How do professional teams translate QTI 3 assessment items and tests without changing the correct answers, scoring rules, reading demands or opportunities available to candidates? Question and Test Interoperability (QTI) is a 1EdTech standard that lets authoring systems, item banks and delivery platforms exchange structured assessment content. A translated question is still an assessment instrument; fluent language is necessary, but it is not sufficient evidence of equivalent measurement.
This guide explains QTI 3 localization, multilingual assessment item authoring, QTI 3.0.1 XML testing, response-processing preservation, digital test accessibility, language variants, portable custom interactions and result-reporting review. It is designed for international examination teams, educational publishers, psychometricians, item writers, software engineers, accessibility specialists and translators who work together on high-stakes or classroom assessments.
1EdTech released QTI 3 as a final standard and publishes versioned implementation material; its QTI 3.0.1 information-model and binding documents add technical detail. QTI 3 improves web-compatible markup and accessibility and incorporates capabilities formerly associated with APIP. This article separates source item meaning from translation choices and implementation behaviour. No example here authorizes a team to alter approved answer keys, scoring algorithms, candidate accommodations or legal examination rules.
The central rule: the candidate must face the intended question
A legitimate translation may use different words, but it must not silently demand a different inference, grant an extra hint, remove a useful accommodation or reinterpret a result. Treat wording as one part of the test design and verify both candidate experience and the machine’s response interpretation.
QTI is an assessment contract, not a word-processing format
A QTI item can contain a prompt, response declarations, outcome declarations, response-processing rules, feedback, metadata and media. A test can establish sections, navigation, selection and delivery conditions. A translator sees readable prose, but those words are embedded in a decision system that records candidate responses and determines outcomes. Editing only the visible words without understanding the response model can make a test misleading or mathematically wrong.
The three approvals that should never be merged
Language review asks whether the candidate understands the intended question in the target language. Subject review asks whether the same scientific, mathematical or literary construct is being tested. Measurement review asks whether difficulty, distractors, response patterns, accommodations and score interpretation remain defensible. A passage can be grammatically excellent and still fail the second or third test. Maintain separate sign-offs and involve the original test owner when an adaptation rather than an exact translation is necessary.
Item content, interaction logic and evidence
Item bodies may be localized, but identifiers, data bindings, response declarations, correct-response values and processing expressions require technical control. In multiple-choice items, a translated label does not license changing which choice maps to a given response identifier. In numerical items, punctuation conventions affect perception, while the numeric value and tolerance define scoring. In interactive items, a drag target or hotspot may be paired with hidden geometry and event logic that must continue to work.
Accessibility is a candidate experience
QTI 3 provides mechanisms relevant to alternate or supplemental content, language variants and accessibility needs, but the presence of technical markup does not prove a particular test is accessible. Alternative text must convey the information a visually impaired candidate needs without giving away the answer. Recorded audio, on-screen reading order, mathematical speech, bidirectional text and personal-needs settings must be tested in the actual delivery engine. Accommodations may have legal significance; do not treat them as optional styling.
Localization versus adaptation
Replacing a familiar sport, currency or measurement with a different local example may help comprehension, but it can also change background knowledge, difficulty and score interpretation. A translation tries to preserve the intended construct, often with authorized linguistic adjustments. A content adaptation changes stimulus or context and needs a documented review path. Keep version records that distinguish these two activities instead of labeling every changed item a simple translation.
One source, multiple delivery systems
An item may look correct inside an authoring tool yet fail after export, import or rendering in another product. Valid XML cannot guarantee semantic equivalence, and a good preview cannot guarantee preserved scoring. A reliable review sequence checks package structure, content rendering, response processing, accessible alternatives, test assembly and results. Use fictional candidate responses and deliberately wrong answers as well as correct ones; otherwise scoring defects can remain invisible.
Forty-three QTI localization decisions with worked risks
Item identifiers
A bilingual item bank contains hundreds of translated versions of the same assessment task and their revision history. Renaming response or item identifiers as natural-language words can break links, result analytics and re-imports. The point is not to force a word-for-word rendering. It is to preserve the evidence the assessment was designed to collect. A reviewer must know whether the difference is linguistic, subject-specific or part of the scoring contract. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Preserve stable identifiers and compare source-target mappings at the package boundary. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Response declarations
A single-choice question declares response identifiers which may not resemble the visible choice labels. Changing an identifier to match translated wording may produce an unscorable response. This is a good example of why candidate-facing fluency and technical validity must be assessed separately. One can be excellent while the other fails, so a green check beside the XML parser is not sufficient. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Check the declaration and the selected response value with a known test attempt. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Correct responses
An item uses a defined answer value and displays a longer natural-language explanation to candidates. A translator can accidentally move the correct answer to a different option while preserving polished wording. The consequences depend on the stakes. In practice exercises the error may confuse one learner; in certification the same ambiguity can change a consequential decision. Make the approval process proportionate to the potential harm. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Use an answer-key audit independent of the language review. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Distractor design
A science question contains three incorrect answers designed to reflect different misconceptions. Making a distractor less plausible in another language changes difficulty and reduces diagnostic value. Do not compensate for an unclear translation by silently changing an answer key or hidden processing rule. Escalate the mismatch to the item author and agree on a new reviewed version if adaptation is genuinely necessary. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Ask a subject specialist to classify the misconception behind every translated choice. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Negation
The source prompt asks which statement is not correct, with emphasis on the negative instruction. If the negative is weakened, many candidates may answer the opposite question. Candidate expectations are formed by what is displayed and spoken, not by metadata alone. Have a person unfamiliar with the source language explain the target-language instruction before deciding it communicates the intended task. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Verify the negative using a plain-language paraphrase and candidate comprehension test. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Quantifiers
The prompt distinguishes always, sometimes, never, at least and at most. A single swapped quantifier can change the set of valid answers. Good evidence includes both ordinary responses and boundary cases. A familiar sample answer can pass even when another valid response, assistive setting or imported interaction has stopped working. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Back-test with counterexamples that expose whether the translated condition has shifted. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Numbers and decimal punctuation
An engineering item displays 1,250 and 1.250 under distinct locale conventions. Ambiguous grouping can turn the intended quantity into a different number. The point is not to force a word-for-word rendering. It is to preserve the evidence the assessment was designed to collect. A reviewer must know whether the difference is linguistic, subject-specific or part of the scoring contract. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Review all numeric literals with the approved target-locale format and scoring value. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Units and conversions
A physics item uses metres per second while a target audience commonly sees kilometres per hour. Converting units without adjusting values and tolerances changes the arithmetic task. This is a good example of why candidate-facing fluency and technical validity must be assessed separately. One can be excellent while the other fails, so a green check beside the XML parser is not sufficient. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Keep original units unless the examination owner authorizes a fully checked adaptation. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Rounding and precision
A chemistry answer expects a specified number of significant figures. Translating the instruction as ordinary rounding could alter acceptable response ranges. The consequences depend on the stakes. In practice exercises the error may confuse one learner; in certification the same ambiguity can change a consequential decision. Make the approval process proportionate to the potential harm. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Test the exact precision wording against the response-processing rules. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Mathematical symbols
A question contains inequalities, exponents, radicals and variables named with Latin letters. Replacing symbols for typographic preference can alter meaning or break accessible speech. Do not compensate for an unclear translation by silently changing an answer key or hidden processing rule. Escalate the mismatch to the item author and agree on a new reviewed version if adaptation is genuinely necessary. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Compare rendered and accessible math with the reference expressions. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
MathML and mathematical speech
A screen reader should read an expression in a way that communicates its grouping and operators. A visual preview may look correct while spoken reading changes order or scope. Candidate expectations are formed by what is displayed and spoken, not by metadata alone. Have a person unfamiliar with the source language explain the target-language instruction before deciding it communicates the intended task. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Test actual screen-reader output in the target language and inspect the math source. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Rubrics and partial credit
An extended-response item contains analytic criteria and multiple score levels. A translation that adds a criterion or broadens an achievement label changes scoring fairness. Good evidence includes both ordinary responses and boundary cases. A familiar sample answer can pass even when another valid response, assistive setting or imported interaction has stopped working. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Review rubric descriptors against scored examples at each boundary. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Response processing
The item uses conditional processing that maps candidate values to outcomes. A language editor may unintentionally touch processing logic when rewriting markup. The point is not to force a word-for-word rendering. It is to preserve the evidence the assessment was designed to collect. A reviewer must know whether the difference is linguistic, subject-specific or part of the scoring contract. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Run source and localized fixtures through identical responses and compare outcomes. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Feedback timing
The source item reveals feedback only after submission or within a defined practice mode. Moving explanatory text into the visible prompt can leak answers. This is a good example of why candidate-facing fluency and technical validity must be assessed separately. One can be excellent while the other fails, so a green check beside the XML parser is not sufficient. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Test the sequence of views and ensure feedback appears under the intended conditions. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Item-level feedback
The platform shows one message for correct responses and another for a particular misconception. A target-language message can name the answer too directly before resubmission. The consequences depend on the stakes. In practice exercises the error may confuse one learner; in certification the same ambiguity can change a consequential decision. Make the approval process proportionate to the potential harm. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Check each response branch under practice and examination settings. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Test-level navigation
An assessment allows or forbids returning to earlier questions. A translated ‘Next’ button must not imply that a candidate can go backward if the engine prohibits it. Do not compensate for an unclear translation by silently changing an answer key or hidden processing rule. Escalate the mismatch to the item author and agree on a new reviewed version if adaptation is genuinely necessary. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Confirm navigation language against actual permitted actions. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Timing instructions
Some timed sections have distinct start events, warnings and stop conditions. An ambiguous phrase can make candidates believe there is more or less time available. Candidate expectations are formed by what is displayed and spoken, not by metadata alone. Have a person unfamiliar with the source language explain the target-language instruction before deciding it communicates the intended task. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Test localized countdowns and instructions with the true section timing. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Adaptive selection
Computer-adaptive delivery chooses subsequent items based on previous responses. Reordering translated items manually can compromise an adaptive test design. Good evidence includes both ordinary responses and boundary cases. A familiar sample answer can pass even when another valid response, assistive setting or imported interaction has stopped working. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Do not change adaptive rules and inspect score comparability when language variants differ. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Item selection rules
A test blueprint chooses questions from several pools according to constraints. A translated bank may accidentally omit items in one pool and distort topic coverage. The point is not to force a word-for-word rendering. It is to preserve the evidence the assessment was designed to collect. A reviewer must know whether the difference is linguistic, subject-specific or part of the scoring contract. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Validate identifiers, pool membership and selection behavior for each locale. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Randomized choices
A choice interaction may shuffle the visible choices while answer identifiers remain fixed. A fixed ‘answer C’ translation can fail when presentation order changes. This is a good example of why candidate-facing fluency and technical validity must be assessed separately. One can be excellent while the other fails, so a green check beside the XML parser is not sufficient. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Test multiple random seeds and document answers by stable IDs rather than positions. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Drag-and-drop targets
A language-learning item asks candidates to drag labels onto regions with defined target positions. Long target-language labels can overlap or hide essential information. The consequences depend on the stakes. In practice exercises the error may confuse one learner; in certification the same ambiguity can change a consequential decision. Make the approval process proportionate to the potential harm. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Inspect geometry, touch targets, keyboard access and response mapping. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Hotspot interactions
A geography question expects the learner to select a region of a map. A translated instruction may describe a different geographic area without moving its hotspot. Do not compensate for an unclear translation by silently changing an answer key or hidden processing rule. Escalate the mismatch to the item author and agree on a new reviewed version if adaptation is genuinely necessary. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Compare region coordinates and wording, including accessible alternatives. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Portable custom interactions
A custom interaction uses scripts and a rendering contract beyond ordinary choice controls. Unreviewed translation of hidden state names can break interactions or accessibility APIs. Candidate expectations are formed by what is displayed and spoken, not by metadata alone. Have a person unfamiliar with the source language explain the target-language instruction before deciding it communicates the intended task. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Test the actual packaged interaction, not just a static screenshot. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Images carrying text
An infographic contains embedded language that is not visible in ordinary XML extraction. A text-only translation may leave the source language inside the image. Good evidence includes both ordinary responses and boundary cases. A familiar sample answer can pass even when another valid response, assistive setting or imported interaction has stopped working. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Prepare authorized localized media while preserving the evidence and question scope. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Alternative text
An image of a graph needs accessible description for candidates who cannot see it. Too much detail can reveal the intended answer; too little can deny access. The point is not to force a word-for-word rendering. It is to preserve the evidence the assessment was designed to collect. A reviewer must know whether the difference is linguistic, subject-specific or part of the scoring contract. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Have assessment and accessibility reviewers approve information parity. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Audio prompts
A listening item depends on speech rate, accent, pauses and aural distinctions. A beautiful translation may make an originally difficult listening task trivial. This is a good example of why candidate-facing fluency and technical validity must be assessed separately. One can be excellent while the other fails, so a green check beside the XML parser is not sufficient. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Review audio performance, timing and the assessed listening construct. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Captions and transcripts
An accessibility alternative may be provided for a media item under specific accommodations. Publishing a transcript to every candidate can change what a listening assessment measures. The consequences depend on the stakes. In practice exercises the error may confuse one learner; in certification the same ambiguity can change a consequential decision. Make the approval process proportionate to the potential harm. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Check accommodation settings and the authorized delivery mode. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Ruby and multilingual scripts
A Japanese-language assessment includes reading aids which may be essential or disallowed depending on level. Adding ruby markup without approval can supply hints or modify expected proficiency. Do not compensate for an unclear translation by silently changing an answer key or hidden processing rule. Escalate the mismatch to the item author and agree on a new reviewed version if adaptation is genuinely necessary. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Confirm when reading aids are part of the intended accessibility or instructional design. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Bidirectional text
An Arabic prompt includes Latin chemical abbreviations, numbers and parentheses. Visual directionality bugs may reorder tokens, labels or formula components. Candidate expectations are formed by what is displayed and spoken, not by metadata alone. Have a person unfamiliar with the source language explain the target-language instruction before deciding it communicates the intended task. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Inspect native bidirectional rendering and spoken order in the delivery platform. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Pronunciation and SSML
Text-to-speech alternatives can specify pronunciation for specialized names or symbols. A mistaken pronunciation may change a vocabulary, phonics or listening question. Good evidence includes both ordinary responses and boundary cases. A familiar sample answer can pass even when another valid response, assistive setting or imported interaction has stopped working. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Verify speech output with content experts and language reviewers. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Reading burden
Two translations convey the same proposition but differ greatly in sentence complexity. Unnecessary complexity can raise reading demand in an assessment not designed to measure it. The point is not to force a word-for-word rendering. It is to preserve the evidence the assessment was designed to collect. A reviewer must know whether the difference is linguistic, subject-specific or part of the scoring contract. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Review readability and response times with appropriately selected trial candidates. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Cultural references
A word problem mentions a local transport ticket unfamiliar to international learners. Substitution without approval could change both calculation and background knowledge demand. This is a good example of why candidate-facing fluency and technical validity must be assessed separately. One can be excellent while the other fails, so a green check beside the XML parser is not sufficient. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Record adaptation decisions and check construct equivalence with item writers. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Personal needs settings
A candidate’s preferences may request particular accommodations or alternate content. A translation must not silently override access settings because a target locale behaves differently. The consequences depend on the stakes. In practice exercises the error may confuse one learner; in certification the same ambiguity can change a consequential decision. Make the approval process proportionate to the potential harm. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Test the intended candidate profile across supported preferences. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Metadata language tags
An item bank tracks which language a file represents and which source it came from. Incorrect language tags can cause a delivery platform to select the wrong variant. Do not compensate for an unclear translation by silently changing an answer key or hidden processing rule. Escalate the mismatch to the item author and agree on a new reviewed version if adaptation is genuinely necessary. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Validate language metadata against actual item content and fallback behavior. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Package resources
An imported QTI package references images, stylesheets, scripts and other assets. Renaming or moving assets casually can break relative links after import. Candidate expectations are formed by what is displayed and spoken, not by metadata alone. Have a person unfamiliar with the source language explain the target-language instruction before deciding it communicates the intended task. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Validate the complete package and open every localized resource in the target platform. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
XML validity
A translated explanation contains ampersands, quotation marks and special characters. Improper escaping can make an otherwise correct assessment fail parsing. Good evidence includes both ordinary responses and boundary cases. A familiar sample answer can pass even when another valid response, assistive setting or imported interaction has stopped working. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Validate the XML against the applicable schema and test content rendering. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Namespaces and element names
An XML file contains technical element names that resemble English words. Translating elements rather than content changes the document grammar. The point is not to force a word-for-word rendering. It is to preserve the evidence the assessment was designed to collect. A reviewer must know whether the difference is linguistic, subject-specific or part of the scoring contract. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Keep technical names as required and isolate legitimate translatable nodes. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Score reporting
A results service or export carries candidate outcomes and associated metadata. An explanatory translation that turns provisional results into final certification misleads stakeholders. This is a good example of why candidate-facing fluency and technical validity must be assessed separately. One can be excellent while the other fails, so a green check beside the XML parser is not sufficient. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Match output labels to result states and ensure report data remains unchanged. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Error and incident messaging
A timed assessment disconnects and the candidate sees a recovery notice. A translated notice must not promise that answers were saved if that state is unknown. The consequences depend on the stakes. In practice exercises the error may confuse one learner; in certification the same ambiguity can change a consequential decision. Make the approval process proportionate to the potential harm. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Test interrupted sessions and use precise, condition-aware recovery language. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Psychometric equivalence
A question that works in one language may show different item difficulty in another. Linguistic equivalence alone does not establish statistical equivalence for high-stakes use. Do not compensate for an unclear translation by silently changing an answer key or hidden processing rule. Escalate the mismatch to the item author and agree on a new reviewed version if adaptation is genuinely necessary. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Use appropriately designed field testing and psychometric review when required. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Security and item exposure
A vendor sends real examination content through an unsecured translation workflow. An accurate translation can still compromise an entire assessment program. Candidate expectations are formed by what is displayed and spoken, not by metadata alone. Have a person unfamiliar with the source language explain the target-language instruction before deciding it communicates the intended task. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Restrict access, use controlled samples and verify handling of embargoed materials. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Versioned item banks
The source owner corrects an error while target translations are already in review. Different language editions may silently assess different source versions. Good evidence includes both ordinary responses and boundary cases. A familiar sample answer can pass even when another valid response, assistive setting or imported interaction has stopped working. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Keep revision links and revalidate answer keys and processing after each change. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Conformance claims
A provider advertises QTI 3 support but may not support every interaction or accommodation. Assuming full delivery parity from a generic label can create unequal candidate experiences. The point is not to force a word-for-word rendering. It is to preserve the evidence the assessment was designed to collect. A reviewer must know whether the difference is linguistic, subject-specific or part of the scoring contract. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Check product conformance scope and test representative item types. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Release acceptance
An assessment renders well in one browser but fails an accessibility mode or device. Visual acceptance alone misses evidence-affecting defects. This is a good example of why candidate-facing fluency and technical validity must be assessed separately. One can be excellent while the other fails, so a green check beside the XML parser is not sufficient. The target item should make clear what the candidate is asked to do, with no hidden change to the intended construct.
A responsible validation step is to Require a signed matrix of delivery, scoring, accessibility and language checks. Compare the source and localized versions using the same controlled candidate responses and, where appropriate, candidates who use assistive technology. Document the behavior rather than relying on a translated screenshot. If an apparently minor wording change affects the correct response, permitted assistance, time burden or scoring, hold the item for formal review before it can reach a learner.
Four-week implementation and learning plan
Week one — classify the assessment construct
Collect an authorized item sample across interaction types, content areas and candidate populations. Identify what each item is intended to measure, what evidence counts as a correct or partial response, which features must remain invariant and which language choices might affect difficulty. Prepare source and target traceability before any translation is approved.
Week two — translate with subject and accessibility review
Write candidate-friendly target text in the actual item context. Review options, distractors, math notation, audio, images, alternative text and navigation separately. Where the target audience requires an adaptation, refer the change to the assessment owner. Verify that XML structure, identifiers, score rules and required assets have not been rewritten.
Week three — exercise the real assessment engine
Import packaged items into each designated platform. Try correct, partially correct, incorrect, blank and malformed candidate responses. Check result equivalence, shuffle behaviour, time limits, feedback conditions and supported accommodations. Inspect reading order, mathematical speech, bidirectional scripts, keyboard navigation and display on small screens.
Week four — approve evidence, release and monitor
Require language, subject, technical and assessment-measurement sign-offs proportional to stakes. Record version, reviewer, issue and resolution for every changed item. Prepare a rollback to an approved package. After release, monitor item anomalies, accessibility complaints, unexpected response distributions and confusing wording without changing live scoring on the translator’s authority.
Frequently asked questions
Is translating QTI XML the same as translating a quiz document?
No. QTI content includes processing, interactions, resources and delivery behaviour. Only specifically designated human-readable content should be localized, and imported functionality must be tested. For high-stakes use, retain the assessment owner’s decision and a versioned test record.
Can we change answer order to improve readability?
Not casually. Some interactions randomize choices while preserving response identifiers; others depend on explicit mappings. Any changed ordering must be approved and tested without altering the answer key. For high-stakes use, retain the assessment owner’s decision and a versioned test record.
Does QTI 3 guarantee accessibility?
No. It supports important accessibility capabilities, but authors and delivery systems still need to create, render and test accessible content. For high-stakes use, retain the assessment owner’s decision and a versioned test record.
When is cultural adaptation necessary?
When literal wording introduces irrelevant background-knowledge demands or impossible local references. A subject and measurement reviewer should approve the change and document its effect on the task. For high-stakes use, retain the assessment owner’s decision and a versioned test record.
Can a translator rewrite the scoring rubric?
Not independently. A translated descriptor must preserve the decision boundary; changing criteria or score allocations is an assessment-design decision. For high-stakes use, retain the assessment owner’s decision and a versioned test record.
How should mathematical content be checked?
Compare expressions, quantities, units, tolerances, response values and accessible speech. Test at least a correct answer and nearby incorrect answers. For high-stakes use, retain the assessment owner’s decision and a versioned test record.
How do we know two language versions are equivalent?
Use linguistic, subject-matter, functional and, for consequential assessments, empirical evidence. A single back-translation or XML validator is insufficient. For high-stakes use, retain the assessment owner’s decision and a versioned test record.
What is the safest publication check?
Verify the exact exported package in the target delivery environment and reproduce expected scores, accessibility experiences and navigation with controlled test candidates. For high-stakes use, retain the assessment owner’s decision and a versioned test record.
Official references and adjacent topics
1EdTech QTI specification documents, 1EdTech QTI accessibility guidance and the QTI 3 best practices and implementation guide describe technical capabilities, versioned bindings and accessibility considerations. Confirm the exact adopted edition and product capability before applying a conformance label.
For related but non-identical topics, see CASE competency-framework localization and Common Cartridge course package localization. The short QTI localization QA guide is a narrower inspection workflow; this article covers the full item, test, accessibility and scoring design problem. The education translation architecture connects these layers.
Conclusion
A professional QTI translation is not finished when every sentence sounds fluent. It is finished when the translated candidate experience, approved question construct, response processing, accessible alternatives and result reporting have been shown to agree. Keep each technical contract intact, document necessary adaptations, and test the delivered exam rather than merely the source file.