Qadg: Generating question–answer-distractors pairs for real examination

Reading comprehension question generation aims to generate questions from a given article, while distractor generation involves generating multiple distractors from a given article, question, and answer. Most existing research has mainly focused on one of the above tasks, with limited attention to the joint task of Question–Answer-Distractor (QAD) generation. While previous work has achieved success in the joint generation of answer-aware questions and distractors, applying these answer-aware approaches to practical applications in the education domain remains challenging. In this study, we propose a unified and high-performance Question–Answer-Distractors Generation model, named QADG. Our model comprises two components: Question–Answer Generation (QAG) and Distractor Generation (DG). This model is capable of generating Question–Answer pairs based on a given context and then generating distractors based on the context and QA pairs. To address the unconstrained nature of question-and-answer generation in QAG, we employ a key phrase extraction as reported by Willis (in: proceedings of the Sixth ACM Conference on Learning@ Scale, 2019) module to extract key phrases from the article. The extracted key phrases, as the constraints that can be used to match answers. To enhance the quality of distractors, we propose a novel ranking-rewriting mechanism. We employ a fine-tuned model to rank distractors and introduce a rewriting module to improve the quality of distractors. Furthermore, the Knowledge-Dependent-Answerability (KDA) as reported by Moon (Evaluating the knowledge dependency of questions, 2022) is used as a filter to ensure the answerability of the generated QAD pairs. Experiments on SQuAD and RACE datasets demonstrate that the proposed QADG exhibits superior performance, particularly in the DG phase. Additionally, human evaluations also confirm the effectiveness and educational relevance of our model. © The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2024.