Parallel Corpora Preparation for bi-Directional Amharic-Kistanigna Machine Translation

Abstract

This research focused on creating a parallel corpus for the Amharic and Kistanigna languages and conducting Machine Translation (MT) tests on it. The goal was to increase Kistanigna language content online, address its endangered status, and facilitate information sharing between the two languages. To achieve this, a parallel corpus was developed, and machine translation models were tested, including LSTM, BiLSTM, LSTM with attention, CNN with attention, and Transformer models. Experiments were conducted with both word and morpheme-based translation units, using the morfessor tool for morpheme segmentation. The best results were achieved with morpheme-based bidirectional machine translation using a Transformer, with BLEU scores of 21.31 for Amharic-Kistanigna and 22.40 for Kistanigna-Amharic translations. The resulting corpus contains 9, 225 parallel sentences and will be made available to the community. This is the first parallel corpus for these languages, and more such corpora are needed for further research.