Nisma Fadillah Abdul Rahman Onge, Tirana Noor Fatyanosa, Irawati Nurmala Sari
Communication barriers faced by the deaf community in Indonesia significantly impact their access to healthcare, education, and employment. While Indonesian Sign Language (BISINDO) serves as the primary communication method, the lack of automated translation systems limits broader accessibility. This paper presents an optimization approach for IndoBART, a pre-trained transformer model specifically designed for Indonesian language tasks. IndoBART was selected due to its proven effectiveness in Indonesian NLP tasks and its bidirectional encoder-autoregressive decoder architecture, which is well-suited for sequence-to-sequence translation. Previous research has explored hyperparameter optimization for text-togloss translation and achieved strong results using mBART on other sign languages, but limited work exists on Indonesianspecific pretrained models for BISINDO. Although augmentation techniques have shown promise in low-resource settings, their systematic evaluation for Indonesian sign language remains underexplored. To address this gap, back translation data augmentation at several proportions (30%, 50%, 80%, and 100%) is evaluated in combination with Optuna-based hyperparameter tuning. The fine-tuning process involves training IndoBART's encoder-decoder architecture on text-gloss pairs, in which the encoder processes Indonesian text through bidirectional attention layers and the decoder autoregressively generates BISINDO gloss sequences. The combined use of 100% back translation augmentation and optimized hyperparameters yields the best performance, representing substantial improvements over the baseline. These findings establish a foundation for advancing accessible communication technology for the Indonesian deaf community. © 2026 IEEE.
Universitas Brawijaya, Department of Informatics Engineering, Malang, Indonesia
