Finetuning transformer-based MT using syntactic guides

Incorporating syntactic information into machine translation (MT) systems can enhance their performance by providing richer linguistic context. This paper investigates the impact of integrating dependency informations into the fine-tuning process of Transformer-based MT models. We evaluate two translation scenarios: multilingual source-to-English and English-to-multilingual targets. Our experiments demonstrate that leveraging dependency tags in the target language improves translation quality significantly, achieving up to a 50% increase in validation scores compared to models without such enhancements. Additionally, dependency informations consistently boost performance across diverse language pairs, except in cases where datasets lacking such annotations. These results underscore the effectiveness of dependency tags integration to enhance translation accuracy, minimize reliance on large training datasets, and improve the overall efficiency of modern MT systems.