Implementation of Neural Machine Transliteration from Komering Language to Indonesian Language Using Recurrent Neural Network (RNN) Model

Language is an important tool for communicating and understanding expressed ideas. However, there are many different regional languages in the world, making translation tools necessary to connect speakers of different languages. This research aims to develop a Neural Machine Translation (NMT) model for translating from Komering to Indonesian and vice versa using the Recurrent Neural Network (RNN) algorithm. The development of the NMT model begins with the creation of the dataset, preprocessing, the development of the RNN model with Gated Recurrent Unit (GRU) components, and finally evaluation to obtain the best model. This study uses four architectures of RNN-GRU with a combination of network structures and batch size numbers. The accuracy metric is used to provide an initial overview of the performance of each evaluated RNN model. The testing results for translating from Komering to Indonesian and vice versa from Indonesian to Komering show that Model 3 with 3 GRU layers performs exceptionally well with the lowest loss and the highest accuracy of 81.32% and 89.96%. The evaluation results demonstrate that adding GRU layers improves NMT accuracy. In addition, the results obtained from testing all models for NMT indicate that larger batch size values yield better performance. The best model developed can subsequently be implemented in translating other languages.