GE’EZ-AMHARIC MACHINE TRANSLATION USING DEEP LEARNING

Abstract

Neural machine translation (NMT) which has come to be the breakthrough in the field of machine translation is now greatly being used by many translation services such as google translate. This generic deep learning approach of machine translation (MT) with the help of attention mechanism is used as the core method of our Ge’ez-Amharic translation. Despite their low accuracy, low speed of translation, and involvement of linguistic professionals, other methods of Ge’ez-Amharic translation such as using Statistical Machine Translation (SMT), morpheme-based have already been researched. The limitation of these approaches is the complex nature of the design of the models. We used a unidirectional model which only translates from Ge’ez to Amharic. We designed an NMT encoder-decoder translation model based on attention mechanism that contains two Long Short-Term Memory (LSTM) layers with 500 hidden units both in the encoder and decoder parts. The model takes source sentence as input in the encoder side and generates a target sentence as output in the decoder side, generating a single word at a time. We used attention mechanism to handle long term dependencies in long sentences by paying attention to the parts of the input sentence which contain relevant information in generating a single word in the target sentence. We have collected 50k Ge’ez-Amharic parallel sentences and used different portions of this data for the different experiments. OpenNMT was employed for developing our model and Bilingual Evaluation Under Study (BLEU) is used for evaluating the translation quality of our model. The proposed model was trained using different experiments on Colab and we found the best performing translation with BLEU score of 15.4%. Despite the hungry nature of NMT models for data, the model has performed well with the collected corpus.