DEEP LEARNING BASED AFAAN OROMO-AMHARIC MACHINE TRANSLATION

Abstract

Machine translation is the process of translating text from one human languages to others languages using computer. It has gained much attention currently. There are many approaches available for Machine translation used for European languages, which may not works for Ethiopian languages because of the structure of the languages. Among those machine translation approaches, neural machine translation approaches, that uses artificial neural network is selected in this thesis. In this thesis, we propose bidirectional machine translation between Afaan Oromo language and Amharic language pair. The study was carried out by investigation the capabilities of deep learning approaches on the algorithm of RNN (LSTM) with attention and Transformer model. A total of 22,609 parallel sentences are collected from different sources and used for the experiment with percentage split of 80% for training and 20% for testing. The experiment was conducted to compare performance of two model and select the better model for bidirectional Afaan Oromo -Amharic languages machine translation. Bleu score evaluation metrics have been used to compare the performance of the two models. The result scored from the experiment with BLEU score of 18.5% and 21.2% by the transformer model when translating from Afaan Oromo to Amharic and from Amharic to Afaan Oromo respectively. While using RNN (LSTM) model, we got bleu score of 8.72 % and 10.39% from Afaan Oromo to Amharic and from Amharic to Afaan Oromo respectively. Therefore, it is concluded that transformer model was better in bleu score performance in both direction for languages Afaan Oromo and Amharic languages than RNN (LSTM) model. The transformer model mainly reduce number of trainable parameters, memory needed for training, training time and perform best result even if the length of the sentence increased when compared to LSTM.