Bi-directional Neural machine Translation

The main objective of this research is to develop a bi-directional machine translation model for Amharic and Kistanigna languages. It is critical to develop machine translation between Amharic and Kistanigna to share information between the language, to increase the content of the language on the web, to address the issues of the endangered Kistanigna language. The study explores five models, including LSTM, Bi-LSTM, LSTM with attention, CNN with attention,and Transformer encoder- decoder model. The experiments are conducted on 9,225 parallel sentences with word-based units, considering factors such as training time, memory usage, and BLEU score to find the optimal model. The research suggests bidirectional machine translation using Transformer, which achieves BLEU scores of 7.73 and 4.43 for Amharic-Kistanigna and Kistanigna- Amharic translation, respectively. However, due to a lack of adequate datasets, the study's significant limitation is the inability to conduct extensive experiments. As a result, we recommend preparing parallel corpora to facilitate similar studies in the future.