Auto MT Quality Prediction Solution and Best Practices


Track: Technical | T5 |
Wednesday, January 27, 2021, 1:30pm – 2:00pm
Held in: Stream 2
Presenters:
York Jin - VMware 
Martin Xiao - VMware

In this presentation we will share the solution and best practices of automatically evaluating machine translation (MT) output quality and deploying it into production. This includes how to train the MT quality prediction models by using a natural language process algorithm and how to deploy these models into production. We will also include two use cases in production: golden MT detection and smart MT selection. For golden MT detection, an index (patent applied) has been invented as a benchmark to identify golden MT with high accuracy which then applies it directly without human post-editing. For smart MT selection with multiple MT output, it uses prediction results to evaluate the quality then selects the best one.

Takeaways: Attendees will learn how to train the MT quality machine learning prediction models; learn about an MT output quality index that has been invented to identify golden MT output; and hear two use cases: golden MT detection and smart MT selection and their benefits.