Enhancing Vietnamese Automatic Question Generation with Synthetic Data and Preference Optimization

Automatic Question Generation (AQG) systems have long been recognized for their diversity of applications, particularly in generating questions for education and machine reading comprehension tasks. A key challenge in AQG is ambiguity, where multiple questions can lead to the same answer given a context. Existing works in Vietnamese Automatic Question Generation (ViAQG) mainly focus on exploiting the connection between context, answer, and question, with little attention to the problem that there can be multiple acceptable questions for a given context and answer, and only a few of these are of high quality. Therefore, we aim to enhance the question generation system to generate high-quality questions from an acceptable set frequently. In this paper, we propose a selection method for the preference dataset which is then utilized for the Direct Preference Optimization (DPO) algorithm to align the Supervised Fine-tuning (SFT) question generation model. Additionally, we expand the training dataset by using an innovative prompting technique to generate synthetic data with a Large Language Model (LLM). Our approach enhances the model's stability on the UIT-ViQUaD dataset across different sampling temperatures, generating more relevant questions. It also performs better in automatic and AI-based evaluation metrics, outperforming several model alignment methods. Our work contributes to developing a better AQG system by improving accuracy thus adding value to its applications.