Kreol Morisien to English and English to Kreol Morisien Translation System using Attention and Transformer Model

Abstract

Translation from one language to another is extensively carried out by travellers, students and many other people. Unfortunately, less popular languages, such as Kreol Morisien (KM), are not catered for by popular translation systems. The objective of this work is to develop a system that translates English into Kreol Morisien and vice-versa. The size of the dataset which was used in the model consists of 19,650 pairs of English and Kreol Morisien sentences. The Kreol Morisien words are up to date since they were taken from the third edition of the Diksioner Morisien. A neural machine translation system has been used in this work. A transformer model with attention was then developed which was trained for several iterations. Evaluation of the system is done using the standard BLEU scores. The English to Kreol Morisien model achieves a BLEU score of 0.20 and the Kreol Morisien to English model achieves a BLEU score of 0.23. Both are much higher than existing systems. Furthermore, user evaluation has been carried out in the form of two surveys. Each survey consisted of 25 pairs of sentences in the source language and the target language. Responses were gathered from people with different age groups living in both rural and urban regions and including both students and professionals. We received more than 90 responses for each survey. Evaluation and testing of the translation model using the BLEU score showed that the model can produce satisfactory translations.