Pivoted Low Resource Multilingual Translation with NER Optimization

Machine translation (MT) has advanced significantly with neural machine translation (NMT) models like BERT, GPT, and MarianMT, which leverage deep learning to provide more accurate and natural translations. However, low-resource languages such as Javanese and Madurese still face challenges due to limited parallel datasets. Pivot-based translation, which uses an intermediary language like Indonesian, has improved translation quality but struggles with preserving named entities, such as names of people or places. Previous models without Named Entity Recognition (NER) often produce inconsistent or inaccurate translations, losing these critical elements. We propose a hybrid approach to address this, combining NER with pivot-based NMT. By identifying and preserving named entities throughout the translation process, this method enhances translation consistency and accuracy for Javanese-Madurese translations. Our experiments demonstrate that NER-Pivot-NMT achieves perfect preservation rates for named entities, although translation accuracy fluctuates between 0.32 and 0.54 for Javanese-to-Madurese and 0.21 to 0.45 for Madurese-to-Javanese. Despite these variations, the overall translation quality improves compared to traditional direct and pivot translation methods. This study highlights the potential of integrating NER with pivot translation to solve challenges in low-resource language translation. It suggests that future research should further optimize NER models and expand datasets to enhance accuracy and fluency. The findings hold promise for improving translation across other regional languages facing similar constraints.