question_generation FINE-TUNING VIT5 MODEL FOR VIETNAMESE MULTIPLE-CHOICE QUESTION GENERATION

Abstract

This paper addresses the demand for automated question generation in EdTech, overcoming the limitations of current models that struggle to generate structured JSON data for system integration. The author proposes fine-tuning the ViT5 model on the Custom MCQ dataset comprising 13,342 samples, constructed with a Bloom taxonomy balancing strategy and strict data leakage prevention procedures. Key findings indicate that ViT5's SentencePiece tokenizer significantly outperforms BARTpho's syllable-based mechanism in preserving JSON syntax integrity. Based on this, the study applies Label Smoothing to enhance distractor diversity and an Early Stopping strategy to identify the optimal convergence point at epoch 6. Final experimental results show the model achieving a METEOR score of 45.59, confirming its practical applicability.