Review of Hierarchical Transfer Learning Architecture in Low-Resource Machine Translation

Abstract

Machine translation is a field of study that has attracted significant attention in recent years. The success of a model built on a language pair depends mainly on the number of parallel sentences between languages. Unlike high-resource languages, low-resource languages suffer from (lack of)/(limited) parallel data. In this study, an examination was made of the hierarchical transfer learning architecture used in the field of low-resource neural machine translation. The study was conducted on the Turkish-English language pair, German-English was used as a high-resource language pair, and Kazakh-English was used as a language pair similar to the Turkish-English language pair. When the results obtained were examined, it was seen that the language pair that had the main impact on translation success in the hierarchical transfer learning approach was the language pair that showed similar characteristics to the low-resource language pair.