1. Optimizing for User Benefit
Designing User-Centric MachineTranslation Systems
Chris Hokamp
Dublin City University
chokamp@computing.dcu.ie
@Mystical_Wiz
5. Post-editing time, Post-editing Effort
Interactive Prediction Quality
How much will our predictions help?
How can we measure how much they help?
Measuring Translation Utility
6. Post-process Machine Translation
Learn from translator feedback using Translation
Memories or Online/ Incremental Retraining
Use Quality Estimation to guide translators to
“problem spots”
State of the Art
7. Task-Based Optimization
Formalize the purpose of
Machine Translation for the task
Take advantage of the user’s
context to improve predictions
Explicitly train Machine
Translation systems to maximize
user benefit
9. Optimizing Interactive Prediction
Neural Machine Translation
for Interactive Translation
tasks
Neural Machine Translation
for Post-editing tasks
10. Task based optimization is not limited to machine
translation systems
This model can be applied to any scenario where a
user interacts with an agent to achieve a shared
goal
The Vision