Improving Low-Resource Machine Translation Using Syntactic Dependencies
DeLin Deng, Liqing Wang · 2024
Transformer-based neural network machine translation has recently demonstrated remarkable performance across various tasks. However, for low-resource languages, achieving satisfactory results poses a significant challenge due to limited data availability. Existing approaches targeting low-resource machine translation often rely on augmenting training data with additional resources. To enhance the performance of low-resource machine translation models, we propose leveraging syntactic structures extracted from the source language training data. By incorporating these structures into the translation model training process and employing a pre-training strategy, we aim to facilitate more accurate translations by enabling the model to better understand syntactic nuances. Experimental results on Chinese-Thai translation demonstrate a notable improvement of 1.32 BLEU over the baseline model. Furthermore, by integrating the proposed method into Transformer optimization, we achieve an additional enhancement of 0.85 BLEU.