firefox-translations-training
firefox-translations-training copied to clipboard
Dig deeper in sentences we translate badly to try and identify common failure themes
For example, by using tools such as compare-mt, or by comparing translations with a “good” known one and sorting by BLEU/COMET.
Another interesting tool in this space: https://github.com/ondrejklejch/MT-ComparEval.
We can also use the new eXplainable COMET models for this.
We could also use the same tools to compare teacher vs student and see if there are common cases where the student fails and the teacher succeeds.