<![CDATA[This research aims to address the challenges of preserving the Angkola language in the digital era, which are exacerbated by the lack of an adequate digital data corpus, by developing an accurate and efficient automatic Angkola-to-Indonesian machine translation system. The proposed method focuses on a fine-tuning approach for the Multilingual Text-to-Text Transfer Transformer (mT5-base) model using an Angkola-Indonesian text data corpus.The initial dataset, consisting of Angkola-Indonesian sentence pairs, was cleaned, resulting in 28,775 sentence pairs used for training. The data was subsequently split into 70% training data (20,142 lines), 15% validation data (4,316 lines), and 15% test data (4,317 lines). Intelligent model performance optimization was conducted using Optuna Hyperparameter Tuning to find the best hyperparameter combination. Optuna's objective function was designed to maximize a composite score based on the BLEU and chrF metrics from the validation evaluation results. The optimization process yielded the best Trial (Trial 50) with key hyperparameters: learning rate = 0.0004316 and num beams = 4. The best model obtained from the fine-tuning process was then evaluated on a separate Test dataset. The final evaluation on the test data using standard translation metrics demonstrated excellent performance, achieving a BLEU score of 73.84 and a chrF score of 83.34. Overall, this research successfully implemented hyperparameter optimization using Optuna for the mT5 model, resulting in an Angkola-to-Indonesian translation model that exhibits high accuracy and more efficient performance. These results provide a tangible contribution to the preservation of the Angkola language by offering a modern and accurate translation tool.]]>
Copyrights © 2026
