Quality assurance in translation has a fundamental problem: it requires humans to read everything.

Linguistic Quality Assessment (LQA) evaluates translation quality through systematic error detection and categorization. A reviewer reads each segment, identifies problems, classifies them by type and severity, and scores the overall quality. This produces valuable data about translation performance.

It’s also exhausting, time-consuming, and doesn’t scale.

The review fatigue problem

LQA reviewers face cognitive challenges that undermine quality:

Attention decay. Reviewers catch more errors in their first hour than their fourth. Studies show error detection rates dropping 20-40% over extended review sessions.

Consistency drift. What counts as a “minor” vs. “major” error varies depending on reviewer fatigue, time pressure, and accumulating frustration.

Systematic blind spots. Individual reviewers miss certain error types consistently. Their particular blind spots become QA gaps.

Throughput pressure. When deadlines loom, review depth suffers. “Just check it’s not terrible” replaces systematic evaluation.

The result: QA quality varies based on when review happens, who does it, and how rushed they are. This variability undermines the purpose of quality assessment.

Rule-based automation

The first layer of automated QA: rules-based checking. These catch errors that follow predictable patterns:

Tag validation. Do opening tags have matching closing tags? Are tags in the same order? Are required tags present?

Terminology consistency. Does the translation use approved terms from the glossary? Are prohibited terms avoided?

Number matching. Do numbers in the translation match numbers in the source? Are units converted correctly?

Punctuation checking. Does sentence-ending punctuation match? Are quotation marks and brackets balanced?

Length validation. Is the translation within acceptable length limits for the context?

Consistency rules. Are repeated segments translated consistently throughout?

These checks run instantly across entire projects. They catch a substantial fraction of mechanical errors with zero reviewer fatigue.

AI-powered quality evaluation

Rule-based checks have limits. They can’t evaluate whether a translation:

• Accurately conveys the source meaning
• Uses appropriate register for the audience
• Flows naturally in the target language
• Makes sense in context

AI-powered evaluation addresses these by using LLMs to assess translation quality:

Semantic accuracy. Does the translation mean the same thing as the source? Are there omissions, additions, or distortions?

Fluency evaluation. Does the translation read naturally? Would a native speaker phrase it this way?

Register assessment. Is the formality level appropriate for the content type?

Contextual coherence. Does the segment fit logically with surrounding content?

The LLM provides both scores and explanations, enabling targeted review rather than blanket examination.

Dual-model evaluation

Single-model evaluation has bias risks. The same tendencies that affect translation can affect evaluation.

Dual-model evaluation uses two different LLMs as evaluators:

1. Primary model evaluates each segment
2. Secondary model provides a second opinion
3. Agreement increases confidence; disagreement flags for human review

When multiple frontier models both rate a segment as high quality, confidence is high. When they disagree significantly, human judgment resolves the ambiguity.

Error categorization

Useful QA produces structured data about what’s wrong, not just that something is wrong.

Standard error categories:

Accuracy errors:

• Mistranslation (wrong meaning)
• Omission (missing content)
• Addition (extra content not in source)
• Untranslated (source left in target)

Fluency errors:

• Grammar
• Spelling
• Punctuation
• Unnatural phrasing

Terminology errors:

• Wrong term used
• Inconsistent terminology
• Non-standard translation

Style errors:

• Register mismatch
• Style guide violations
• Voice inconsistency

Technical errors:

• Tag problems
• Formatting issues
• Length violations

Each detected error is categorized, enabling analysis of where quality problems concentrate.

Severity scoring

Not all errors are equal. A mistranslated safety warning is critical. A minor punctuation variation is trivial.

Severity levels:

Critical. Affects meaning, could cause harm, must be fixed. Examples: safety information wrong, legal terms mistranslated.

Major. Clearly incorrect, affects quality perception, should be fixed. Examples: obvious grammar errors, terminology violations.

Minor. Suboptimal but acceptable, fix if time allows. Examples: style variations, minor fluency issues.

Preferential. Matter of opinion, not objectively wrong. Note for awareness, don’t require fixes.

Severity weighting enables prioritization. Review critical and major issues; process minors in batch; accept preferential variations.

Aggregate quality metrics

Beyond segment-level evaluation, aggregate metrics show project-wide quality:

Error density. Errors per 1,000 words, by category and severity.

Quality score. Weighted calculation based on error counts and severity.

Category distribution. Which error types appear most frequently?

Trend analysis. Is quality improving or declining over time? Across vendors? By content type?

These metrics enable quality management at scale—understanding not just individual translation quality but systematic patterns.

The reviewer focus shift

Automated LQA doesn’t eliminate human review. It transforms it.

Without automation: Reviewers read everything, find whatever they find, time pressure limits thoroughness.

With automation: Reviewers receive pre-screened content with issues flagged, can focus on segments actually needing attention, see systematic data about what to look for.

The shift is from “find all the errors” (impossible at scale) to “validate flagged issues and spot-check automated passes” (achievable).

Implementation requirements

Effective automated LQA requires:

Configurable rules. Different projects need different checks. A marketing project cares about different things than a technical documentation project.

Quality thresholds. What scores require human review? What scores can pass automatically?

Workflow integration. QA results need to route to appropriate people with actionable information.

Reporting. Aggregate data for quality management decisions.

The infrastructure investment enables QA at volumes that would be impossible manually.

The quality economics

Manual LQA costs: reviewer time × volume = linear cost scaling.

Automated LQA costs: setup + API costs (marginal) = sublinear cost scaling.

The crossover point depends on volume. For small volumes, manual review may be simpler. For significant ongoing translation, automation produces better quality at lower cost.

Language Ops provides rule-based QA, dual-model LLM evaluation, and configurable quality scoring with detailed error categorization. Run QA on sample content to see the evaluation depth.