| Abstract: | A language can be described by its rules or its symbols. Making computers understand sentences or words written in human languages is the goal of natural language processing (NLP). Machine translation (MT) is area of NLP where computers are used to translate one natural language into another. One of the languages that needs such translation systems is Tigrigna. Tigrinya is a Semitic language spoken in northern Ethiopia in the Tigray Region, as well as in Eritrea. Previously, some studies were conducted on machine translation of Tigrigna and English languages. However most of the studies were only one directional which is English to Tigrigna languages. Some studies that proposed bidirectional Tigrigna-English machine translation are also domain specific. In this study, the researcher developed a bidirectional Tigrigna-English machine translation model using different machine translation approaches. In the study we collected, 31000 Tigrigna-English parallel corpus from different sources and by translating English text to Tigrigna. We then preprocessed the dataset through cleaning, normalizing and tokenization stages. Using our dataset, we have experimented different machine translation approaches. We have experimented approaches of encoder decoder model and attention based models using LSTM, Bi-LSTM and GRU deep learning algorithms. Based on the result of our experiments, our encoder decoder model using the Bi-LSTM algorithm has a better BLEU score. The encoder decoder model using the Bi- LSTM algorithm scored 24.8 for English-Tigrigna translation and 24.4 for Tigrigna-English translation. The model achieves a BLEU score of +0.8 from a baseline translation model on the area. |
