Analisis Pengembangan Model Neural Machine Translation (NMT) dengan Transformer untuk Penerjemah Bahasa Indonesia ke Bahasa Gayo

Abstract

The Gayo language has existed alongside the Gayo people for around 7500 years, but its usage has been continually declining, and its digital documentation is very limited. Preserving the Gayo language is crucial for maintaining the cultural heritage and traditions of the Gayo people. To address this challenge, several efforts need to be made to document, preserve, and develop the Gayo language to keep it relevant for future generations. This research conducts an experiment by training a Neural Machine Translation (NMT) model using the OpenNMT framework and the transformer architecture to translate from Indonesian to Gayo. The research adopts a training approach using a parallel corpus collected from the “Kamus Bahasa Indonesia - Bahasa Gayo II”. This dataset undergoes a series of preprocessing steps to prepare it for model training. The preprocessing steps include case folding and removing punctuation. The dataset is divided into three parts, training data, validation data, and test data. The researchers also perform data augmentation using the MixSeq method to enhance the diversity and size of the training data. By utilizing an optimized transformer architecture, the experimental results show an accuracy rate of 90% for the augmented training data. Evaluation using the BLEU score yields a result of 79.90.