Improving Edit Distance: MT Quality Estimation and the Rise of the “Super Segment”

Track: Multilingual AI | TA1 |   Everyone |
Wednesday, June 7, 2023, 9:00am – 9:30am
Held in: Live 4-5
Presenter:
Jay Marciano - Lengoo
Host: Johan Sporre

A virtuous cycle of retraining neural machine translation models with post-edited content can result in such high translation quality that up to half of new segments require no edits. Identifying which half of the segments don’t require correction — we call these “super segments” — is the tricky part and a potential game changer! The solution: trust and transparent communication between the language service provider and customer, and an automated quality estimation system that is also trained on the customer’s data. Depending on the customer’s requirements, super segments can be passed directly to the final quality assurance phase or even delivered without human review, offering substantial time and cost savings.

Takeaways: Attendees will learn that reliable and responsible application of AI can benefit all stakeholders in the translation process.