ABSTRACT
This paper introduces a curriculum-based fine-tuning approach for Neural Machine Translation (NMT) that leverages Quality Estimation (QE) feedback to dynamically select and order training examples by domain relevance and translation difficulty, enabling effective domain specialization while preserving general translation capabilities.
PAPER · PDF
Loading PDF...
Key findings
Proposed a novel curriculum-based fine-tuning approach for domain-adaptive NMT.
Achieved 36.6% improvement on target medical domain while maintaining performance on general-domain text.
Strong correlation (R2=0.72) between QE-estimated difficulty and final translation quality.
Limitations & open questions
The approach is evaluated on simulated datasets, and its effectiveness on real-world data remains to be tested.