Multilingual Assessment for International Training Providers: How to Translate and Adapt Exams

Multilingual assessment, as this guide uses the term, means delivering the same exam to learners in more than one language so that a pass means the same thing in every version. It is not language-proficiency testing, which is what most academic results for the phrase cover. This guide is for training providers who run exams for learners in several languages or countries.

The short answer: pick a design that matches the stakes (translate, adapt, develop separately, or offer a language choice), run a documented review and pilot, and do not compare scores across languages until you have evidence they are comparable. The International Test Commission (ITC) publishes the reference method, and as of October 2026 its second edition (cited as 2017) is the newest edition the ITC lists.

Below you will find a decision table, a seven-step workflow tied to the ITC guidelines, practical notes on scripts and formats, and an honest look at item bias and pass marks when your volume per language is small.

Key takeaways

• A translated exam is a new version of the test, so it needs its own evidence, and a fluent translation alone does not show it measures the same thing.
• Choose the design by stakes and by how much the job content differs between countries: course quizzes need far less evidence than a certification exam.
• Back translation is a spot check, not a validation method, and the ITC Guidelines say so directly.
• Until you can show the language versions are comparable, report results by version and set the pass mark for each version.

What is multilingual assessment for a training provider?

For a training provider, multilingual assessment means the same exam, built to the same blueprint, is available in more than one language, and its results can be trusted in each. Three things can vary by language: the interface, the instructions and the items. Each needs its own decision.

Several vendor help pages on this topic stop at the first of those. They explain how to switch button labels, which is useful, but it says nothing about whether a question about a safety procedure is equally clear and equally hard in Spanish and in English. That second question is the one a client, regulator or accrediting body will ask.

Should you translate, adapt, develop separately or offer a language choice?

Choose by stakes and by how much the job content changes between countries. Translate when content is identical and stakes are low to medium, adapt when only references differ, develop separately when local rules differ, and offer a language choice only among versions you have already validated.

The table below shows the five approaches, when each fits and the minimum evidence to keep. The stakes column is editorial judgment, not a quotation from a standard.

The ITC calls adaptation “the broader term” and says translation “has a more limited meaning.” Those guidelines were written for tests in general, including psychological and educational instruments, so scale the effort to your stakes. A course quiz does not need a bias study. A credential that affects someone’s job does.

Which language should the interface, instructions and exam items use?

Treat them as three separate settings. The interface, the instructions and the items can each use a different language, and each should follow a written rule. Letting a learner choose the interface language is usually harmless. The item language is a measurement decision tied to what the job requires.

Instructions deserve more care than they get. The ITC’s administration guidelines list the clarity of translated test instructions as something to check, and suggest rewriting instructions in a standardized way that suits the new language and culture rather than translating them word for word.

Among SimpliAssess’s published pages, two news posts cover languages. One says the platform is available in Spanish, with a Spanish interface, a language switch in account settings, and educators able to create assessments and analyze results in Spanish. The other says users can build assessments in Portuguese. Both pages show only a last-modified date of July 22, 2024 and no publication date. Neither says whether the Portuguese interface is translated, how language versions of one exam are linked, or anything about right-to-left scripts, so confirm those details in a trial before you rely on them.

The same ideas apply to feedback surveys. SimpliAssess’s July 2024 post on multilingual surveys recommends professional translators for idioms and layouts adjusted for text length and reading direction.

What does a defensible workflow for multilingual assessment look like?

A defensible workflow has seven steps: choose the design, translate twice and reconcile, review with bilingual subject experts, use back translation only as a check, pilot with real learners, test equivalence when volume allows, and document every change. The ITC Guidelines for Translating and Adapting Tests set out 18 guidelines in six areas.

The table maps each step to the guideline it comes from and the record to keep. The step order is a practical sequence, not the guidelines’ own numbering.

Back translation tempts teams because it feels like a test. The ITC warns that in its narrowest form the translated version itself is never reviewed, and that rarely would translation designs alone provide sufficient evidence to validate an adapted test. A literal translation can back-translate perfectly and still read awkwardly to the learner.

For a worked example from a high-stakes exam, NCSBN says its Canadian French translation of the NCLEX follows the ITC Guidelines. Approved items are translated, a panel of three to six bilingual nurses reviews them, and translated items are monitored and, once enough data exists, analyzed for DIF.

Pro tip: Freeze the Source Before You Translate. Lock the English items and give each a permanent ID that every language version shares. If the source changes after translation starts, you get silent mismatches that no later check will catch.

How do you handle right-to-left scripts, date and number formats, and glossaries?

Set language and direction in the page markup rather than patching them with styling, fix one date and number format per language version, and keep a bilingual glossary so each technical term is translated the same way everywhere. Then test every version on the devices learners actually use.

Practical checks to run on each version:

• Direction: W3C advises adding dir=”rtl” to the html element when the whole document is right-to-left, and using the dir attribute on individual blocks for mixed content. It cautions that automatic direction detection is not foolproof.
• Language tags: WCAG 2.2 success criterion 3.1.2 (Level AA) asks that the language of each passage can be programmatically determined, which matters for items that mix languages.
• Dates and numbers: write the month as a word, because 03/04 means different dates in different countries, and check decimal commas and periods in numeric answers.
• Glossary: list terms to translate, terms to leave in the source language, and approved translations, and share it with every translator and reviewer.
• Length and timing: translated text is often longer or shorter, so check that layouts hold and that the time limit is fair for the longer version.

How do you check for item bias across language versions?

Compare how learners of similar overall ability perform on each item in each language. That is differential item functioning, or DIF. A flagged item needs review by bilingual experts, not automatic deletion, and DIF analysis needs enough candidates per version, which many training providers do not have.

The ITC defines DIF as test takers of equal trait level from different groups having different response probabilities on an item, and says Mantel-Haenszel “may be the most straightforward” method for right-or-wrong items. It also sets a floor: studies to identify potentially biased items need “a minimum 200 persons per version.” The Standards for Educational and Psychological Testing (AERA, APA and NCME, 2014) treat test adaptation under fairness in testing, in Chapter 3.

Here is the arithmetic for a hypothetical provider. If 40 learners a year take the Portuguese version, reaching 200 takes five years. That is why a small provider should lean on the other tools: expert review, the pilot, and item analysis run separately for each language. Look for items whose difficulty differs sharply by version and for distractors that attract learners in one language only.

Treat any gap in pass rates between languages as a prompt to investigate, not proof of bias. The ITC says to interpret group score differences with reference to all relevant available information. Groups can differ in experience or prior training, not only in how the items work.

Can you use the same cut score for every language version?

Only after you have evidence the versions are comparable. The ITC Guidelines say to compare scores across populations only when the level of invariance has been established. A cautious default is to treat each language version as its own form with its own pass-mark decision until you can show otherwise.

The ITC warns that linking score scales across versions “will always be problematic” and calls the assumption that versions are equivalent and can be used interchangeably “highly problematic.” The Standards deal with score linking and cut scores in Chapter 5. Three practical options follow, from most to least cautious:

• Separate panels: run a standard-setting study for each language version with judges who can read that version.
• One panel, several versions: use bilingual judges who review the source and each translation, then record a cut score per version.
• One shared cut score: keep it only for identical translated items with documented equivalence evidence, and review it as data arrives.

Our post on setting cut scores for professional training courses walks through a modified Angoff study you can repeat per version.

Pro tip: Report Results by Language Version From Day One. Add the exam version to every result record before the first translated cohort sits. When someone asks about fairness later, you can answer from your own data instead of rebuilding it.

What should you do next?

Pick one exam and one target language, starting with your highest-volume language pair. Fill in the decision table for that pair, write down why you chose translate, adapt or develop separately, and name the two translators and the bilingual reviewers before any work starts.

Then run the pilot with a small group of real learners, interview a few, and keep every change in a log. That file is what you will show a client or accrediting body, and few vendor help pages describe it.

If you also deliver the training itself, SimpliTrain, a separate training platform from the same company, has a page on multi-language support that says quizzes, tests and certifications can be offered in multiple languages. It does not describe the translation method or right-to-left support, so ask about both.

Frequently asked questions

Is back translation enough to validate a translated exam?

No. The ITC Guidelines say rarely would forward or backward translation designs alone provide enough evidence to validate a translated and adapted test, and the narrow form of back translation never reviews the target version itself. Use it as a spot check alongside expert review, a pilot and item analysis.

What is the difference between translating and adapting a test?

Translation moves the wording into another language while keeping its meaning. Adaptation is the broader term: it also changes examples, units, laws or formats so the item measures the same thing in the new setting. A correct translation of a reference to a US regulation, for example, may still be irrelevant to learners abroad.

How many candidates do you need to check a translated exam for bias?

The ITC suggests a modest pilot sample, for example 100, for item analysis, and at least 200 people per version for studies that identify potentially biased items. Analyses based on item response theory need at least 500. With lower volume, rely on expert review and item statistics monitored by language, and avoid strong conclusions.

Can you use machine translation for exam items?

It can produce a first draft, but it does not replace the process. The ITC Guidelines call for qualified experts who understand the language, the culture and the content, and for evidence that items mean the same thing across populations. Budget for bilingual subject experts to review every item and for a pilot before scores count.

Should learners choose their own exam language?

Usually yes for the interface. For items, yes only when you offer validated versions and the job does not require a specific language. If learners must read manuals in English at work, an English exam can be the valid choice. Record each learner’s language and report results by version.

Do translated exams need separate item analysis?

Yes. Run item analysis for each language version as well as overall. Combined statistics can hide an item that works in one language and fails in another, such as a distractor that attracts many learners only in the translated version. Small groups give noisy numbers, so treat flags as prompts for review.