Implementasi Augmentasi Data Pada Konversi Text To-Gloss Bisindo Menggunakan Flan-T5 Dengan Low-Rank Adaptation

Abstract

Text-to-gloss system is one of the components in developing inclusive communication technology for the hearing-impaired community, particularly for Indonesian Sign Language (BISINDO). This research implements the FLAN-T5 model with Low-Rank Adaptation (LoRA) technique to address dataset limitations in BISINDO text-to-gloss conversion tasks. Data augmentation based on combining preprocessing was applied to enrich the limited dataset through lemmatization, producing an augmented dataset of 3000 text-gloss pairs from 1500 original data. The FLAN-T5 Small model was configured with LoRA rank 8 on Query and Value matrices, training only 0.44% of total parameters. Evaluation was conducted using BLEU, ROUGE-1, ROUGE-2, and ROUGE-L metrics at several epoch checkpoints to compare model performance with original and augmented data. Experimental results demonstrated significant improvement across all evaluation metrics for the model trained with augmented data compared to the model with original data. Qualitative analysis of 50 prediction samples identified three main error patterns: word duplication indicating decoder issues, loss of important words showing semantic understanding limitations, and lemmatization errors in Indonesian morphology preprocessing. This study demonstrates that the combination of FLANT5 with LoRA and data augmentation can improve the performance of BISINDO text-to-gloss systems by reducing overfitting risk under low-resource dataset conditions, although it requires larger dataset development and more diverse augmentation methods to achieve better results.