Privacy and preparation
Professional translation revision data: what a rejected term reveals
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
A rejected translation reveals a terminology constraint when the record keeps the struck-out rendering, the replacement, the reviser's reason and the authority behind it, such as a termbase entry or client instruction. Professional translation revision data built that way shows where fluent, correct-looking choices were still wrong for a client and scope, which approved term lists alone never show. Clear client and linguist rights first.
Countries and languages
Professional translation
Key takeaways
- A before and after pair shows a preference; the reason code and cited authority are what make it a terminology constraint.
- Label each rejection by type, because style fixes, accuracy fixes and termbase defects teach different lessons from term conflicts.
- Keep the rejected term verbatim and scope every constraint by client, document type and date.
- Bilingual files with change tracking keep the rejected rendering; final target files and translation memories usually do not.
- Client source text, linguist renderings and reviser comments can each carry different rights, so map them before licensing.
How can a rejected translation reveal a terminology constraint?
A rejected translation reveals a terminology constraint when three things are recorded together: the rendering the reviser struck out, the rendering that replaced it, and the reason and authority behind the change. The before and after pair alone only shows that someone preferred one wording. The reason code and the authority, such as a client termbase entry or a style guide clause, turn that preference into a rule with a scope.
That is why professional translation revision data is worth more than its row count suggests. A termbase lists the terms a client approved. A revision log shows where a fluent, dictionary-correct choice was still wrong for this client, product line or document type, which is exactly the kind of error a general translation model keeps making. The rejected rendering is the negative example that the approved term list never contains.
The catch is that most revision history is noisy. Revisers also fix grammar, style, omissions and their own preferences, and many tools store the change without any reason at all. This guide sets out which rejections carry a real constraint, the record that captures one, and the rights screen a language service provider should run before any of it leaves the firm.
What a revision record shows that a termbase or a post-edit does not
A revision record is the only artifact that pairs a plausible human translation with an expert's documented correction of it. That combination is different from the two closest neighbors, and buyers will ask which of the three they are looking at.
A termbase is concept-oriented. TBX-Basic, the dialect maintained by LTAC Global, gives each term an administrative status such as preferred, admitted, deprecated or superseded. That status tells a model which term is approved, but not which wrong choices translators actually make under deadline, or in which contexts an admitted synonym was rejected anyway.
Machine translation post-editing data is a separate workflow with its own standard. ISO 17100 sets requirements for the processes, resources and other aspects of a quality translation service, and its scope excludes raw machine translation output plus post-editing, which ISO 18587 covers. A post-edit corrects machine output; a revision corrects a human translator. The errors differ, and so does what a model can learn from them.
The file formats help you tell them apart. XLIFF 2.1, the OASIS interchange format, gives each segment a state that moves from initial to translated to reviewed to final, and its change tracking module stores revision elements with an optional author, datetime and version. When a bilingual file still carries those revisions, you have the raw material. When it only carries the final target, the rejection is gone.
Which rejections are terminology constraints, and which are something else?
Only rejections whose stated reason points to a terminology resource, a client instruction or a documented domain convention should be labeled as terminology constraints. Everything else is still useful revision data, but it teaches a different lesson and must carry a different label.
The MQM error typology draws the line clearly. Its terminology dimension separates a term that is inconsistent with a specified termbase or other resource from inconsistent use of several terms for one concept, and from a term that is simply wrong for the domain. MQM's list of root causes is just as useful here: it includes no termbase provided, the wrong termbase applied, a source term missing from the termbase and an incorrect target equivalent in the termbase. In other words, a rejection can expose a defect in the termbase itself rather than a translator error.
| Rejection pattern | What the evidence usually shows | Is it a terminology constraint? | Label to use |
|---|---|---|---|
| Reviser replaces a synonym with the client's approved term and cites the termbase entry | Term conflicts with a specified terminology resource | Yes, scoped to that client and resource | termbase_conflict |
| Reviser makes one term consistent with earlier segments in the same document | Several terms used for one concept where consistency was required | Yes, if the project specification required consistency | inconsistent_use |
| Reviser swaps a general-language word for the domain's accepted term with no resource cited | Domain convention, but authority undocumented | Only after a terminologist confirms the convention | domain_term_unverified |
| Reviser restores a term the translator changed, citing a forbidden or deprecated status | Deprecated or superseded term used | Yes, and the rejected term is the negative example | deprecated_term |
| Reviser notes the termbase entry itself is wrong or missing | Defect in the resource, not the translation | No; route it to the termbase owner | termbase_defect |
| Reviser rewrites for tone, length or word order | Style or fluency preference | No | style_or_fluency |
| Reviser corrects a mistranslation or omission of meaning | Accuracy error | No, unless the meaning error came from a term choice | accuracy |
The revision and rationale record: change and context fields
Capture each rejection as one row that keeps the source segment, both target renderings and the context a reviser saw. If a field cannot be filled from the bilingual file or the project record, leave it empty and flag it rather than reconstructing it from memory.
| Field | What it holds | Where it usually comes from | Required? |
|---|---|---|---|
| revision_id | Stable ID for this change, unique across projects | Generated at export | Required |
| project_ref and file_ref | Pseudonymous project and file keys, never client names | Project management system, mapped at export | Required |
| source_lang and target_lang | Language tags with region where it matters, such as en-US and de-DE | XLIFF srcLang and trgLang attributes | Required |
| source_segment | Full source sentence or segment, after redaction | Bilingual file | Required |
| rejected_target | The translator's rendering as submitted to revision | XLIFF change tracking revision or tool history | Required |
| accepted_target | The rendering after revision | Final bilingual file | Required |
| changed_span | The exact term or phrase that changed, in both versions | Diff of the two targets, checked by a person | Required |
| context_window | Neighboring segments and document type, such as operator manual or UI string | Bilingual file and job ticket | Recommended |
| segment_state_trail | State history such as translated, reviewed, final, with dates | XLIFF state attribute and tool audit log | Recommended |
The revision and rationale record: reason, authority and rights fields
The second half of the record is what turns a diff into a constraint. Without a reason and an authority, a buyer cannot tell a client rule from one reviser's habit, and the row should be labeled accordingly.
At a glance
The revision and rationale record: reason, authority and rights fields
| Field | What it holds | Where it usually comes from | Required? |
|---|---|---|---|
| reason_code | One label from the table above, such as termbase_conflict or deprecated_term | Reviser comment, LQA form or later annotation | Required |
| reason_text | The reviser's own words, redacted, in the language they wrote them | Comment thread or LQA report | Recommended |
| authority_type | Termbase entry, client style guide, client query answer, regulation or domain reference | Project reference materials | Required for constraint labels |
| authority_ref | Entry ID, guide section or query ticket number, with version and date | Termbase export, style guide revision table | Required for constraint labels |
| term_status_after | Status of the accepted term in the resource, such as preferred or admitted | TBX administrative status | Recommended |
| scope_of_rule | Client, product line, document type and date range the rule covers | Authority document and project dates | Required for constraint labels |
| reviser_role | Role and qualification level, never a name | Vendor or staff records, generalized | Recommended |
| dispute_flag | Whether the translator contested the change and how it was resolved | Feedback and query threads | Recommended |
| rights_basis | Contract or permission that allows this row to be licensed | Rights screen, described below | Required |
Turning one rejection into a constraint statement a model can use
Write every confirmed constraint as a short, testable statement that names the context, the required term, the rejected term and the authority. A model, an evaluator and a human reviewer should all be able to check a new translation against it without reading the original project.
A workable template is: in this scope, when the source term appears in this context, use the accepted term with this status; do not use the rejected term, because of this reason; authority and version; valid from this date. Mapping the accepted term's status to the TBX-Basic values keeps the statement compatible with the termbase the client may already hold.
- Scope first: a constraint that holds for one client's operator manuals may be wrong for the same client's marketing copy.
- Keep the rejected term verbatim, including inflection, because that is the form a model will actually produce.
- Record context triggers, such as a product family or a regulatory section, when the same source term is translated differently elsewhere.
- Mark a constraint as superseded, never delete it, when the client changes the approved term, so older revisions stay explainable.
- Link each statement back to every revision_id that supports it; one rejection is an observation, several consistent rejections are a pattern.
- Send termbase_defect rows to the termbase owner as a correction request instead of turning them into training rules.
Worked example (illustrative): six rejections in a maintenance manual batch
Consider an illustrative batch: a 40-segment chapter of an equipment maintenance manual, translated from U.S. English by one linguist and revised by a second. The reviser made six changes. The terms are shown as term A and term B so the example does not depend on any particular language pair.
Two changes replaced the translator's term B with the client's preferred term A for the same component, citing the client termbase entry. Both become one termbase_conflict constraint, scoped to that client's maintenance documentation, with term B kept as the negative example.
One change made a part name consistent with the three earlier segments where the translator had used it correctly. That is inconsistent_use, and the constraint only applies because the job specification required consistent terminology.
One change replaced a deprecated product name that the client retired the previous year. That is deprecated_term, and the scope must start on the retirement date, since older manuals legitimately used the earlier name.
The last two changes were a word-order rewrite and a restored omission. They belong in the dataset as style_or_fluency and accuracy rows, but they produce no terminology constraint.
The exception arrived later. In a safety section of the same manual, the client's query answer said term B was correct there because it matched the wording on the certified warning label. The constraint statement gained a context trigger: use term A except in safety warnings that reproduce label text. A dataset that dropped the query thread would have taught a model the wrong rule for exactly the segments where the error matters most.
Clearing rights before revision history leaves the language service provider
Revision data usually involves at least three rights holders, so the language service provider should map who holds what before anything is licensed. Holding the bilingual files is not the same as having the right to license them.
Under U.S. copyright law, a translation is listed as one form of derivative work, and the copyright in a derivative work covers only the material its author contributed, not the preexisting work. The client's source text therefore stays the client's, and the translation and revision may be covered by work-for-hire or assignment terms in the client contract. Counsel should read those terms before any segment is included.
| Item in the record | Who may hold rights or interests | Evidence to collect | Typical action |
|---|---|---|---|
| Source segments | The client that owns the source documents | Client master agreement, confidentiality terms, any reuse clause | Include only with written client permission or exclude |
| Translator and reviser renderings | Provider, client or freelancer depending on assignment terms | Freelancer agreements, client assignment clauses | Confirm chain of title for each linguist and client |
| Termbase entries and style guides | Usually the client | Delivery terms for reference materials | Cite by ID only unless the client permits release |
| Reviser comments | Provider; may contain personal or client details | Internal policy, comment content review | Redact names, contact details and client identifiers |
| Project metadata | Provider; may reveal client lists and volumes | Project management export | Pseudonymize clients and generalize dates |
Rejections that should never become training constraints
Some rejections are real and still unsafe or misleading to generalize. Keep them out of the constraint set, label them for exclusion or route them to a specialist before a buyer sees them.
- One reviser's habitual preference with no authority cited and no second reviser making the same change.
- Changes the translator disputed where the dispute was never resolved, or was resolved by a project manager rather than the client.
- Regulated text, such as product safety labels, legal notices or financial disclosures, where the authority is a regulation or certified wording that a domain specialist should confirm.
- Segments whose source text is confidential to the client, even after redaction, because the term itself can reveal an unreleased product or process.
- Rows where the termbase was later corrected, unless the record shows which version applied at the time.
- Very small clients or niche products, where the combination of terms alone could identify the client to a buyer.
If your revision history could become a licensed dataset
For a language service provider, a cleaned revision and rationale dataset is a candidate for licensing only after the client and linguist rights above are settled. SourceX licenses selected business records directly from providers, handling sourcing, rights review, contracting, data preparation, delivery and payment.
Nothing moves before the provider approves the scope and terms and signs, and that agreement governs how SourceX may later use, sublicense and distribute the de-identified dataset. On geography, SourceX works with U.S., Canadian and EU providers that had 50 or more full-time employees at their peak, contractors excluded, and can consider providers in the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia only when at least 90% of the offered records are in English, as measured on the actual dataset. A translation archive with non-English target segments needs that measurement before anything else. Its first step is a brief company assessment, and no records need to be uploaded for it.
Frequently asked questions
Is translation revision data the same as post-editing data?
No. Revision corrects a human translator's work, while post-editing corrects machine translation output, and ISO 17100 explicitly leaves raw machine output plus post-editing to a separate standard, ISO 18587. The errors differ, so label the two record types separately even when the same tool stored both.
Do revisions without a reviser comment have any value?
Yes, as general revision pairs for fluency or accuracy work, but not as terminology constraints. Without a reason and an authority, a buyer cannot tell a client rule from one reviser's preference. You can add reasons later only if a terminologist checks each change against the reference materials of that project.
Which file formats keep the rejected rendering?
Bilingual XLIFF files that retain change tracking keep it, since XLIFF 2.1 stores revision elements with optional author, datetime and version. Many tools also keep segment history in their own databases. A delivered target file or a translation memory usually keeps only the final accepted text.
Who owns a translation that a client paid for?
It depends on the contract. Under U.S. law a translation is a derivative work, the client keeps rights in its source text, and the translation's own copyright covers only the new material. Work-for-hire or assignment clauses may give the translation to the client, so counsel should review each client agreement.
Can a rejected term reveal confidential client information?
Yes. A new product name, a retired component or an internal code can identify a client or an unreleased plan even when the surrounding text is redacted. Review the changed span itself, not only the segment, before any row leaves the firm.
How many rejections make a reliable terminology constraint?
There is no fixed count. One rejection with a cited termbase entry is a documented rule for its scope, while several consistent rejections without a citation are a pattern a terminologist should confirm. Link every constraint to all the revisions that support it so a reviewer can judge the evidence.
Sources
- The MQM Full typology's terminology dimension distinguishes a term that is inconsistent with a specified termbase or other resource from inconsistent use of several terms for one concept. Source
- MQM's list of error root causes includes no termbase provided, the wrong termbase applied, a source term missing from the termbase and an incorrect target equivalent in the termbase. Source
- XLIFF 2.1 defines segment states initial, translated, reviewed and final, and a change tracking module whose revision elements carry optional author, datetime and version attributes. Source
- The TBX-Basic dialect defines administrative status values for preferred, admitted, deprecated and superseded terms. Source
- ISO 17100:2015 sets requirements for the core processes, resources and other aspects of a quality translation service, and its scope excludes raw machine translation output plus post-editing. Source
- ISO 18587:2017 sets requirements for post-editing of machine translation output. Source
- Under 17 U.S.C. 101 a translation is a form of derivative work, and under section 103(b) the copyright in a derivative work extends only to the material contributed by its author. Source
See if your company qualifies
A short company assessment. No data uploads are needed.