A Study for Enhancing Low-resource Thai-Myanmar-English Neural Machine Translation

• DOI: 10.1145/3645111
• Publication Date: 2024-02-13T13:51:05Z
• Authors: Mya Ei San; Sasiporn Usanavasin; Ye Kyaw Thu; Manabu Okumura
• Keywords: 0202 electrical engineering, electronic engineering, information engineering; 02 engineering and technology

Several methodologies have recently been proposed to enhance the performance of low-resource Neural Machine Translation (NMT). However, these techniques yet be explored thoroughly in Thai and Myanmar languages. Therefore, we first applied augmentation such as SwitchOut Ciphertext Based Data Augmentation (CipherDAug) improve NMT Second, enhanced by fine-tuning pre-trained Multilingual Denoising BART model (mBART), where denotes Bidirectional Auto-Regressive Transformer. We implemented three systems: namely, Transformer+SwitchOut, Multi-Source Transformer+CipherDAug, fine-tuned mBART bidirectional translations Thai-English-Myanmar language pairs from ASEAN-MT corpus. Experimental results showed that Transformer+CipherDAug significantly improved Bilingual Evaluation Understudy (BLEU), Character n-gram F-score (ChrF) , Error Rate (TER) scores over baseline Transformer second Edit-Based The achieved notable BLEU scores: 37.9 (English-to-Thai), 42.7 (Thai-to-English), 28.9 (English-to-Myanmar), 31.2 (Myanmar-to-English), 25.3 (Thai-to-Myanmar), 25.5 (Myanmar-to-Thai). also considerably outperformed two baselines, except for Myanmar-to-English pair. all performed similarly most cases. Last, detailed analyses verifying CipherDAug models potentially facilitate improving