Quality estimation
Quality estimation (QE) is predicting how good a translation is without access to a human reference translation. A model scores the output from the source and the translation alone. Because it needs no reference, it can judge live machine translation in production, where references never exist, and flag segments that need review.
How it works
A QE model is trained on translations that humans have scored, learning to predict that score from the source and output alone. In use it takes a fresh translation with no reference and estimates its quality, or the risk that it is wrong.
That reference-free ability is the point: real production translation has no gold answer to compare against, so QE is how you triage machine output at scale before a human sees it.
How SourceTarget uses it
SourceTarget uses quality estimation to score and route live machine output where no reference exists, which is almost always. Low estimates push a segment up the review queue, complementing the engine's own confidence score.
Quality estimation compared with BLEU score
| Quality estimation | BLEU score | |
|---|---|---|
| Needs a reference | No | Yes |
| Works in production | Yes, on live output | Only where a reference exists |
| Used for | Triaging what to review | Benchmarking systems offline |