(IJACSA) International Journal of Advanced Computer Science and Applications, Vol. 11, No. 5, 2020 602 | Page www.ijacsa.thesai.org Development of a Recurrent Neural Network Model for English to Yorùbá Machine Translation Adebimpe Esan 1 *, John Oladosu 2 *, Christopher Oyeleye 3 *, Ibrahim Adeyanju 4 Olatayo Olaniyan 5 , Nnamdi Okomba 6 , Bolaji Omodunbi 7 , Opeyemi Adanigbo 8 Department of Computer Engineering, Federal University Oye-Ekiti, Ekiti state, Nigeria 1,4,5,6,7,8 Ladoke Akintola University of Technology, Ogbomoso 2,3 Abstract—This research developed a recurrent neural network model for English to Yoruba machine translation. Parallel corpus was obtained from the English and Yoruba bible corpus. The developed model was tested and evaluated using both manual and automatic evaluation techniques. Results from manual evaluation by ten human evaluators show that the system is adequate and fluent. Also, results from automatic evaluation shows that the developed model has decent and good translation as well as higher accuracy because it has better correlation with human judgment. Keywords—Recurrent; tokenizer; corpus; translation; evaluation; correlation I. INTRODUCTION The demand for translation and translation tools currently exceeds the capacity of available solution [1], hence, the need to intensify research in the field of machine translation [2]. Machine Translators (MT) accept characters of source language and map to the characters of the target language to generate the words with the help of various rules and other learning process techniques [3]. Previous researchers have employed various approaches to develop machine translators and the approaches were categorized into two by [4], namely; single and hybrid approaches. Single approaches include: rule- based, knowledge-based, statistical and direct approaches while Hybrid approaches are: word-based, phrase-based, syntax-based, forest-based and neural machine translation models. Neural Machine Translation (NMT) is an improvement in the field of machine translation where a large neural network is built and trained to read a sentence and output a correct translation [5]. The approach consists of the encoder and the decoder for encoding a source sentence and decoding it to a target sentence [6]. Neural machine translators have shown promising results than previous MT approaches through the incorporation of some neural components to existing translation systems like phrase-based systems [7]. In addition, research revealed that NMT produces automatic translations that are significantly preferred by humans when compared to other machine translation approaches. However, the most widely used model for NMT is the Recurrent Neural Network model which is a supervised machine learning model that is made of artificial neurons with one or more feedback loops. In order to train a RNN, a parallel corpus is trained so as to minimize the difference between the output and target pairs by optimizing the weights of the network [8]. In addition, a portion of the corpus is used as the validation dataset [9] to watch the procedure during training and prevent the network from underfitting or overfitting. RNNs have distributed hidden states used for storing information about the past efficiently and non-linear dynamics for updating their hidden state [10]. Hence, this research developed a recurrent neural network model for English to Yoruba machine translation. II. RELATED WORKS Neural Machine Translation (NMT) is an improvement in the field of machine translation and it is based purely on deep neural networks. The encoder–decoder architecture [5] which is a conventional approach to neural machine translation, encodes a whole input sentence into a fixed-length vector from which a translation was decoded. Research show that the use of a fixed-length context vector is a challenge for the translation of longer sentences, hence, the research was extended by developing a model that soft- search for a set of input words, or their annotations computed by an encoder, when generating each target word [11]. The method prevents the model from encoding all the source sentences into a fixed- length vector but focuses only on relevant information that will help to generate target word. This approach outperformed the conventional encoder-decoder model significantly. However, as training and decoding complexities increase proportionally to the number of target words in previous NMT systems, the size of the target vocabulary was extended by using an approach that enables training a model with much larger target vocabulary without substantial increase in computational complexity [12]. Decoding was efficiently done using a very large target vocabulary by selecting a small portion of the target vocabulary. Research show that the models trained outperformed the baseline models with a small vocabulary size. Though, it is unable to translate words which could not be found in the vocabulary. Therefore, alignment- based technique was used by [13] to mitigate this problem. The technique was carried out by training the model on data that is augmented by the output of a word alignment algorithm, allowing the NMT system to emit, for each Out of Vocabulary (OOV) word in the target sentence, the position of its corresponding word in the source sentence. Moreover, [14] developed a multi-task learning model by training a unified neural machine translation model. In the research, an encoder is shared across different language pairs and each target language has a separate decoder. The *Corresponding Author