Sampling and Filtering of Neural Machine Translation Distillation Data
This work addresses the incremental improvement of model distillation efficiency for machine translation practitioners.
The paper tackled the problem of improving neural machine translation distillation by exploring various importance sampling methods, such as pruning, hypothesis upsampling, and deduplication, and found that careful upsampling combined with original data leads to better performance compared to using only original or synthesized data.
In most of neural machine translation distillation or stealing scenarios, the goal is to preserve the performance of the target model (teacher). The highest-scoring hypothesis of the teacher model is commonly used to train a new model (student). If reference translations are also available, then better hypotheses (with respect to the references) can be upsampled and poor hypotheses either removed or undersampled. This paper explores the importance sampling method landscape (pruning, hypothesis upsampling and undersampling, deduplication and their combination) with English to Czech and English to German MT models using standard MT evaluation metrics. We show that careful upsampling and combination with the original data leads to better performance when compared to training only on the original or synthesized data or their direct combination.