is extraction against a requirement, performed in a model that holds both languages at once.

03Why the score is deliberately not the clever part

It would be straightforward to ask a language model for a match percentage and present it. Expertini does not, and the reason matters more in a multilingual setting than a monolingual one. The Candidate Match Score is computed by a fixed published formula from the extracted evidence — AI reads, arithmetic scores. Same evidence, same number, every time.

In one language that property buys reproducibility. Across languages it buys something stronger: it makes the language question separable. If a German candidate scores lower than an English one, that difference is traceable to which evidence was found, not to how a model happened to feel about the two documents. You can inspect the dimensions, see what was and was not evidenced, and argue with it. A single generated number gives you nothing to argue with, and gives a regulator nothing to audit — see auditable hiring reports.

It also bounds the blast radius of a translation problem. If cross-lingual reading misses something, it shows up as missing evidence on a named dimension, which is visible and correctable. It does not silently redistribute weight across the whole score.

04The limits, stated plainly

Cross-lingual extraction is a model capability, not a guarantee. It is strongest for languages with abundant training data and weaker for those without, and Expertini's 75 supported languages are emphatically not equal in this respect. A CV in a widely-written language will be read more reliably than one in a language with a thin digital corpus. We would rather name that asymmetry than let a language count imply uniform quality.

Domain vocabulary is the sharpest edge. Professional and legal titles are often untranslatable rather than merely untranslated — a German *Meister*, a French *cadre*, a Japanese 主任 carry structural meaning about seniority and qualification that has no clean English equivalent. A system that maps them to the nearest English word loses exactly the information a hiring manager needed. Where the evidence is genuinely ambiguous, the honest output is an unevidenced dimension a human can look at, not a confident guess.

And none of this removes the human. Cross-lingual screening widens the pool a hiring team can see; it does not decide who to hire, and our own published methodology says so explicitly.

05What this means for hiring in more than one market

The practical consequence is that language stops being a filter applied before assessment. A company hiring in six countries can run one requirement against candidates who wrote in six languages, and read the results in one. The alternative — a separate localised pipeline per market, or an implicit English-only requirement — either multiplies the operational cost or narrows the field to people comfortable in a second language.

That second option is worth naming for what it is. Requiring a CV in English is a proficiency filter, applied at the earliest stage, to a skill that many roles do not need. Sometimes it is a legitimate requirement. Frequently it is an artefact of the tooling, and it should not be mistaken for a hiring standard.