Abstract

In this work, we investigate methods for the challenging task of translatingbetween low-resource language pairs that exhibit some level of similarity. Inparticular, we consider the utility of transfer learning for translatingbetween several Indo-European low-resource languages from the Germanic andRomance language families. In particular, we build two main classes oftransfer-based systems to study how relatedness can benefit the translationperformance. The primary system fine-tunes a model pre-trained on a relatedlanguage pair and the contrastive system fine-tunes one pre-trained on anunrelated language pair. Our experiments show that although relatedness is notnecessary for transfer learning to work, it does benefit model performance.

 

Quick Read (beta)