Transliterating Nom Script into Vietnamese National Script Using Multilingual Neural Machine Translation

Traditional methods for Nom transliteration have been relying on statistical machine translation. While there have been initial attempts to apply neural machine translation (NMT) to Nom transliteration, these approaches have not effectively utilized the shared characteristics of multilingualism. To overcome this, we propose a new method that employs a highly shared parameter multilingual NMT model, using a unified parameter set the encoder and decoder. This approach utilizes a single encoder-decoder pair for all translation directions. Experimental evaluation of our proposed method demonstrates a significant improvement of +6.3 BLEU score compared to the NMT baseline. We also conduct preliminary experiments on incorporating BERT-NMT into multilingual machine translation to enhance the performance of low-resource translation tasks in general.