G-Transformer for Document-level Machine Translation
This addresses the problem of unstable training in document-level machine translation for NLP researchers, offering an incremental improvement over existing methods.
The paper tackled the failure of Transformer models in document-level machine translation by identifying that the issue stems from local minima during training due to complex target-to-source attention, and proposed G-Transformer with locality inductive bias, achieving new state-of-the-art BLEU scores on three benchmark datasets.
Document-level MT models are still far from satisfactory. Existing work extend translation unit from single sentence to multiple sentences. However, study shows that when we further enlarge the translation unit to a whole document, supervised training of Transformer can fail. In this paper, we find such failure is not caused by overfitting, but by sticking around local minima during training. Our analysis shows that the increased complexity of target-to-source attention is a reason for the failure. As a solution, we propose G-Transformer, introducing locality assumption as an inductive bias into Transformer, reducing the hypothesis space of the attention from target to source. Experiments show that G-Transformer converges faster and more stably than Transformer, achieving new state-of-the-art BLEU scores for both non-pretraining and pre-training settings on three benchmark datasets.