The Effect of Bahdanau Attention Mechanism on Javanese to Indonesian Translation

Innovations in Natural Language Processing (NLP), particularly in Machine Translation (MT), are continuously developed to enhance efficiency. MT has significant potential to facilitate cross-language communication and increase productivity. The need for Machine Translation is particularly relevant in Indonesia, due to the diversity of regional languages such as Javanese, Sundanese, and Madurese. Javanese, as one of the largest spoken regional languages, has unique complexities such as levels of politeness (ngoko, krama and krama inggil). This complexity presents a challenge in creating an accurate translation model, given the significant structural and sociolinguistic differences between the two languages. The research developed Machine Translation from Javanese to Indonesian using Neural Machine Translation (NMT), with Bidirectional Long Short Term Memory (Bi-LSTM) model and Bi-LSTM with Attention Bahdanau mechanism. This study aims to compare the two models performance by highlighting the Attention mechanism's effectiveness in improving translation accuracy and quality. The dataset includes 23,015 sentences from various sources. The results show that the Bi-LSTM with Attention (64 Units) achieved 99% training accuracy and 82% validation accuracy, outperforming the Bi-LSTM model without Attention (64 units), which obtained 89% and 71%, respectively.