GEEZ TO AFAAN OROMOO MACHINE TRANSLATION USING RECURRENT NEURAL NETWORK

Advisor: Dr. M. Kumarasa Machine translation is a sub field of natural language processing that investigates the use of computer software to translate text or speech from one natural language to another. Since the world in which we are living today is occupied by massive languages, about six thousand languages worldwide, the speakers of these much languages need to interact with each other for different global issues. In today’s era where the global population communicates easily with any angle of the world using different communication platforms the need for translation between different languages is vital. This communication gap is solved by using an expert translator. The use of manual translation is expensive and inconvenient. Many researches were done to resolve this problem using machine translation techniques for some of resourced languages. However, the researches done on our local languages are very low. The intension of this thesis is to design and implement Geez-Afaan Oromoo neural machine translation, based on the encoder-decoder Recurrent Neural Network approach. Geez is classical South Semitic language which was used in many inscriptions including religious history, philosophy, medical and other since the early 4th century. Today Geez remains only as a spoken language and the liturgy language of the Ethiopian Orthodox Tewahedo Church and Ethiopian Catholic Church in our country. Where, Afaan Oromoo is the most spoken language in Ethiopia and the official working language of Oromia regional state, and it is primary school language in Oromia, Finfinnee and Dirree Dawaa administrations. Afaan Oromoo is the fourth most widely spoken African language after Arabic, Hausa and Swahili. The machine translation of Geez document into Afaan Oromoo will be of paramount importance in order to enable Afaan Oromoo user to easily access the invaluable indigenous knowledge decoded in Geez language. To train the model, two experiments were conducted using two different RNN algorithms. The first experiment is conducted by using GRU to translate Geez to Afaan Oromoo and has a BLEU score of 73.75%. The second experiment is carried out by using LSTM and has a BLEU score of 77.55%. The result shows that the LSTM approach is slightly better than the GRU approach. Keywords: Machine Translation, Recurrent Neural Network, Gated Recurrent Unit, Long Short-Term Memory, Geez, Afaan Oromoo