Syntax-aware Data Augmentation for Neural Machine Translation

Data augmentation is an effective method for the performance enhancement of neural machine translation (NMT) by generating additional bilingual data. In this paper, we propose a novel data strategy translation. Unlike existing methods that simply modify words with same probability across different sentences, introduce sentence-specific approach word selection based on syntactic roles in sentence. Our motivation to consider linguistics-motivated obtain more ingenious language generation rather than relying computation-motivated approaches only. We argue high-quality aligned crucial NMT, and only insufficient provide good enough extra leverages dependency parse trees input sentences determine each sentence using three functions calculate probabilities depths. Besides, our also revises considering length. evaluate multiple tasks. The experimental results demonstrate proposed does effectively boost sentence-independent significant improvement Furthermore, ablation study shows select fewer essential preserves structure.