You might have heard the term ‘quality estimation’ (QE) before - it’s something that’s being used increasingly frequently across the translation industry. At Unbabel, we’re extremely proud of our QE models, and we use them across our translations to inform us of the quality of a translation after it’s been through our engines, but before the job has been corrected by a member of our community.
QE is complex, but essentially it has been developed to identify potential errors in translation and therefore inform us what needs correcting, and, by extension, which segments need to be prioritised for translation. It does this by learning from the work our community does, but also the work of our community pro members, who look at sample translations and examine their quality, annotating them with their assessment, which is then fed back into the translation engine to improve its output. It also helps our customers understand the levels of translation quality they’re currently getting from us, so it’s a really important tool and one of the things that helps keep Unbabel on the cutting edge of translation technology.
For the most part, QE goes on behind the scenes, so it’s not something you’ll necessarily have been aware of. However, we are now starting to use it to change what you see when you’re working on some reviews. The way it works is that when our QE systems detect that a segment is of good quality, it’ll be marked as a medium confidence segment on a review, which will then allow reviewers to assign their time more efficiently, as explained in this article.
If you’re interested in learning more about QE, you can check out this article written for our clients which talks about our QE systems.