DEEP LEARNING-BASED GURAGIGNA TO AMHARIC MACHINE TRANSLATION

Abstract

Machine translation is an application of NLP, which can be used to translate text from onenatural language to another natural language. In this study, we aimed to develop Deep Learning Based Guragigna to Amharic Translation, recognizing Natural Language Processing as a pivotal domain within AI facilitating human-computer language interaction. Previously, there is no research conducted on machine translation between Guragigna and Amharic. Given the abundance of information in Amharic across various domains in Ethiopia, including legal, media, religious, educational, and governmental documents, it becomes imperative to bridge the language gap for the growing Guragigna-speaking population. Neural Machine Translation (NMT) is a recently proposed approach to machine translation (MT) that has achieved the state-of-the-art translation quality in recent years. Unlike traditional MT approaches, NMT aims to create a single neural network that can be tuned collaboratively to maximize translation performance. So, the aim of this study is to develop Deep learning Amharic-guragigna bi-directional machine translation.To conducted experiments employing six encoder-decoder models: LSTM, Bi-LSTM, LSTM+attention, CNN+attention, GRU and Transformers. Collected a dataset of 9,515 parallel sentences, and evaluated the models based on efficiency metrics, including training time, memory usage, and BLEU score, to propose an optimal translation model and utilize the 80/20 splitting technique for dividing the dataset into training and testing sets. Achieving among those models, the transformer model outperforms other models by 99.4% accuracy, 0.0113 loss and a BLEU score of 9.93 for Amharic-Guragigna translation and 9.99 for Guragigna-Amharic machin translation. Because transformer process the whole sentence simultaneously, which reduces training time and it computes similarity scores between words in a sentence by itself means self attention. Due to the problem of unavailable parallel corpus, we have trained our model with minimum corpus, though NMT requires huge data for training and create an optimal model that learn the different features of the two languages and also challenges with LSTM, Bi-LSTM, LSTM+attention and GRU models, which required significant memory resources.