OnlineDocTranslatorGuides

Answer

Translation Quality: What It Means and How Production Teams Measure It

Translation quality is the measurable degree to which a target-language file preserves source meaning, terminology, locale and style — scored per 1,000 words

Translation production teams treat quality as a numeric property, not a verdict — every finished file crosses an LQA gate that converts defects into a per-1,000-word score before release. This page situates that score inside a certified, ISO 17100-aligned workflow and sits downstream of the broader AI Proofreading Tools Worth Using in a Translation Workflow cluster.

Key points

  • Release thresholds on certified and regulated lanes require under 2 errors per 1,000 words; product, app and support lanes require under 3; marketing lanes require under 5.
  • The reviewer scoring a file cannot be the translator of the same segments, otherwise the per-1,000-word score is not auditable.
  • Translation quality is distinct from translation preference (taste on synonymous rewordings) and from localization fit (the full locale build around the sentence-level transfer).
  • The US LQA gate price band sits at $40–$90 per 1,000 words against an MQM-style scorecard aligned to ISO 17100.

What is translation quality in a production context?

Translation quality is the measurable degree to which a target-language file preserves the meaning, terminology, locale conventions, and style of the source text, scored per 1,000 words against a weighted MQM-style rubric aligned to ISO 17100 before release. The score is computed inside an LQA gate, compared to a published lane threshold, and used as a release decision — not as a stylistic opinion.

The term covers four adjacent concepts that buyers and junior producers confuse. The table below separates them so the rest of the page can stay on the measurable property.

TermWhat it isWhen it applies
Translation qualityDefect count per 1,000 words against an MQM-style rubricAt the LQA gate, before release
Translation preferenceReviewer taste on semantically equivalent rewordingsDuring style-guide revision, not release
Localization fitEngineering, legal, cultural and layout build around the sentenceAt locale QA, after translation quality clears
ReadabilityTarget-audience comprehension score (e.g. Flesch)During style authoring, not release

Why is translation quality a measurable property, not a feeling?

Every defect in a target-language file can be tagged against an MQM category and severity level, summed to a per-1,000-word score, and compared to a published lane threshold. Subjective impressions of a "good translation" do not clear a production gate. A reviewer who rejects a file without a tagged defect record cannot block release — the score is the control, not the mood.

The measurement chain has four fixed steps:

  1. Tag each defect against one of five MQM categories.
  2. Assign a severity: critical, major, or minor.
  3. Sum the weighted defects for the file.
  4. Divide by source-word volume in thousands.

How does translation quality differ from translation preference?

Translation quality counts defects against a rubric; translation preference captures reviewer taste on wordings that are semantically equivalent. A preference edit does not change the per-1,000-word error score, so it does not change translation quality — it changes house style. Preference belongs in the style guide revision cycle, not in the LQA scorecard.

  • Quality edit: a mistranslated clinical term flagged as a critical accuracy defect.
  • Preference edit: "utilize" rewritten as "use" where both are semantically equivalent.
  • Quality edit: a currency left in source locale, flagged as a major locale defect.
  • Preference edit: a comma replaced by a semicolon with no grammar rule broken.

What are the core dimensions of translation quality?

Translation quality has five production dimensions drawn from the MQM typology: accuracy, fluency, terminology, locale convention, and style. Each defect is classified into one of these buckets and assigned a severity level — critical, major, or minor — before the per-1,000-word score is computed. The weighted scorecard assigns Accuracy 30%, Terminology 20%, Functional integrity 20%, Style 15%, and Locale conventions 15%.

DimensionWeightDominates on laneExample defect
Accuracy30%Certified / legalMistranslated clinical finding
Terminology20%Product / app / supportLocked UI label rephrased
Functional integrity20%Product / appMissing or broken tag
Style15%MarketingBrand voice register mismatch
Locale conventions15%Marketing / consumerDate in source-locale format

How is accuracy defined as a translation quality dimension?

Accuracy is the dimension that captures mistranslation, omission, addition, and untranslated segments — defects that change the meaning of the source text. On certified and legal lanes, accuracy carries the highest MQM weight at 30% and a single critical accuracy error can fail an otherwise clean file.

  • Mistranslation: a numeric dosage changed from 5 mg to 50 mg.
  • Omission: a required disclaimer dropped from a consent form.
  • Addition: a clause not present in the source text.
  • Untranslated: a source-language segment left in the target file.

How does fluency contribute to translation quality?

Fluency captures grammar, punctuation, spelling, and syntax defects in the target language that make the translated text read as non-native, even when the meaning is correct. Fluency failures do not change what the source said; they change whether the target audience trusts the output. On certified lanes a fluency defect sits below accuracy in severity; on marketing lanes it climbs.

Why is terminology its own translation quality dimension?

Terminology is scored separately because locked brand terms, approved glossaries, and regulatory boilerplate must match exactly across every segment of a project. Terminology weight sits at 20% on an MQM-style scorecard and dominates product, app and support lanes where inconsistency breaks the UI. A product menu label translated two ways inside one build registers as two separate terminology defects.

What does locale convention cover in a translation quality score?

Locale convention covers date, time, number, currency, address and measurement formats that must match the target locale, not the source locale. A locale defect is a quality error even when the underlying number is mathematically correct, because the format misreads to the target audience.

  • Date: "03/04/2025" left in US format inside a UK-English file.
  • Currency: "$1,000" in a target file priced in euros.
  • Number: decimal comma not switched to decimal point for US output.
  • Measurement: meters left in a US consumer file calling for feet.
  • Address: ZIP code format preserved where the target uses postcodes.

How is style weighted in a translation quality rubric?

Style covers register, tone, voice, and adherence to a published style guide, weighted at 15% on an MQM-style scorecard. Style defects dominate marketing lanes, where brand voice carries more risk than a single terminology slip, and sit near the floor on certified lanes where accuracy rules. A style defect without a documented rule in the style guide is reclassified as preference.

How is translation quality measured per 1,000 words?

Translation quality is measured by tagging every defect against the five MQM categories, assigning each a severity weight, summing the weighted defects, and dividing by the volume of translated source words in thousands. The resulting per-1,000-word score is compared to a published lane threshold. Files above the threshold return to the translator before release, not after.

flowchart TD
 A[Finished target file] --> B[Automated pre-LQA checks]
 B --> C[Human LQA reviewer]
 C --> D{Tag defect: MQM category + severity}
 D --> E[Sum weighted defects]
 E --> F[Divide by source words / 1000]
 F --> G{Score vs lane threshold}
 G -- Below --> H[Release]
 G -- Above --> I[Return to translator]

Back-translation sits alongside the LQA gate as a complementary check on high-risk segments — the Back Translation: How It Works, When to Use It, and Its Limits page covers when it earns its cost.

Which release threshold applies to each production lane?

Builds above the lane threshold return to the translator before release, not after. The full lane breakdown sits on the Localization Management: How Translators and Agencies Run Multi-Locale Programs page.

LaneThreshold / 1,000 wordsDominant dimensions
Certified / regulated< 2Accuracy, terminology, compliance
Legal / compliance< 2Accuracy, terminology
Product / app / support< 3Terminology, functional integrity
Marketing< 5Style, locale conventions

Why must the reviewer be independent of the translator?

An MQM-style score is only auditable when the reviewer scoring the file did not translate the segments being scored. Independence turns the per-1,000-word score into a production control; without it, the number records the translator's confidence in their own work, not the file's quality. The rule is operational, not ceremonial — a batch scored by its own translator does not clear the gate.

How do automated checks and human review combine in a translation quality measurement?

Automated checks catch deterministic defects — missing tags, number mismatches, forbidden terms, segment-length outliers — before a human reviewer opens the file. Human review then scores the remaining accuracy, terminology, style and locale defects that only a bilingual linguist judges.

LayerCatchesMisses
Automated pre-LQATag integrity, number mismatch, forbidden terms, length outliersMistranslation, register, locale fit
Human LQAAccuracy, terminology judgment, style, locale conventionScales linearly with word volume

How it works

What causes translation quality failures in professional workflows?

1

How does terminology drift inflate the per-1,000-word score?

Terminology drift happens when a locked brand term or glossary entry is translated differently in two segments of the same file. Each drift counts as a terminology defect, and three drifts on a 1,000-word file push a product lane score to the release threshold on their own. Drift traces to CAT tool concordance gaps, not translator inattention.

2

Why do locale conventions get missed on files that otherwise read well?

Locale conventions get missed because CAT segmentation hides date, currency and number formats inside tagged source segments that translators pass through without reformatting. The file reads well sentence by sentence but fails locale checks at the whole-file level. Automated pre-LQA number checks catch the arithmetic but not the format convention.

3

When do reviewer edits record preference instead of defects?

Reviewer edits record preference when the reviewer rewrites a segment that was already semantically equivalent to the source, without tagging an MQM category. Preference edits still cost the translator a rework cycle but do not improve the per-1,000-word quality score. The fix is a scorecard rule: no edit ships without a category tag.

Where does translation quality end and localization fit begin?

Translation quality ends at sentence-level transfer of meaning scored by the MQM rubric; localization fit begins at everything built around that transfer — engineering, cultural adaptation, legal boilerplate, layout, and editorial judgment. A file can clear under 2 errors per 1,000 words and still ship as bad localization when the layout truncates, the legal boilerplate is wrong for the target jurisdiction, or the imagery misreads. The Localisation Project: How Translators and Agencies Scope, Run, and Deliver One page scopes the full build.

  • Inside the quality score: accuracy, fluency, terminology, locale convention, style — at sentence level.
  • Outside the quality score: UI truncation, legal boilerplate selection, image swap, SEO adaptation, pricing format, payment method compatibility.

How does translation quality fit into the certified production workflow?

Translation quality is enforced by an LQA gate that scores every finished file at $40–$90 per 1,000 words against an MQM-style scorecard aligned to ISO 17100, sitting between human post-edit and release sign-off. On certified lanes the gate blocks publish until the file clears under 2 errors per 1,000 words. The scorecard mechanics live on the LQA: How Linguistic Quality Assurance Works in Practice page; the end-to-end workflow sits on Translation Processes: How Agencies Run Certified Production.

ISO 17100 sets the process and personnel requirements — translator qualification, independent revision, project management — that the MQM-style rubric then converts into numeric defect-per-1,000-words scores. One standardizes the workflow; the other measures its output.

How do production teams keep improving translation quality over time?

Production teams improve translation quality by closing the feedback loop between LQA defects, glossary updates, style guide revisions, and translator onboarding. Each failed file feeds specific error markup back to the post-editor, so the next batch ships with fewer defects per 1,000 words. A Translation Management System centralizes glossary, style guide, translation memory and LQA scorecard so measurement is reproducible across batches.

How do glossaries and style guides translate into fewer defects?

Glossaries convert terminology defects into automated pre-LQA catches, and style guides convert style defects into reviewable rules. Both move defects from the human reviewer's score to the automated check layer, lowering the per-1,000-word score the LQA gate sees. A glossary entry a CAT tool can enforce is worth ten style-guide paragraphs a reviewer has to remember.

How does the feedback loop between LQA and translators compound over batches?

The LQA reviewer returns annotated files with MQM-tagged defects to the translator who produced them, by segment and severity. Over 4–6 batches, translators internalize the recurring error categories, and the file-level score drops below the lane threshold without a rework cycle. The loop only compounds when the annotations carry category tags — ungrounded "please reword" notes do not teach.

How we treat translation quality

We run every finished file through an LQA gate scored against a weighted MQM-style rubric aligned to ISO 17100, with a reviewer who did not post-edit the segments. The score is the release decision, not an advisory note. On certified lanes a file that scores above 2 errors per 1,000 words returns to the post-editor with tagged annotations before it leaves the production system.

We decline to issue a single universal "quality score" across lanes. A marketing file at 4 errors per 1,000 words is a passing build; a certified file at the same score is a failing one. Flattening that difference into one headline number reads as marketing and misleads the receiving authority.

We refuse to count a reviewer rewrite as a defect when the original segment was already semantically equivalent to the source. Preference edits go to the style-guide backlog, not the translator's rework queue. Counting them as defects inflates the per-1,000-word score and trains translators to write for the reviewer instead of the source.

We do not let the same person translate and score the same segment. The independence rule is not a formality — a score written by the author of the work is a self-report, not a control.

Working translators and agency leads ask a short set of recurring questions about translation quality: what the release threshold is, how automation changes the number, how the score behaves on machine-translated segments, and when the score fails to capture a real defect.

Frequently asked questions

How do I know if a translation is certified?

A certified translation includes a signed statement from the translator or agency attesting that the translation is complete and accurate to the best of their knowledge. This certificate appears on agency letterhead and includes the translator's name, credentials, signature, date, and sometimes a seal or notary stamp. Requirements vary by jurisdiction and use case — USCIS, court systems, and academic institutions each have slightly different expectations — but the signed attestation is the universal minimum. Always verify that the certification matches the requirements of the receiving institution.

What are the ISO standards for translation?

The two primary ISO standards governing translation quality are ISO 17100:2015 and ISO 21998:2020. ISO 17100 defines requirements for translation service providers — covering competence of human resources, revision workflows, and project management — and is the benchmark most agencies seek certification against. ISO 21998 addresses quality management for language services more broadly. For machine translation post-editing specifically, ISO 18587:2017 sets competence and process requirements. Together, these standards form the framework that professional translation agencies and production teams use to design and audit their QA processes.

How to evaluate the quality of a translation?

Translation quality is evaluated by measuring accuracy, fluency, terminology consistency, and fitness for purpose against a defined standard. The most widely used structured method is MQM (Multidimensional Quality Metrics), which categorizes errors by type — accuracy, fluency, style, locale convention — and severity (minor, major, critical). BLEU and other automated metrics can flag gross errors in machine output quickly, but human review remains essential for certified translations. Production teams combine automated QA checks (terminology, number, tag consistency) with a human revision pass that follows ISO 17100's two-step translate-then-revise workflow.

What is QA in translation?

QA in translation refers to the systematic process of checking translated content against defined quality criteria before delivery. It operates at two levels: automated QA (software checks for missing segments, number mismatches, inconsistent terminology, tag errors, and formatting issues) and human QA (a qualified reviser comparing the target text against the source for accuracy and fluency). For certified translation production, QA is not optional — it is a process requirement under ISO 17100, which mandates a revision step performed by a second qualified linguist. Many agencies also add a final proofread or spot-check as a third gate.

Can I translate without being certified?

In most countries, there is no universal legal requirement for translators to hold a specific certification to work professionally. The United States, for example, has no federal licensing requirement — anyone can produce a translation and sign a certification statement. However, certain contexts impose stricter rules: sworn or official translators are legally required in countries like Spain, Germany, and Brazil, and courts or government agencies may specify accredited translators only. The practical risk of translating without recognized credentials is reputational and legal liability when translations are rejected or challenged.

What are the four types of quality assurance?

In translation production, QA is categorized into four types: linguistic QA (accuracy, fluency, terminology), functional QA (does the content work in context — links, UI strings, placeholders), technical QA (file integrity, tags, encoding, formatting), and process QA (workflow compliance — were the right steps followed by qualified people). These map roughly to general QA categories used in software and manufacturing: prevention, detection, correction, and verification. For certified translation teams, process QA — ensuring the right credentials and revision steps were applied — carries particular weight because it underpins the legal validity of the certification.

Related