Experiment on a Transformer Model Indonesian-to-Sundanese Neural Machine Translation with Sundanese Speech Level Evaluation
Restu Bias Primandhika, Muhammad Nadzeri Munawar, Aceng Ruhendi Saifullah · Advances in Social Science, Education and Humanities Research/Advances in social science, education and humanities research · 2021
Speech level is one of the essential Sundanese language elements.As Indonesian mixed within Sundanese language use, the usage of speech level is gradually degrading.Indonesian, in order to get correct word choice in Sundanese language, social contexts may refer to many sources such as a dictionary, or thesaurus.However, for better translation in syntax and context, machine translation is offered.Based on the fact, this experiment focuses on the problem when translating Indonesian to Sundanese and the evaluation of Sundanese speech level in the translated texts.Neural machine translation (NMT) was chosen as the current technology in machine translation, which worked by combining recurrent neural network encoder-decoder.The experiment started with building 50.000Sundanese-Indonesian sentences as a parallel corpus to build and train NMT models.The experiment on sentence training in Transformer NMT without outof-vocabulary (OOV) shows 42.72% BLEU Score, and Average Training Loss was 1.77 while for speech level was dominated by 56% basa loma (coarse) of the whole testing result.