A Bi-LSTM Neural Machine Translation for Tigrigna-Kunama Languages

Abstract

This paper describes a novel bidirectional neural machine translation system between the Tigrigna and Kunama languages using a Bidirectional Long Short-Term Memory network. It seeks to address severe deficiencies in the digital language resource and the absence of existing MT systems for this very low-resource language pair. This study, therefore, attempts to design an MT system trained on a thoroughly prepared parallel corpus. This corpus consists of 4,712 Tigrigna sentences with their Kunama translations, all manually created by language experts. The choice of the Bi-LSTM architecture is motivated by its ability to capture long-range dependencies and contextual subtleties, which are important in producing high-quality translations. The model showed encouraging results, achieving BLEU scores of 88% for Tigrigna-to-Kunama and 89% for Kunama to Tigrigna translations. It then elaborates on the system architecture, pre-processing of data, training methodology, and BLEU score for performance evaluation in detail. The present study helps to fill the important NLP gap in the underrepresented group of languages and emphasizes the role of deep learning techniques in overcoming linguistic barriers for low-resource languages.