Bilingual Evidence and Source Conflicts

A bilingual AI service can retrieve fluent Arabic and fluent English while giving different answers to the same business question. The problem may begin before a model is involved: two source versions can disagree about a limit, an exception, a date or the person authorised to decide. Translation quality and source authority therefore need to be assessed together before the organisation relies on bilingual answers.

The objective is not identical sentences. It is consistent business meaning under the same facts, with the correct source and an explicit route for unresolved differences. A polished translation can change a condition. A literal translation can preserve words while making the rule difficult to apply. Both situations require a controlled decision by people responsible for the source and its use.

This guide is for Saudi organisations preparing bilingual knowledge services, customer guidance or internal assistants. It provides an original conflict register, a paired-case method and a worked policy example. Use it as part of an enterprise data and AI readiness assessment, with actual legal or contractual interpretation assigned to the appropriate qualified owner rather than delegated to a language model.

Identify the source relationship before comparing wording

Two documents in different languages may be translations, separate policies, summaries for different audiences or versions from different periods. A shared title does not establish a translation pair. Confirm the document family, issuing authority, covered entity, audience, effective interval and approval relationship. Then determine whether one document translates the other or whether both have independent authority for different circumstances.

Record which language or document governs the particular question, if that has been established through the applicable source or authorised process. Do not assume Arabic always governs every business document in Saudi Arabia, or that English governs because a software product was configured in English. The answer depends on the relevant document and context, and may require professional interpretation beyond a technical readiness assessment.

Give the relationship a status: approved translation, translation awaiting review, independently issued version, historical counterpart or relationship unconfirmed. Retain the evidence behind the status. A file named “final-English” is not proof of approval. Similarly, a translation stored beside an approved source may have been created for convenience and never reviewed for operational use.

Establish this foundation through source authority and business meaning before building a conflict detector. A model that accurately identifies a wording difference cannot decide whether the difference is legitimate, outdated or unauthorised without the surrounding evidence. The result should be an issue for the responsible owner, not an invented new policy.

Worked example: a guest limit that changes in translation

Consider a fictional airport-lounge operator preparing a bilingual enquiry assistant. Its approved Arabic rule for membership category B permits one adult guest and one child under twelve, provided the member accompanies them. An English translation awaiting approval says “one guest in total.” The example concerns an invented private service policy, not an aviation rule, an actual lounge or a legal interpretation of consumer rights.

The two texts produce different answers for a member arriving with one adult and one eligible child. The Arabic rule permits that composition under the example’s conditions; the English translation limits the total to one guest. Better retrieval cannot resolve the policy difference. Selecting whichever passage is in the user’s language would make eligibility depend on the language of the question.

The source owner must confirm the intended rule, the governing version and the corrected translation. Until then, the readiness team can exclude the disputed English guidance, route the question to authorised staff or limit the pilot to approved material. It should not silently blend the two clauses into a compromise such as “usually one guest, sometimes two.” That wording would create a third rule with no approved source.

Even after correcting the guest count, other conditions remain. “Under twelve” differs from “twelve and under.” A member’s accompaniment is different from merely having a booking. Category B cannot be substituted for every membership category. The correction must be reviewed as a complete rule, with its predicates and exceptions, not only as a replacement of one number.

Compare meaning using explicit propositions

Break the relevant passage into propositions that a reviewer can test: who is covered, what is permitted, how many, under which conditions, at what time and subject to which exception. Preserve the relationship between them. The guest count applies only when the member accompanies the guests; it is not a standalone allowance available to anyone holding a copied membership reference.

Original bilingual comparison record for the illustrative lounge rule

Proposition	Approved source meaning	Difference to investigate	Decision evidence
Eligible membership	Category B	Translation omits the category	Scope confirmation by the policy owner
Guest composition	One adult and one eligible child	Translation states one guest in total	Corrected approved wording
Age boundary	Child under twelve	Alternative wording includes age twelve	Explicit boundary interpretation
Accompaniment	Member present with the guests	Translation refers only to a reservation	Operational condition and approved explanation
Effective version	Named policy revision and interval	English file retains an earlier interval	Version linkage and publication correction

Preserve differences in obligation and discretion. “May request,” “is eligible,” “is approved” and “will receive” are different states. A translation that turns a request into an entitlement can create a misleading customer promise. The same caution applies to negation, conditions such as “unless,” and qualifiers placed in a footnote or table header rather than in the main sentence.

Use a terminology record for recurring concepts, with an approved definition and contextual examples. The W3C SKOS reference supports language-tagged labels for concepts. Such a structure can organise Arabic and English terms, but it does not certify that two policy passages have equivalent legal or operational effect. The business definition and approval process remain necessary.

Do not mistake formatting and extraction errors for policy disagreements

Inspect the original source before escalating a semantic conflict. A table header may have been detached during extraction. A negation may be missing from OCR text. A footnote can apply to several rows even though a parser associates it with only one. In these cases the underlying approved documents may agree while the machine-readable versions do not.

Retain an inspectable location in each original document. A reviewer should be able to compare the actual text, table layout and qualifying note with the extracted passage. Where the source is a scan, assess whether image quality permits a reliable reading at all. A model’s confident interpretation of unclear characters should not become the expected reference answer without specialist confirmation.

Separate the findings into source-content conflict, version mismatch, translation defect, extraction defect and unresolved interpretation. This classification determines who can fix the problem. The policy owner resolves meaning; the translator corrects the approved linguistic expression; the data engineer repairs extraction or metadata. One team may perform several roles, but the decisions and evidence should remain distinguishable.

Use Arabic OCR and document-quality checks where text recovery is the problem. Do not rewrite a policy merely to compensate for a poor extraction pipeline. Likewise, improving scan quality cannot resolve two genuinely different approved statements. The readiness report should connect each proposed correction to the mechanism it is intended to address.

Recognise legitimate audience and entity differences

An internal procedure and a customer explanation can contain different detail without contradicting each other. Staff may need instructions for checking membership evidence, while customers need a concise account of the entry conditions. Record that relationship instead of demanding sentence-by-sentence equivalence. The shorter explanation must still preserve the conditions material to the question it answers, and the assistant must not extend it into a complete operational procedure merely because it is easier to read.

Two entities under one brand can also have independently approved service conditions. An Arabic document for one entity and an English document for another are not necessarily competing translations. The readiness requirement may be to identify the relevant entity and route the question correctly. Forcing both sources into one bilingual answer can erase an authorised difference. Where the user has not supplied the necessary entity or location context, clarification is preferable to selecting the language-matching document by default.

Finally, establish the meaning of shared terms. A business may use “guest” as a category that includes children in one document and as shorthand for adult companions in another. The remedy is an approved concept definition and a clear relationship to the actual rule. A dictionary translation cannot decide the organisation’s categories. Keep the original use visible until the owner confirms the intended meaning, then update the terminology record and the affected examples together. This avoids making a wording correction that leaves the underlying classification disagreement untouched.

Maintain a conflict register that can support a decision

A conflict item should identify the question affected, the exact source versions and passages, the differing propositions, the potential consequence and the accountable decision owner. Add the current operating restriction, the required resolution evidence and the derivatives that will need updating. A description such as “Arabic inconsistent” gives neither the translator nor the sponsor enough information to act.

Preserve the source passages before changing them. Record who made the resolution, the scope of that authority and the effective date of the corrected wording. Where the issue affects existing customer-facing material, the owner should decide what review or communication is needed. The technical team should not quietly erase the earlier version and leave downstream teams unable to explain a previous answer.

Unknown authority must remain visible. If two approved versions appear inconsistent and no governing relationship is established, the register should not select a winner through a confidence score. Record what additional evidence or decision is required. A narrow pilot may proceed on unaffected questions, but its exclusion list needs to reach the retrieval corpus, evaluation cases and user guidance.

Connect the register to catalogue and lineage records. A corrected clause can affect an extracted-text store, translated summary, question-answer example, search index and cached answer. Closing the language issue requires verification that the intended approved material is available where the pilot will use it, not only that an editor has saved a new document.

Build paired cases from facts, not translated questions alone

A bilingual test pair should contain the same underlying facts and decision question in both languages. Translating the wording of a question is not sufficient if one version accidentally adds an eligibility condition. Establish the shared case facts separately, then have competent reviewers confirm that the Arabic and English questions ask for the same decision or explanation.

For the lounge example, an original sixteen-pair set could cover six guest compositions, four age-boundary cases, two accompaniment cases, two cases with essential facts missing and two policy-version cases. Those counts are illustrative coverage choices. The resulting thirty-two answers are not thirty-two independent policy situations: each pair deliberately shares facts, and several pairs share the same governing rule.

Write the acceptable reasoning and uncertainty for each case before evaluating a model. A case missing the membership category should ask for clarification or explain the boundary, not assume category B. A question about a historical visit may require the version effective at that time. An answer can differ naturally in phrasing between languages while preserving the same conditions, limitations and authorised source.

Inspect retrieval as well as the generated answer. An English question may retrieve an older English translation while the Arabic question retrieves the current approved original. A final answer that happens to be correct can conceal that source-selection defect. Arabic retrieval testing should therefore distinguish retrieval coverage, source applicability and answer support rather than collapsing them into one fluency judgement.

Choose an operating policy for mixed-language evidence

Agree whether the service may answer in one language using an authoritative source in another. If so, define how the translated explanation is reviewed, how the source is shown and which questions need specialist escalation. A bilingual retrieval architecture can support this behaviour, but a technical capability does not establish that the organisation has accepted the associated interpretation risk.

Decide how much of the evidence must be visible to the reviewer. A person approving an English answer may not be able to evaluate an Arabic exception clause without suitable language support. Providing a link that the reviewer cannot understand does not make oversight effective. Match reviewer competence, translated evidence and decision responsibility to the actual task.

Search aliases should help users find the right concept without weakening the rule. Dialect terms, transliterations, spelling variants and product names may all be useful query forms. Keep them connected to a controlled concept identity and approved definition. An alias for a common phrase should not cause a different membership category, entity or policy condition to be treated as equivalent.

Be careful with numbers and dates. A date such as 03/04 is ambiguous without a documented convention; neither the question language nor the user’s location resolves it reliably. Identify the calendar, time zone where relevant and effective-date rule. Preserve original representations and controlled conversions. Similarly, a translated amount must retain its currency and whether it represents a limit, an estimate or a payable charge.

Use acceptance evidence that measures the intended consistency

Define consistency at the proposition level. Does each answer identify the same eligible category, count, age boundary, accompaniment requirement and version under the shared facts? Does each withhold a conclusion when the same essential fact is missing? A text-similarity score may reward similar wording while missing a changed negation, and may penalise good natural translations that express the same rule differently.

Separate source readiness from demonstrated application performance. Before a prototype exists, reviewers can establish approved source pairs, resolved conflicts and expected case outcomes. They cannot claim that a future model will retrieve or apply them correctly. After the prototype is available, evaluate the end-to-end behaviour under the agreed conditions, preserving failures and their causes rather than retrospectively adjusting the expected answers to match the model.

Use severity to guide the next step. A stylistic difference may be acceptable after linguistic review. An altered eligibility condition may block the affected customer guidance. An unapproved translation may permit an internal exploratory exercise but not publication. State the permitted activity and the required evidence for moving beyond it; a blanket bilingual readiness percentage conceals these materially different outcomes.

Retest the corrected proposition and nearby cases. Fixing “one guest in total” should not accidentally remove the accompaniment condition or change the age boundary. Include the old wording in a controlled regression case where useful, while ensuring that obsolete material is not reintroduced as current authority. The purpose is to verify the corrected behaviour, not to assume that editing one sentence repairs every dependent answer.

Make bilingual ownership sustainable

Assign responsibility for source approval, translation approval, terminology maintenance and technical publication. These responsibilities may be held by different people. Define what happens when the approved original changes before the translation is ready: which questions are withheld, which existing translation is marked historical and how users are told about the limitation.

Include source changes, new audience groups and revised operational definitions as triggers for review. A translation can remain linguistically accurate while its business context changes. For example, a membership category can be replaced or a service introduced in another entity with different conditions. Detecting the new words alone will not identify the changed applicability.

Place unresolved work in the data remediation backlog with an owner, required evidence and a verification method. The readiness decision should identify the approved language pairs, excluded questions and the procedure for future disagreements. Bring one disputed pair, its approval history and the intended user questions to a readiness assessment discussion about bilingual source conflicts. That evidence allows the assessment to distinguish a translation task from a policy decision that must be resolved before reliable bilingual use.