Neural Machine Translation Between Myanmar and English Languages

This study aims to present a thorough overview of cutting-edge machine translation solutions and evaluate their effectiveness in translating between Myanmar and English language pairs.Translating Myanmar poses challenges due to its unique language characteristics and limited resources.These challenges include the complex Burmese script, the tonal nature of the language, its agglutinative structure, and the scarcity of parallel corpora for training.The integration of attention mechanisms, Transformer architecture, and transfer learning in Neural Machine Translation (NMT) models demonstrates their potential to deliver accurate, context-aware, and adaptable translations.A key achievement of this study is highlighting the significant roles of hyperparameter optimization and the use of subword units in enhancing the effectiveness of Transformer-based NMT for Myanmar-English and English-Myanmar language pairs, despite resource constraints.The experimental results showed that using a Transformer-optimized model with a 32k Byte Pair Encoding (BPE) subword model resulted in significant improvements in Bilingual Evaluation Understudy (BLEU), Translation Error Rate (TER), and Character n-gram F-score (ChrF) scores compared to the baseline Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN) and fine-tuned models such as mBART and mT5.The proposed transformer model achieved impressive BLEU scores of 50.77 for English to Myanmar and 48.92 for Myanmar to English.The comparative studies of various machine translation models have confirmed that both the Transformer and fine-tuned models demonstrate promising results in improving low-resource NMT performance for the Myanmar and English languages.