QUARTZ Unbabel -ELISE Project Open Call 1
from Finnish Center for Artificial Intelligence FCAI Hosted by DASER
Within the period of the project, we developed a quality-aware machine translation (MT) approach that optimizes the MT system to produce fewer severe translation errors as evaluated by professional translators. Our approach is effectively reducing on average up to 40% translation errors made by MT systems that might compromise the understanding of a translation or that can lead to health, safety, legal or financial implications to the end user or reader of the translation. Furthermore, we made steps also towards a faster automatic machine-learning-based MT evaluation metric based on COMET (37% faster than previous models) and a model that performs estimation of critical translation errors better (improvement of more than 70% in precision and recall).
Embed this episode
No transcript yet.