Quality score
A quality score is a composite rating of a finished translation, usually from 0 to 100, that combines several automatic metrics. On SourceTarget the weighting is set by content type, terminology counts for more on legal documents, fluency for more on campaigns, and the score arrives with an explanation and the segments a human should check.
How it works
A quality score combines several automatic metrics into one number, then weights them for the kind of content being judged. Word-overlap and character metrics catch surface accuracy, neural metrics estimate meaning, and terminology checks confirm the right terms were used. The weighting decides which of these matters most for this job.
Because it is a composite, it is more robust than any single metric, and it can be explained: the score comes with the components behind it and the segments dragging it down, so a reviewer knows where to look.
How SourceTarget uses it
On SourceTarget every delivered job carries a quality score, weighted by content type: terminology counts for more on legal and medical work, fluency for more on campaigns. The score arrives with an explanation and a list of segments to check, and it drives how work is routed for review.
It is a decision aid, not a verdict. It tells a human where to concentrate; the human decides what ships.
Quality score compared with Confidence score
| Quality score | Confidence score | |
|---|---|---|
| Timing | After the translation is produced | At generation, before any check |
| Judges | The finished translation | The engine's certainty in itself |
| Built from | Several weighted metrics | The engine's own probabilities |
| Used to | Decide what is fit to publish | Triage what to review first |