eduKate Singapore
Auto mode follows the Route Engine until you choose a viewpoint.
YOU ARE HERE
CONNECTED TO
WHAT NEXT
Use the canonical route for this room, or HELP if you are unsure.
USEFUL ROOMS IN THIS HUB
—
by
On this page
QTI 3 assessment item localization QA asks whether a translated digital question still measures the same skill and produces the same intended score. Translating QTI items is not simply translating a question stem. In a portable Question and Test Interoperability package, the visible prompt, response choices, identifiers, correct-response declarations, scoring rules, feedback and media can occupy different parts of a structured XML artifact. A plausible translation can be unusable when those parts no longer agree.
People searching for QTI item translation, 1EdTech QTI 3 response processing, multilingual quiz item validation, translated assessment scoring or correct response QA usually need an operational handoff between linguists, item writers and assessment-platform engineers. The difficulty is that language carries the learning target while machine-readable identifiers carry the system’s answer logic. Changing either without checking the other can make a question misleading or the marking wrong.
This guide owns one reader job: accept or reject one localized QTI 3 assessment item without changing what it assesses, how it accepts a response or how it calculates and reports a score. It is not a general guide to translating classroom worksheets, not an assessment psychometrics textbook and not a replacement for a receiving platform’s import documentation. The method is useful to tutors, publishers and learning-technology teams that exchange questions between compatible systems.
A reliable QTI 3 localization review starts with a source item and a documented scoring expectation. Freeze a copy of the source XML and any media or package resources. Identify the item body a learner sees, the response declarations and identifiers used by the system, the correct or mapped outcomes, the response-processing rules and the feedback that appears after submission. Translate the user-facing language with context while protecting the machine-readable relationships unless an assessment engineer explicitly changes the item design.
Then render the source and target in an actual player or supported preview. Submit correct, incorrect and boundary responses. Record whether the item behaves as intended in both versions. Inspect feedback as well as the numeric score. Finally export and import the target through the receiving platform, because a locally valid XML file is not automatically supported by every assessment engine. Keep a result sheet showing one item ID, one source version, one target locale, test answers and observed outcomes.
The primary acceptance rule is stronger than “the translation sounds natural.” It asks whether a learner who understands the same concept has a fair opportunity to select the corresponding answer and receive the intended score. Two versions can have different words while being equivalent in the skill they measure. Two versions can also have nearly identical words while one accidentally reveals the answer, narrows the set of choices or changes a negative into a positive.
The first layer is visible instructional content: the prompt, reading passage, answer choices, hints, diagrams, audio transcripts and feedback. Linguists usually work here. The second layer is response structure: the kind of interaction offered to a learner, such as choosing among options, entering a text response or selecting a point. The third layer is scoring behaviour: correct responses, mappings, processing templates or custom rules. The fourth layer is packaging and delivery: how the item and its resources travel into another system.
The 1EdTech QTI 3 beginner’s guide explains that a choice interaction uses a response-identifier to connect the interaction to a response declaration. Simple choices have identifiers, and the selected identifier is used in response processing. This means the human-facing option wording and the machine-facing option identifier are deliberately separate. Translating a sentence does not require renaming its answer ID; renaming an ID may instead break the connection to the correct-response declaration.
response-identifier
A beginner may assume the answer key is stored next to the visible correct sentence. In a structured item the answer key can live in a response declaration or processing rule. For example, the target choice labelled “Evaporation” might carry identifier choice_B while the correct response declaration also refers to choice_B . If a designer replaces the visible text but changes the identifier to choice_C , the user-facing wording may look fine while the scoring engine checks a different value. The correct task is to verify the correspondence, not preserve arbitrary letters as if they were answer content.
choice_B
choice_C
These four layers also explain why ownership boundaries matter. A translator can recommend that a distractor be rewritten because the literal translation becomes nonsensical. An item writer decides whether the revised distractor still serves its educational purpose. An engineer confirms whether the XML, response processing and rendering support the edited item. A final reviewer checks the complete experience. No single test substitutes for all those judgments.
The construct is the knowledge or skill the item is intended to measure. A mathematics item might assess ordering fractions. A science question might assess identifying the role of evaporation in a cycle. A language item might measure recognition of implied meaning. Write that objective down before translating because it helps distinguish language needed for the construct from language that is merely a vehicle for the question.
Suppose the source asks, “Which container cooled fastest?” and a chart displays three temperature curves. The intended construct is interpreting cooling rates from evidence. If a target version says “Which container ended at the lowest temperature?” the visible words still concern cooling, but the question now assesses final temperature rather than rate. The chart might support different answers for those two questions. A translator cannot establish equivalence by replacing technical nouns alone; the requested operation must remain the same.
At the same time, avoid making language unnecessarily complex when the item is not assessing language ability. An awkward translation may turn a science assessment into a reading-comprehension obstacle. Review vocabulary at the intended age and proficiency, but do not remove essential scientific conditions. In “Which statements must be true?” the phrase “must be” establishes necessity. Replacing it with “could be” changes logic. Simplification is acceptable only when it leaves the target inference unchanged.
For high-stakes assessment, translated forms may require formal adaptation, equating, pilot testing and qualified psychometric analysis beyond item-level localization QA. This guide addresses the engineering and content acceptance of a portable item; it does not claim that matching answer keys proves two language forms have identical statistical difficulty or fairness. Keep that distinction explicit when results affect progression or certification.
Build an item ledger with separate columns for source ID, visible text, semantic role, target text, invariant identifier, expected response and test status. Semantic roles include stem, instruction, choice, hint, feedback, alt text, rubric and metadata. A mixed list of arbitrary XML strings is not enough because the same sentence can function differently depending on where it appears. A hint may be allowed to guide thinking; a distractor should not accidentally provide a tutorial for the correct answer.
Lock response identifiers, correct-response references, scoring templates, data types and outcome identifiers unless a controlled engineering change is approved. Also lock image references, mathematical values, constraints and randomisation rules. Be wary of treating numerical examples as localizable style. A decimal separator can be formatted for a locale, but the underlying value must remain the same. Replacing a problem’s numbers merely to make them familiar is item adaptation, not a silent translation edit.
Record the source package, QTI version, exporting system, importing system and supported interaction types. QTI is an interoperability specification, but platform implementations may differ in supported features. An item can be syntactically plausible and still rely on a feature the destination does not render or score. A small compatibility test with the receiving engine should precede a large translation batch, especially if the package includes custom response processing, rich interactions or external media.
Make a separate note of the answer rationale. A simple item that appears to ask one obvious thing may rely on a particular distinction between two near-synonyms. A translator who only sees the stem and choices can erase that distinction. Giving the linguist the learning objective, correct answer and a concise explanation helps protect the purpose of plausible distractors without turning the target sentence into an explicit clue.
A stem establishes the operation the learner must perform: identify, compute, compare, infer, justify, select all, select one or arrange in order. Negation and quantifiers are especially important. “Which option is not supported?” is the inverse of “Which option is supported?”; “all” is not “some”; “least” is not “most.” During review, highlight these logical operators and compare them explicitly against the source.
Translate any answer-count instruction alongside the interaction configuration. If the source permits one choice and the target says “Select two answers,” the assessment has become inconsistent even though the words themselves could be grammatical. If the source uses a multiple-response interaction, a target instruction saying “Choose the best answer” may suggest a single response. Check the displayed instruction against the interaction’s actual cardinality and permitted response behaviour in the receiving player.
Many item stems contain deictic references such as “the figure below,” “the highlighted cell” or “the second paragraph.” Localized layouts can move media above the text or place content on another screen. Verify that those references still point to the right element in the target rendering. If the source layout changes during adaptation, revise the instructional reference with editorial approval rather than preserving a phrase that is now literally false.
Use a deliberate ambiguity check. Give a bilingual subject reviewer the target item without the source and ask what the question requires. Ask for a plain-language paraphrase of the operation and the evidence needed to answer. Then compare with the source construct statement. This independent interpretation catches mistakes that a side-by-side proofreading pass may miss because the reviewer subconsciously supplies the source meaning.
A choice interaction may contain three incorrect but plausible distractors and one correct response. Localization can damage a distractor in several ways. A wrong option may become an obvious grammatical outlier. Two options may collapse into the same meaning. A culturally specific example may become unfamiliar, but removing it might change the intended difficulty. One option might become much longer and attract attention simply because it is more detailed. These are assessment-design issues as well as language issues.
Keep the stable machine identifier with the same logical option while translating the display text. Do not assume choice_A must appear first in the interface; a shuffled interaction may display choices in a different order. The scoring logic identifies choices by their declared identifiers, not by visual position alone. To test an option, use the target displayed wording to determine its identifier and verify that the outcome for selecting it matches the declared scoring rule.
choice_A
A deceptively small failure occurs when two source distractors use different degrees of certainty. For example, one states that a claim “always” holds and another says it “sometimes” holds. If both become a target phrase equivalent to “usually,” the learner loses the intended distinction. This may make the answer set non-exclusive or alter the number of defensible answers. A subject reviewer should inspect contrasts across the whole set, not approve each translated option independently.
If a choice contains an equation, image or code snippet, the translator must distinguish which part is human language and which part carries the answer logic. A correct variable name or symbol may need to remain unchanged. Reformatting an expression in a visually attractive way can silently move parentheses and change the evaluated value. For technical problems, review source and target expressions symbol by symbol rather than relying on similarity of rendered appearance.
A source science item asks, “Which change increases the rate of evaporation?” The choices are lowering temperature, reducing exposed surface area, increasing moving air over the liquid and placing a lid on the container. The third choice is intended as correct in the stated classroom model. The item checks recognition of factors affecting evaporation rate. A useful translation preserves the contrast between each option’s causal effect rather than merely repeating the word “evaporation” correctly.
Imagine the target translation changes “moving air” into a phrase meaning “air pressure rises.” The sentence is fluent but introduces a different physical mechanism. A reviewer might still spot that the third choice is identified as correct and assume the answer key survived. In reality, the learner now sees an option not supported by the original teaching objective. The failure is semantic, even though the QTI correct-response identifier remains intact. Locking the XML alone would not catch it.
Now suppose the target phrasing accurately describes airflow but two distractors both become “cover the liquid.” That makes the options less discriminating. The repair should preserve the difference between reducing area and using a lid. The subject reviewer might approve a more explicit target wording for each, provided the resulting item still assesses the same principle and does not reveal the answer through unnatural precision. This is why the full set of choices must be reviewed together.
Finally submit each option in the target player. Confirm the intended correct choice receives the expected score and all distractors follow the same source policy. Test the player with shuffling enabled if the item supports it. A mismatch where the correct target wording receives zero points shows an identifier or processing problem; a target distractor that actually becomes scientifically true is a content problem. Both block release, but their fixes belong to different owners.
A history item asks, “All of the following were consequences of the policy, except one. Which one?” The correct choice is the exception. A literal translation may move the negative construction to the end of a long sentence, making it easy for younger readers to miss. A more natural target formulation might ask, “Which outcome did not result from the policy?” That change can be valid if the logical exclusion and choices remain the same.
Check the item at three levels. The translated stem must still request the exception. The choices must represent the same claims. The correct-response identifier must still point to the item that is not a consequence. A reviewer who checks only the stem can miss a separately edited option. A reviewer who checks only the XML will see no defect if the target language has reversed the polarity of the question.
A useful testing method is to produce a truth table. For each option, write whether it meets the target predicate: consequence or not consequence. Compare the truth table with the source. If the marked correct response is the only negative outcome in one version but a positive outcome in the other, the translation has changed the assessment. This little table is faster than debating whether the sentence “feels equivalent” when the issue is logical scope.
A mathematics item requires a numeric answer of 1.5 . In some locales, the display convention uses a decimal comma. The linguist may adapt visible prose to “one point five” or the culturally expected equivalent, but the QTI response base type, numeric comparison and scoring tolerance are not ordinary text. A target learner entering 1,5 may be accepted or rejected depending on the player and locale configuration. The acceptance test must therefore exercise the actual input control.
1.5
1,5
First record the mathematical invariant: the expected real number is 1.5, not fifteen. Then review the displayed instruction, any example of accepted input syntax and the player’s parser. Test reasonable locale variants, including the intended decimal separator. If the system accepts only a period, the target guidance may need to make that explicit or the platform may need localization work. Merely translating the instruction without checking input behaviour creates an unfair mismatch between language and interaction.
The same logic applies to thousands separators, minus signs, percentage formatting and units. A prompt saying “Enter the answer in centimetres” cannot be translated into a target asking for metres without adjusting the intended numeric response. If conversion is officially part of the adaptation, document it as a redesigned item and revalidate scoring. Silent unit changes are a scoring defect waiting to happen.
In an ordering task, the learner arranges four stages of a process. The item may represent each stage with a stable identifier and use response processing to evaluate their order. Target-language phrases can be much longer, so dragging tiles may become awkward or overlap. The correct sequence is a content invariant; the visible tile width is a layout variable. Review both. A translated tile should still be recognisable while being draggable through the intended interface.
For a matching item, each premise and response may be a separate labelled element. Translation can create two near-identical labels that make an otherwise straightforward task ambiguous. Test every pair and ask the subject reviewer whether each match remains uniquely defensible. If two responses are synonymous in the target, the item may need controlled adaptation, not a last-minute answer-key change by a translator.
For graphics-based drag-and-drop tasks, referential integrity matters. A response zone associated with the heart in a biology picture must remain attached to that organ after target-language labels have been moved. A correct string in an incorrect drop zone is not equivalent. These interactions are especially dependent on platform rendering, pointer or keyboard support, and accessibility. A read-only XML check should be supplemented with hands-on testing in the actual player.
The QTI 3 beginner’s guide describes the connection between a response declaration and an interaction’s response-identifier . For a conventional choice example, a declaration might identify RESPONSE , while the interaction refers to RESPONSE and the declared correct value identifies the right simple choice. The important thing is consistency of references. A human translator generally should not modify those attributes while localizing visible words.
RESPONSE
The same official guide describes response-processing templates for common scoring behaviours, including match-correct and map-response patterns. With match-correct, the engine evaluates a response against a declared correct response. With mapped responses, the item can award partial credit according to its mapping. A localized item can appear to work on one correct response yet fail partial-credit cases, so the test plan must reflect whichever policy the source item uses.
If a localization pipeline regenerates XML, compare identifiers, cardinalities, base types, mapping entries and referenced resources before and after transformation. Not every byte must remain identical, because some metadata and translated content legitimately change. Instead compare a protected subset representing operational invariants. Any unexpected change becomes an engineering review item, not an assumption that different target language requires different internal IDs.
Avoid claiming that XML schema validation alone proves scoring. A file may be well-formed and conformant while containing the wrong correct-response identifier or a logically mistaken mapping. Conversely an engine can import a structurally valid feature yet display it differently from the author’s intention. Treat structural validation, programmatic diff and behavioural test as complementary evidence.
Feedback appears after a response or at another stage defined by the assessment. It can explain why a choice is correct, why a distractor fails or what concept to revise. Localization must preserve which feedback appears for which response or outcome. A correctly scored option paired with feedback for another distractor confuses learning and may undermine a tutor’s explanation of the result.
A feedback translation needs to preserve the stated reason, not merely the praise or discouragement. “Correct, because the surface area increased” is not equivalent to “Correct, because the temperature increased” even if both sound encouraging. If the item uses the same feedback block for several responses, check that the target wording remains compatible with all those paths. A description that names one specific error may be inappropriate for a generic incorrect-response branch.
Check whether feedback includes links, diagrams or audio. A link labelled “Review the graph” should lead to the intended resource in the target learning environment, not an unrelated source-language page. If the target course has no equivalent resource, the team should document that limitation or create an approved fallback. A translated label cannot magically make the linked evidence localized.
For a practice item, detailed explanation may be useful. For a secure test, an item could intentionally withhold feedback until a permitted review period. Translation must not change that disclosure policy. A well-meaning editor who adds a revealing explanation in a target-language hint may compromise assessment integrity. Respect the timing and visibility rules defined by the original item and deployment.
A QTI item can include images, charts, formulas and audio. These may carry essential language or culturally specific information. Inventory all resource references from the item and package. An image may contain an English legend that a plain text extraction misses. A listening assessment may require separately recorded target-language audio, while an item assessing recognition of English speech should not automatically have its source recording replaced.
Ask whether the media is part of the construct. In an English-language listening test, changing the audio language fundamentally changes what is being tested. In a science item where audio narrates instructions, an approved target narration may be necessary for accessibility and language parity. The correct decision follows the learning objective, not a blanket rule that every recording must be dubbed.
Mathematical expressions require special care. Translating surrounding instructions is usually separate from changing MathML, LaTeX content, variable names and operator precedence. If the target typographic convention uses a different decimal or multiplication symbol, review both mathematical equivalence and the renderer’s implementation. Validate superscripts, fractions, roots and signs in the actual item display, including mobile presentation if students take the assessment on tablets.
Alternative text for images should preserve what the learner needs without giving away a response that sighted learners have to infer. The challenge is especially acute when an image is itself the evidence from which a student draws a conclusion. A descriptive alternative might state chart values or relevant relationships, but the accessibility adaptation should be designed with assessment specialists so it measures the intended skill fairly. Generic machine-generated image captions are not sufficient for this decision.
Word count is a poor measure of assessment equivalence. An English stem may use six words while a natural target sentence needs twelve. Forcing the target down to six can introduce ambiguity. Conversely a very verbose translation may overload working memory in a task meant to assess a simple concept. Review the cognitive job: what information must the student hold, what inference must be made and what evidence distinguishes a correct answer from an incorrect one?
Watch culturally specific references. A question about coins, school years, sports or holidays may be more familiar to one audience than another. Sometimes localization can preserve the original reference with a short explanatory note. Sometimes the item requires a controlled adaptation and a fresh fairness review. The translator should flag the issue rather than independently replacing a reference with something that alters numerical relationships or clues.
Do not equate a single successful pilot response with measurement equivalence. Educators may need cognitive interviews, item analysis or statistical differential item functioning methods for important comparative examinations. The technical acceptance gate in this article verifies a coherent item and its scoring behaviour; fairness evaluation is a wider research task. Maintaining that boundary protects both the educator and the learner from overclaiming what QA can prove.
Start with the simplest matrix: choose the intended correct response, choose each distractor and leave the item unanswered if the player permits it. Record the returned score, feedback and response state for each. For multiple-response items, include partial selections, extra choices and an empty selection. For mapped scoring, test at least one partial-credit path and one zero-credit path. For numeric responses, test accepted and rejected boundary values according to the item’s tolerance policy.
The matrix should use the intended outcomes rather than a generic demand that every item score zero or one. A source item may permit partial credit or multiple attempts. The target must preserve those rules unless a documented assessment redesign changes them. If a test engine reports a different value from the author’s expected matrix, pause the translation release and investigate the declarative scoring data or import behaviour.
Behavioural tests should be run after the target text is loaded, not before. A localization edit can affect a control, especially if an answer choice becomes too long to click, a drag tile disappears off-screen or a button label is hidden. When a tester checks only the backend response function with hard-coded IDs, they may prove that scoring still works while overlooking that real students cannot reach the correct option through the interface.
Where automation is available, combine machine and human steps. A script can compare the invariant XML fields and run candidate responses against a sandbox player. A bilingual tutor can verify whether the target wording still elicits the same reasoning. A specialist can inspect accessibility and fairness concerns. Automation increases repeatability; expert judgment determines what the response means.
A QTI package can include several files linked through a manifest or resource references. Translation work can accidentally alter filenames, paths, character encoding or media references. Before handoff, ensure every referenced resource exists and every required file is included. A valid standalone item file that points to a missing diagram will fail the learner’s experience once moved to another platform.
Test the version the destination system actually imports. Not every LMS has the same QTI 3 support, and some systems transform items into their own format. A preview in the exporting authoring tool may display a choice interaction correctly while the receiving engine ignores a feedback rule or changes a shuffle behaviour. A small round trip—export one representative target item, import it, answer it and re-export if possible—can uncover those gaps before a batch of hundreds is delivered.
If an engine does not support a required interaction, avoid silent substitution. Converting a complex hotspot item into a plain multiple-choice question changes how learners respond and may change the construct. The content owner and assessment engineer should approve any conversion as a new item version. Record the original item ID, derivative version and scoring differences so the institution can track results responsibly.
Release records should also identify supported player, browser and input methods. A question that works with a mouse may still be inaccessible to a keyboard user. Translation is not an excuse to ignore existing accessibility requirements. If the localised text makes focus order confusing, the issue may be layout and interaction design rather than the linguistic content itself.
If the correct target choice scores zero, examine whether its stable identifier matches the declared correct response and whether the imported target item uses the same processing policy. If a scientifically true option is marked wrong, examine the language and answer rationale before touching code. If both the language and XML are sound but the click does not register, examine the renderer and accessible control state. Route the defect according to evidence.
If the target shows two equivalent choices, ask a subject reviewer to compare semantic alternatives and revise distractors within the permitted adaptation policy. If text overlaps a button, ask the designer or renderer owner to adjust the layout before forcing unnatural shortening. If feedback appears at the wrong time, inspect assessment presentation settings and feedback relationships. If a linked image is missing, inspect the content package and resource path rather than translating the filename.
A particularly dangerous failure is quietly editing the source answer key to make a mistaken translation seem valid. That may resolve a visible mismatch but destroys traceability and can invalidate results. The correct repair is to restore the target item to the intended proposition or formally revise the item design with a new version and approved scoring rationale. Do not use technical convenience to redefine what the question measures.
Keep an exception register for deviations that cannot be fixed in the current release. A minor typographic limitation may be documented and accepted. A scoring discrepancy, unanswerable prompt or critical accessibility barrier should block release. The acceptance threshold follows the consequence for learners, not the number of remaining tickets.
Scale begins with role clarity. Item authors supply constructs, answer rationales and invariants. Translators receive context for stems, choices and feedback. A QTI-aware validator checks protected identifiers and package references. Subject reviewers examine target meaning and distractor quality. Testers submit representative responses in the actual delivery engine. Release managers tie approved targets to source versions. This division prevents every new locale from reinventing the method.
Prioritise representative item types before mass translation. A simple single-choice question is not proof that numeric entry, drag matching, media or partial-credit processing will work. Build a compatibility set that samples each interaction family used by the institution. For each target locale, run the invariant and behaviour tests against that set first. Only then scale the workflow to the full bank, reserving deeper manual review for complex or high-stakes content.
A change to the source item must trigger a target impact check. If an item writer revises the correct response or adds an option, do not assume previously approved translations remain valid. Identify which targets were based on the old source version, compare the new construct and scoring relationships, and revalidate affected items. This is not merely translation memory maintenance; it protects the integrity of assessment evidence over time.
Teachers can use a simplified version even without writing QTI XML. For any digital multiple-choice activity, write the measured skill, identify the correct answer, check the meaning of every distractor, submit each response and inspect the displayed feedback. The QTI framework simply makes the underlying contracts more explicit. The transferable lesson is that language, interface and marking must tell the same educational story.
Fast QTI 3 localization does not mean racing through stems while trusting the rest of the package. It means isolating language from machine references, translating with the answer rationale in view, checking distractor relationships, preserving the intended scoring contract and testing the real player. A linguistically polished question with a broken key is not ready. A technically perfect key attached to misleading wording is not ready either.
The final acceptance question is: does the target item ask the same substantive question, offer defensible corresponding responses, score those responses as intended and present the result in the receiving environment? When the evidence says yes, the localized item can join the bank with a clear version and test record. When any part says no, the workflow should stop at the precise layer that failed and repair that layer without contaminating the others.
Generally, user-facing wording can be translated while machine-facing choice identifiers should remain stable. Any change to identifiers must preserve all references in response declarations and processing rules and be controlled by the assessment engineering workflow.
No. Structural validation does not prove the question’s meaning, answer rationale, scoring logic or receiving-player behaviour. Test correct, incorrect and relevant boundary responses in the actual target environment.
Stop release. Ask the item author and bilingual subject reviewer to revise the distractor or formally adapt the item. Do not silently change the answer key just to accommodate a bad translation.
No. Equivalent constructs matter more than equal length. But target language must not introduce unnecessary complexity, clues or unfamiliar context that changes the student’s task.
No. Technical and linguistic QA can establish coherent content and scoring behaviour. High-stakes cross-language form equivalence may also require formal validation and statistical studies.
Connect this article to the How People Translate Quickly series index as the specialist destination for portable assessment item acceptance. The previously prepared H5P interactive-activity QA draft addresses a different platform and should not be copied into this URL. For general educational explanation, direct readers to existing subject-specific eduKate learning pages when the actual final internal URL has been verified.
