Confidence score
A confidence score is a machine translation engine's own estimate, made as it generates, of how reliable its output is likely to be. Unlike a quality score, which evaluates a finished translation, a confidence score is available immediately and is used to decide whether a segment needs human review before anyone has assessed it.
How it works
As the engine generates each segment, it also produces a measure of how probable its own output was. Low probability, roughly, means the engine was unsure, often because the source was ambiguous, rare, or unlike its training data. That measure, normalised, is the confidence score.
Crucially it is available before any human or metric has judged the result, so it can be used to triage: send low-confidence segments to review first. It says how sure the engine is, not whether the engine is right.
How SourceTarget uses it
SourceTarget uses confidence scores to route work. Low-confidence segments are prioritised for human attention, so effort concentrates where the engine is most likely to have slipped.
Because confidence is the engine's opinion of itself, it is paired with the after-the-fact quality score rather than trusted alone.
Confidence score compared with Quality score
| Confidence score | Quality score | |
|---|---|---|
| When it is available | At generation, before review | After the translation is produced |
| What it reflects | The engine's certainty in its own output | An assessment of the finished translation |
| Main use | Triaging what to review first | Judging whether output is fit to publish |
| Blind spot | Can be confident and wrong | Depends on the metrics behind it |