Automatic Distractor Generation in Multiple-Choice Questions Using Large Language Models with Expert-Informed Distractor Strategies

Abstract

In recent years, automatic generation of reading-comprehension questions with artificial intelligence has attracted considerable attention. In particular, producing high-quality distractors remains a critical challenge when generating multiple-choice questions (MCQs). Recent studies have increasingly employed large language models (LLMs) to generate distractors for MCQs. However, prior research has relied solely on the implicit, black-box knowledge of LLMs and has seldom exploited human expertise in distractor design. Therefore, in this study, we propose an LLM-based distractor-generation method that explicitly incorporates expert-informed distractor strategies, which represent typical heuristics used by human experts when crafting distractors. Experiments demonstrate that our method produces distractors of higher quality than those generated by previous approaches.