"The translation feels inconsistent" is a useful warning, but a difficult instruction to act on. Which passages are wrong? How serious are the problems? What should change before publication?
MQM translation quality assessment uses a structured framework to identify, classify, and evaluate translation errors against agreed requirements. MQM stands for Multidimensional Quality Metrics. It can support evaluation of both human and machine-generated translation. (MQM Council: The MQM framework)
The value is not simply a score. It is a defensible explanation of what needs attention and why.
Decide what the assessment needs to tell you
Before selecting categories, define the decision. Are you checking release readiness, comparing suppliers, assessing AI-generated Dutch, or investigating recurring complaints?
Those questions require different evaluation designs. A broad supplier sample and a review of ten high-risk interface messages should not automatically be treated as equivalent evidence.
My recommendation is to document the content type, intended audience, source and target languages, applicable reference materials, and the decision the report must support. Agree whether the assignment includes corrections or evaluation only.
Without that agreement, a detailed error report can still answer the wrong question.
Separate error category from severity
An error category describes what is wrong. Severity describes its effect in context.
MQM offers a hierarchy from which an implementation can select appropriate error types. Relevant areas include accuracy, terminology, fluency, and style. The full framework is not a requirement to use every available category on every assignment. (MQM Council: The full error typology)
For example, changing "you can cancel" into "you cannot cancel" is an accuracy issue. Its severity depends on the role of that instruction and the consequences for the reader.
A minor error has limited impact. A major error seriously affects meaning, reliability, or usability. A critical error can make content unfit for purpose or create a risk of serious harm. Some implementations make a critical error an automatic failure; the policy must be explicit. (MQM Council: Scoring models)
A terminology preference is not automatically an error
Suppose a reviewer prefers one Dutch synonym, while the existing translation is accurate, natural, and compatible with the brief. That preference needs to be distinguished from an actual defect.
Conversely, a term that contradicts an approved glossary can be a genuine problem, particularly when users need to match an instruction to an interface label.
A useful review explains its reasoning rather than treating every rewrite as evidence of poor quality.
Understand what an MQM score actually means
Ask how the score is calculated before comparing numbers. Severity weights, category weights, normalization, and thresholds can differ between implementations. The MQM Council describes configurable scoring models, not one compulsory pass mark for all content. (MQM Council: Scoring models)
Here is a deliberately simple illustrative MQM-based calculation, not a universal MQM formula or a client result. Assume 1,000 evaluated source words, a weight of 1 for minor errors, 5 for major errors, and no additional category weights.
| Finding | Count | Weight | Penalty points |
|---|---|---|---|
| Minor errors | 6 | 1 | 6 |
| Major errors | 2 | 5 | 10 |
| Critical errors | 0 | Separate release rule | 0 |
| Total | 16 |
The normalized result is 16 penalty points per 1,000 source words. Lower is better in this example. It does not establish that the translation is "98.4% accurate."
Nor does it establish a pass or fail until the acceptance threshold has been agreed. A penalty rate and an accuracy percentage are not interchangeable descriptions.
Make the sample and reviewer process visible
A report should state how the sample was chosen and what it excludes. If the sample covers only straightforward product descriptions, avoid drawing conclusions about complex support content.
For a diagnostic review, I would include a mix of routine material, recurring terminology, ambiguous source passages, and high-impact content. For a supplier comparison, I would keep the evaluated material and conditions comparable. High-risk content may need full review rather than sampling.
Reviewer calibration also matters. Evaluators should practice on shared examples, discuss disagreements, and apply the same criteria. MQM’s reliability guidance highlights the importance of training for consistent human evaluation. (MQM Council: Evaluation reliability)
Decide how repeated errors are counted as well. Penalizing every occurrence and penalizing an issue only once produce different results; MQM guidance says that choice should be made and communicated in advance. (MQM Council: Repeated errors)
Ask for a report that supports correction
My recommended deliverable includes the source passage, target passage, location, error category, severity, explanation, and proposed correction where correction is in scope.
The summary should distinguish isolated problems from recurring patterns. If several errors arise from one unclear source term, the next step may be a terminology decision rather than a longer list of individual edits.
Likewise, repeated interface problems may call for more context or a change to resource handling. Keep linguistic findings separate from implementation defects, while showing how they affect the same user experience.
The final question should be: what changes, who owns those changes, and what needs checking again?
Is MQM the same as proofreading?
Not necessarily. Proofreading or editing aims to improve the text within an agreed scope. An MQM assessment evaluates quality using a defined framework and can be commissioned without rewriting every sentence.
Ask for both explicitly when both are needed. Also avoid treating a sampled assessment as a guarantee that unreviewed content contains no errors.
Need a clearer view of your Dutch translation quality?
Send a representative source-and-target sample, the content type, and the decision you need to make. Broekman Language Solutions can discuss an MQM assessment with clear criteria, transparent reporting, and an agreed correction scope.
Related reading: AI translation for Dutch content: when human review matters; Dutch software localization: a UI and UX checklist for SaaS teams.