Dayang Alyssa Raisaputri, Tirana Noor Fatyanosa, Irawati Nurmala Sari
This study systematically investigates the potential and limitations of FLAN-T5 with Low-Rank Adaptation (LoRA) for low-resource Indonesia Sign Languange (BISINDO) text-togloss conversion, providing empirical insights to guide future development of Indonesian Sign Language translation systems. To address the challenge of limited BISINDO datasets, we construct a text-to-gloss dataset and apply lemmatization-based data augmentation to expand the training data. Evaluation across three epoch checkpoints on both original and augmented datasets demonstrates that augmentation consistently improves model performance across all metrics. Despite these improvements, qualitative error analysis reveals persistent challenges in word reordering, character-level corruption, and partial structural transformation, indicating that the model has not fully learned the structural mapping between Indonesian sentence structure and BISINDO gloss ordering. These findings establish an empirical baseline and identify key directions for future research, including larger dataset development, more diverse augmentation strategies, and improved fine-tuning approaches for low-resource sign language tasks. © 2026 IEEE.
Brawijaya University, Department of Informatics Engineering, Malang, Indonesia
