Hybrid Artificial Neural Machine Translation using Deep Learning Techniques English-to-Afaan Oromoo

Abstract

Artificial neural networks are at the basis of the most state-of-the-art for a variety of activities and have enjoyed great success in a variety of tasks such as machine translation, image recognition, speech recognition, text summarization, and in another computation. Neural networks with recurrent model so-called NMT, have recently been applied to machine translation and begun to show successive results. However, as a recently appeared method, the approach has some restrictions. A neural machine translation system typically has to apply a vocabulary of some size to prevent the time-usage in training and decoding; consequently, it produces a critical out-ofvocabulary difficulty. Moreover, the decoder lacks a mechanism to assurance all the source words to be translated and usually resulted short translations, resulting in in???uent but insufficient translations. The problems, are addressed by modeling statistical machine translation, components with the basis of NMT model. The main aim of this research is to design a model that significantly reduces translation problems shown in the neural machine translation system. The proposed model is implemented by classifying the model into three layers namely; encoder layer, hybrid layer and decoder layer. Our experiment result shows that the proposed method signi???cantly improved the translation quality of the state-of-the-art neural machine translation system on English-to-Afaan Oromoo translation tasks prepared dataset. The problems; out-of-vocabulary lead to unknown word output, lack of translation problem for long sentences, lack of decoding mechanism, which badly hurt the translation quality in the neural machine translation system is reduced by our proposed hybrid neural machine translation model. Our approach resulted in a gain of up to 2.4 BLEU score on test sets. The BLEU score evaluation of the translations obtained with this combination increased, and this also results in improved translation quality as observed in our experiments. Experiments on English-to-Afaan Oromoo translation tasks show that our system achieves signi???cant improvements over the reference on a small amount of the training dataset collected from the web. The proposed method achieved better performance compared with the previous models (Recurrent Neural Network-search, and Statistical Machine Translation).