Skip to content

Systems and records

Media localization approval dataset: more than a quality score

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

A translation approval task should preserve the decision as well as the score: the locale and policy version that applied, the terminology used, each requested revision and its reason, which flagged issues were waived and by whom, the approver's authority, and the rights, territory and medium limits on the approved text. A media localization approval dataset built that way shows why a line was accepted and where it may be reused.

Provenance trail

Media localization approval

This page covers media localization approval dataset: more than a quality score.Illustrative

Key takeaways

  • A quality score measures errors against a threshold; approval is a decision by someone with authority, and the two often disagree.
  • Record locale as a BCP 47 tag and point to the style guide, brand book and glossary versions in force on the approval date.
  • Every waiver of a logged issue needs the issue, the reason and an approver entitled to accept it.
  • Rights, territory, medium and clearance conditions decide whether approved text may be delivered or reused at all.
  • Deadline sign-offs with open defects should be labeled as such, never treated as preferred output.

What should a translation approval task preserve beyond a language quality score?

A translation approval task should preserve the decision, not just the grade: which locale policy and terminology version applied, who approved the line and with what authority, every revision the client or creative lead requested and why, which flagged issues were waived on purpose, and the rights conditions that limit where and how the approved text may be used. A score says how many errors a reviewer counted; the approval record says what the business accepted.

That distinction matters most in media work, where the best localized line is often not the closest one. A trailer tagline, a dubbed joke or an on-screen graphic gets rewritten until it lands for an audience, and the person who signs it off may be a brand manager rather than a linguist. If a localization archive keeps only the final text and a quality number, anyone later building a media localization approval dataset loses the reason a line was chosen and the limits on reusing it.

This guide gives the record a creative operations team can keep per segment and per asset: a task specification with locale policy, terminology and approval fields, a table of conditions that change what may be delivered or reused, an illustrative worked case, one meaningful exception and the points where the method stops being enough.

Why is a passing score not the same thing as an approved translation?

A passing score is a measurement against a threshold; approval is a decision by someone entitled to make it, and the two can disagree in both directions. Standards keep them apart. The MQM Council's scorecard counts errors by severity, with an example weighting of 0 for neutral, 1 for minor, 5 for major and 25 for critical, and asks evaluators to judge severity by imagining the error surfacing after publication.

XLIFF, the OASIS exchange format most translation tools read and write, records the two ideas in different places in its 2.2 specification. Its state attribute moves a segment through initial, translated, reviewed and final, while a separate approved attribute says whether the segment holds a translation suitable for merging back into the original format. Its ITS module can carry a quality rating score and a threshold for that score. Nothing in the format makes approval follow from the score; a workflow decides that.

In practice, a segment can score well and still be refused because it uses a rival product's slogan or breaks a character's established voice. Another can carry a logged major error, such as a deliberate mistranslation of a pun, and still be approved because the creative lead wanted the joke to work. A dataset that keeps only the score teaches a model the wrong lesson in both cases.

Which locale decisions belong in the record instead of in someone's memory?

Record the exact locale the text was approved for, the policy document and version that governed it, and the terminology list in force on the approval date. Each of these changes over time, and a line approved under last year's policy can be wrong under this year's.

Locale should be a language tag, not a free-text country name. BCP 47 (RFC 5646) uses es-419 for Spanish appropriate to the Latin America and Caribbean region, which is a different target from es-ES even when the words overlap. Storing the tag lets a later reviewer group, filter and compare approvals without guessing what "Spanish" meant on a given project.

Locale policy is more than spelling. Streaming platforms publish a separate timed text style guide per language; Netflix's partner help center, for example, has per-language guides that cover points such as vocabulary for the target market, when italics are used, how on-screen text is marked and whether foreign song lyrics are subtitled. A client's own brand book adds character names, product names that stay in English, register and words the brand never uses. The approval record should point to the version of each document the translator and reviewer were working from.

The locale policy, terminology and approval task specification

Use the specification below as the record for each approved segment or asset. Fields can sit in a sidecar file next to an XLIFF or subtitle file, in a translation management system's custom fields, or in a review tracker; what matters is that each field is filled at the time of the decision and linked by a stable segment or asset ID. Example values are illustrative.

The locale policy, terminology and approval task specification
FieldWhat it holdsIllustrative valueWhy it matters later
Asset and segment IDStable ID of the title, asset, cue or string, plus source file hashPromo-ES-0042, cue 17Joins the approval to the exact source text and timing it applied to
Source and target localeBCP 47 tags for source and targeten-US to es-419Separates regional targets that share a language
Policy versionsStyle guide, brand book and client instructions in force, each with version and dateClient brand book v6; platform style guide revision of recordShows which rules the decision followed
Term decisionsGlossary entries applied, plus any term the reviewer added, changed or rejectedProduct name kept in English; tagline verb changedLets terminology be audited and reused without re-deriving it
Quality findingsEach logged issue with category, severity and whether it was fixed, waived or disputedOne major accuracy issue, waived as intentional wordplayKeeps the score but explains the outcome
Revision historyEach requested change with requester role, reason code and the text before and afterRound 2: creative lead asked for shorter line to fit lip movementPreserves why the final text differs from the first draft
Approver and authorityRole of the approver, what they were authorized to approve, dateRegional marketing lead, approves promotional copy for es-419Separates linguistic sign-off from creative or legal sign-off
Approval decisionApproved, approved with conditions, rejected or escalated, plus conditionsApproved for streaming promo onlyA score never records a conditional approval
Rights and territoryWho owns the source and translation, permitted territories, media, term and any talent or music restrictionsClient owns both; es-419 territories; online promo onlyDetermines whether the line may be delivered, reused or licensed at all
Evidence linksPointers to the review comment, email or ticket that records each decisionReview thread link, round 2 approval emailLets a second reviewer verify the record instead of trusting it

Which approval conditions stop a localized line from shipping or being reused?

Some conditions change only the text; others change whether the text can leave the project. Treat any condition in the second group as a hard gate on delivery or reuse, not as a note.

Ownership is the first gate. The U.S. Copyright Office lists translations among derivative works, and only the owner of the underlying work can prepare a derivative or authorize someone else to; a translator's own claim covers only the new material. For a localization vendor, that usually means the client controls the source and often the translation by contract, so an archive of approved lines is not automatically the vendor's to reuse or license. Counsel should read the master services agreement and statements of work before any reuse.

Which approval conditions stop a localized line from shipping or being reused?
Condition in the recordEffect on deliveryEffect on later reuse or licensingWho decides
Approved with no conditionsDeliver as approvedStill depends on ownership and the services agreementClient approver, then counsel for reuse
Approved for one medium or campaignDeliver only to that mediumExclude or tag the line so it cannot be used elsewhereClient approver
Territory or embargo restrictionDeliver only to listed territories after the embargo dateHold until the restriction lapses and the agreement allows reuseClient legal or distribution team
Talent, music or third-party clearance pendingDo not deliver the affected cueExclude; clearance for broadcast is not clearance for reuseRights and clearances team
Rejected after reviewDo not deliverKeep only as a labeled negative example if the agreement permitsCreative lead or client approver
Score below threshold but approvedDeliver with the waiver recordedKeep the waiver reason so the segment is not mislabeled as an errorApprover named in the record

Worked example (illustrative): one promo line, two Spanish locales

This example is invented to show the record; no real client, title or reviewer is described. A streaming service promo in U.S. English ends with a tagline built on a pun about a heist. The localization vendor delivers two Spanish versions, es-419 and es-ES, through a translation management system that exports XLIFF.

Round 1: both translations are literal. The language reviewer logs no errors, and both segments reach the reviewed state. The regional marketing lead for es-419 rejects the line because the pun disappears and the result reads flat. The es-ES approver accepts the literal version because the brand book for Spain asks for a neutral register on promotional copy.

Round 2: the es-419 translator writes a new line using a local idiom about stealing the show. The language reviewer logs one major accuracy issue, since the line no longer says what the English says, which pushes the segment's quality rating under the project threshold. The marketing lead approves it anyway and records the reason: intentional transcreation requested in round 1. The approval is conditional, limited to online promotion in the listed territories, because a music cue in the same spot is cleared only for that use.

The final record for es-419 therefore holds a below-threshold score, an approved decision, a waiver with a reason code, two revision rounds with the requester's role, a terminology note that the series title stays in English, and a territory and medium restriction. A dataset built from the score alone would label the best line in the project as a failure, and a dataset built from the text alone would lose the restriction that keeps it out of a trailer.

When should a failing segment stay approved, and when should it be dropped?

Keep a below-threshold segment as approved only when the waiver names the issue, the reason and an approver with authority over that kind of content. Drop or relabel it when any of those three is missing.

The common legitimate waiver is transcreation: marketing copy, humor, song adaptations and titles where the brief asks for effect rather than fidelity. A waiver that cites the brief and comes from the creative or marketing approver is a real decision and a valuable training signal. A waiver that comes from the translator, or that says only "approved per client" with no link, is not.

The exception that should not be kept as approved is the deadline waiver. If a reviewer signs off a line with an open terminology or accuracy error because the delivery date arrived, record it as approved for delivery with an open defect, not as an accepted translation. Treating those lines as preferred output would teach a model that missed deadlines produce good copy.

  • Keep as approved: waiver names the issue, cites the brief or policy, and the approver had authority over that content type.
  • Keep but relabel as delivered with open defect: waiver driven by schedule, or the defect is still listed as unresolved.
  • Drop from any reuse set: no traceable approver, missing rights fields, or a pending clearance on the same cue.
  • Escalate to the rights reviewer: the client relationship has ended or the agreement says nothing about reuse of approved text.

Where does this task specification stop being enough?

The specification records decisions; it does not settle who owns them, and it cannot repair approvals that were never written down. Narrow the scope, or bring in a rights reviewer, in the situations below rather than filling gaps by inference.

Process standards help but do not replace the record. ISO 17100 defines revision as a bilingual examination of target content against source content for its suitability for the agreed purpose, and treats the reviser as a role distinct from the translator. A team can follow that process and still lose the creative approval that happened later in an email thread. The specification is meant to catch that later decision.

  • Approvals were given verbally or in calls with no written trace: keep the text, mark approval as unverified and exclude it from any approved set.
  • The archive mixes clients whose agreements differ: split by client and review each agreement before combining anything.
  • Segments contain names, voices or likenesses of performers: treat talent rights as a separate review, since a translation approval says nothing about them.
  • Subtitles for the deaf and hard of hearing, legal disclaimers or regulated product claims are in scope: these follow their own rules and need a reviewer who knows them.
  • The quality findings came from an automated estimator rather than a human: label the source of each finding so model output is not mistaken for reviewer judgment.

Before a localization archive is offered for licensing

Approved localization records with their reasons can be valuable to AI developers building translation, evaluation and preference models, but only the records a company controls under its agreements can be offered. The rights and territory fields in the specification are the first thing a licensing review will read.

SourceX helps localization businesses license selected records as AI datasets and manages the rights review, contracting, preparation and delivery. No record moves before the company approves the scope and terms and signs. The opening step is a short assessment, with nothing uploaded, which is a sensible moment to ask whether a localization archive belongs in scope at all.

Frequently asked questions

Is an MQM score enough to decide whether a translation is approved?

No. An MQM score counts errors by category and severity against a threshold, which is a measurement. Approval is a decision by someone with authority, and it can accept a line with logged errors, such as intentional transcreation, or reject a clean one for brand reasons. Keep both, and record why they differ.

What is the difference between reviewed and approved in XLIFF?

In the XLIFF 2.2 specification the state attribute moves a segment through initial, translated, reviewed and final, while a separate approved attribute says whether the segment's translation is suitable for merging back into the original format. The format does not tie one to the other. Your workflow should record who set each and when.

Who owns an approved translation of a client's content?

Under U.S. copyright law a translation is a derivative work, and only the owner of the original can authorize one; the translator's claim covers only the new material. In localization work the services agreement usually decides who holds the translation. Counsel should read the agreement before any approved text is reused or licensed.

Should rejected translations be kept in a localization dataset?

They can be useful as negative examples when the record says who rejected them and why, and when the agreement allows reuse. Without a reason code, a rejected line teaches little. Drafts rejected only because of a deadline or a clearance problem should not be labeled as poor translations.

How should locale be recorded for localized media?

Use BCP 47 language tags, such as es-419 for Latin American Spanish or es-ES for Spain, rather than country names. Pair the tag with the version of the style guide and brand book that applied. The same words can be correct for one locale and wrong for the other.

Sources

  • The MQM Council's scorecard example uses neutral, minor, major and critical severity levels weighted 0, 1, 5 and 25, and asks evaluators to judge severity by imagining the error found after publication. Source
  • XLIFF 2.2 defines a segment state of initial, translated, reviewed or final as a simple linear state machine and prescribes no particular workflow. Source
  • XLIFF 2.2 defines a separate approved attribute indicating whether a segment holds a translation suitable for merging back into the original format. Source
  • XLIFF 2.2's ITS module carries the ITS localization quality rating score threshold attribute. Source
  • RFC 5646 (BCP 47) gives es-419 as Spanish appropriate to the UN-defined Latin America and Caribbean region. Source
  • Netflix's partner help center publishes a separate timed text style guide for each language, covering rules such as target-market vocabulary, italics and on-screen text. Source
  • The U.S. Copyright Office lists translations as derivative works; only the copyright owner can prepare or authorize one, and a translator's claim covers only the new material. Source
  • ISO 17100:2015 sets requirements for translation services and defines revision as a bilingual examination of target content against source content for its suitability for the agreed purpose. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies