What is machine translation quality estimation (MTQE) and how to use it to quote post-editing

Machine translation quality estimation (MTQE) predicts how good a translation is without comparing it to a reference translation. It looks only at the source and the translation, and returns a score.

How it differs from BLEU or chrF

Metrics such as BLEU or chrF compare the translation with one made by a person. They are useful to evaluate engines in the lab, but not day to day: when a new project arrives, that reference translation does not exist. MTQE works exactly at that moment.

What it does well and what it does not

• It is very good at spotting clearly good and clearly bad sentences.
• In the middle zone it makes more mistakes, which is why it does not replace review: it prioritises it.
• That is why, in TranslaQE, the model's score is combined with fixed checks (numbers, tags, glossary, length and untranslated text) that catch errors a model sometimes misses.

How to use it to quote

1. Run the project through MTQE and get the word split by band: trusted, review and redo.
2. Apply a rate to each band, just as you do with the match bands of your translation memory.
3. Review a sample of the "trusted" sentences to check that the threshold works for that kind of text and language pair.

Over time, every human review helps calibrate the thresholds to your customer and language pair. That is the part TranslaQE does automatically.

Want to try it with your texts? Join the waitlist.