MQM
MQM, Multidimensional Quality Metrics, is a standardised framework for assessing translation quality by human error annotation. Reviewers mark each error against a shared typology, accuracy, fluency, terminology, style and more, and by severity, and the marked errors produce a score. It is the industry reference for rigorous, human-judged quality.
How it works
A trained reviewer reads the translation against the source and tags each problem with a category from the MQM typology and a severity, minor, major or critical. Weights per severity turn the tags into a penalty, and the penalty into a score.
Because the categories are shared and defined, MQM makes quality diagnosable rather than just a number: you can see whether a translation fails on terminology, on accuracy, or on style, and act on it.
How SourceTarget uses it
SourceTarget's language quality assurance uses an MQM-style error typology so that human review produces structured, comparable results, not just a pass or fail. That structure feeds back into which engines and workflows are chosen for a content type.
MQM compared with BLEU score
| MQM | BLEU score | |
|---|---|---|
| Judged by | Trained human annotators | Automatic word overlap |
| Output | Categorised errors and a diagnostic score | A single overlap number |
| Strength | Explains what is wrong | Instant and repeatable |
| Cost | Slower, human effort | Free and immediate |