Refining the Performance of Indonesian-Javanese Bilingual Neural Machine Translation Using Adam Optimizer

This study focuses on creating a Neural Machine Translation (NMT) model for Indonesian and Javanese languages using Long Short-Term Memory (LSTM) architecture. The dataset was sourced from online platforms, containing pairs of parallel sentences in both languages. Training was performed with the Adam optimizer, and its effectiveness was compared to machine translation (MT) conducted without an optimizer. The Adam optimizer was utilized to enhance the convergence speed and stabilize the model by dynamically adjusting the learning rate. Model performance was assessed using BLEU (Bilingual Evaluation Understudy) scores to evaluate translation accuracy across different training epochs. The findings reveal that employing the Adam optimizer led to a significant enhancement in model performance. At epoch 2000, the model using the Adam optimizer achieved the highest BLEU score of 0.989957, reflecting very accurate translations, whereas the model without the optimizer showed lower results. Furthermore, translations from Indonesian to Javanese were found to be more precise than those from Javanese to Indonesian, largely due to the intricate structure and varying speech levels of the Javanese language. In summary, the implementation of the LSTM method with the Adam optimizer significantly improved the accuracy of bidirectional translations between Indonesian and Javanese. This research contributes notably to the advancement of local language translation technologies, supporting language preservation in the digital age and holding promise for applications in other regional languages.