Developing a Neural Machine Translation System for Irish (2019)

In this paper we develop a neural machine translation (NMT) system for translating from English into Irish and vice versa. We evaluate the performance of NMT on the resource-poor English-Irish (EN–GA) language pair, show that we can achieve good translation quality in comparison to previously reported systems, and outperform Google TranslateTM with several BLEU points on a domain-specific test set related to the legal domain. We show that back-translation of monolingual data closely related to the domain of the test set can further increase the model’s performance. Finally, we present a lightweight method for filtering synthetic sentence pairs obtained via back-translation using a tool for misalignment detection. We show that our approach results in a slightly higher BLEU score while requiring less training data.