Translation Quality Assurance: Beyond Basic Proofreading

Translation quality assurance (QA) is the set of checks that makes a translation fit for its purpose: automated checks for tags, numbers and terminology, a review by a second linguist and measurement against agreed error categories and thresholds. "Quality" is easy to promise and hard to define, so the useful question is what it means for a given project and how it will be measured.

In our experience, translation quality isn't a single standard: it's a framework that must be calibrated to each project's purpose, audience and risk profile. Here's how we think about it at El Turco.

The Quality Spectrum

Not all translations need the same quality level, and pretending otherwise wastes resources or creates false expectations. A common model uses three quality tiers:

Tier 1: Premium (Human Translation + Senior Review)

• Use case: Marketing, brand content, legal documents, life sciences, creative transcreation
• Error tolerance: Zero critical errors; minimal minor issues
• Process: Translation by senior linguist → review by second senior linguist → PM quality check → client review

Tier 2: Professional (Human Translation + Light Review)

• Use case: Product descriptions, support documentation, training materials, UI strings
• Error tolerance: Zero critical errors; minor issues acceptable within defined thresholds
• Process: Translation by qualified linguist → review by second linguist → automated QA

Tier 3: Rapid (MT + Professional Post-Editing)

• Use case: High-volume content, internal communications, knowledge base articles, user-generated content
• Error tolerance: Accuracy must be maintained; style and fluency may be functional rather than polished
• Process: MT output → professional post-editing → automated QA

The key insight: quality tiers aren't about caring more or less. They're about matching the process to the content's purpose and risk profile. Turnaround is agreed at scoping, with a 24-hour option for small urgent updates.

The MQM Framework

The Multidimensional Quality Metrics (MQM) framework, developed at DFKI in the EU-funded QTLaunchPad project and later harmonized with the TAUS DQF error typology in the QT21 project, is the industry's most comprehensive approach to translation quality evaluation. It categorizes errors along multiple dimensions:

Accuracy

• Mistranslation: Target text conveys a different meaning than source
• Omission: Source information missing from target
• Addition: Target contains information not in source
• Untranslated: Source text left untranslated in target

Linguistic conventions (formerly Fluency)

• Grammar: Morphological, syntactic or orthographic errors
• Punctuation: Incorrect or missing punctuation
• Spelling: Misspellings in target text

Terminology

• Inconsistent terminology: Same concept translated differently within a project
• Wrong terminology: Incorrect term used per approved glossary

Style

• Register: Inappropriate formality level or style
• Style guide violations: Deviations from documented style preferences
• Awkward phrasing: Grammatically correct but unnatural expressions

Design and markup

• Tag errors: Missing or misplaced formatting tags
• Truncation: Text cut off due to character limits
• Layout issues: Formatting inconsistencies

Each error is further classified by severity:

• Critical: Impacts meaning, creates legal/safety risk or is offensive
• Major: Noticeable to end users and affects quality perception
• Minor: Detectable but doesn't significantly impact user experience

Automated QA: What It Catches and What It Misses

Rule-based QA tools (Xbench, Verifika, QA Distiller, built-in TMS QA) are essential but not sufficient. Here's an honest assessment:

What Rule-Based QA Does Well

• Tag validation: Catches missing, extra or misplaced formatting tags with near-perfect reliability
• Number checking: Verifies that numbers in source appear in target
• Terminology verification: Cross-references translations against approved glossaries
• Consistency checking: Flags when identical source segments have different translations
• Spelling: Standard spell-check functionality
• Length checks: Flags segments that exceed defined character limits

What Rule-Based QA Cannot Do

• Semantic accuracy: A rule-based check cannot determine if "bank" was correctly translated as a financial institution or a river bank
• Cultural appropriateness: No rule-based check can evaluate whether a Turkish marketing translation resonates culturally
• Register evaluation: Detecting whether the formality level is appropriate for the audience
• Naturalness: Determining if a translation, while technically correct, sounds unnatural to a native speaker
• Contextual correctness: Evaluating whether a UI string makes sense in its actual on-screen position

This gap between automated and human QA is where quality is won or lost. AI quality estimation can now flag likely meaning and style errors and label them with MQM categories, but its findings still need a linguist to confirm them.

Building Effective QA Workflows

The Three-Layer Approach

A practical QA model has three layers:

Layer 1: Translator self-review (immediate)

Before delivering any segment, translators run automated QA checks and perform a self-review against the project style guide. This catches the majority of mechanical errors.

Layer 2: Linguistic review (same day)

A second linguist reviews for accuracy, fluency and style. The reviewer has access to the source text, reference materials and the project style guide. They use a standardized review form that maps to MQM categories.

Layer 3: Spot-check and metrics (delivery day)

A quality lead spot-checks a random sample of delivered segments. Results are scored against project-specific quality thresholds and recorded for trend analysis.

Quality Feedback Loops

QA isn't just about finding errors: it's about preventing them. We maintain:

• Translator scorecards: individual quality metrics tracked over time
• Error pattern analysis: identifying recurring error types by linguist, language pair or domain
• Style guide evolution: updating guidelines based on common issues
• Calibration sessions: regular meetings where linguists and reviewers align on quality expectations

Quality for Turkish and Arabic: Language-Specific Considerations

Turkish QA Specifics

• Vowel harmony verification: Turkish suffixes must follow vowel harmony rules. Errors here immediately signal non-native quality.
• Consonant mutation checks: Certain consonant changes at morphological boundaries (e.g., kitap → kitabı) are frequent error sources.
• Formality consistency: Verify that sen/siz usage is consistent throughout the deliverable.

Arabic QA Specifics

• Dialect consistency: Ensure the target dialect is maintained throughout. Mixing MSA with dialectal forms is a common quality issue.
• Hamza placement: Correct hamza (ء) placement on alif carriers is a frequent error that automated tools often miss.
• Tashkeel (diacritics): When required (e.g., educational or religious content), verify diacritical mark accuracy.

Measuring Quality Over Time

Three metrics are worth tracking across projects:

1. Error rate per 1,000 words: total errors weighted by severity
2. Critical error rate: tracked separately because a single critical error can invalidate an otherwise excellent deliverable
3. Client satisfaction score: qualitative feedback that captures dimensions internal metrics might miss

Over time, these metrics reveal patterns: which linguists excel in which domains, which content types are error-prone and where processes need strengthening.

Next steps

For how these checks apply to post-editing, see MTPE vs human translation for Turkish and Arabic. Language-specific notes are on the Turkish and Arabic pages, and the service itself is described under translation and localization. LSP vendor managers can start with a free test translation of up to 300 words. To compare quality on your own content, request a free LQA of up to 1,000 words already live in one of our languages.