Educational Multi-Question Generation for Reading Comprehension

Abstract

Automated question generation has made great advances with the help of large NLP generation models. However, typically only one question is generated for each intended answer. We propose a new task, Multi-Question Generation, aimed at generating multiple semantically similar but lexically diverse questions assessing the same concept. We develop an evaluation framework based on desirable qualities of the resulting questions. Results comparing multiple question generation approaches in the two-question generation condition show a trade-off between question answerability and lexical diversity between the two questions. We also report preliminary results from sampling multiple questions from our model, to explore generating more than two questions. Our task can be used to further explore the educational impact of showing multiple distinct question wordings to students.

Diversity question generation based on contrastive search algorithm

“…To meet the needs of various domain-specific scenarios, a study has been proposed to generate diverse questions, focusing on different expressions while conveying the same meaning, given an answer and its relevant context. Currently, this research can be categorized into two main classes: enhancing DQG by utilizing internal or external knowledge [10][11][12] and improving the quality of DQG by refining the decoding process [9,13]. In the realm of leveraging internal or external knowledge to enhance DQG, Wang et al [10] introduced a novel content selector that utilizes the continuous latent variable modeling technique of a Conditional Variational Encoder (CVE).…”

“…To evaluate the proposed model, we adopted the PL-DQG dataset created by Rathod et al [9]. This dataset is derived from the publicly available SQuAD dataset [14], with questions transformed into a second set by pre-trained language models.…”

“…PINC [9]：A heuristic evaluation metric, PINC, was used to measure the n-gram overlap between pairs of questions [36] and for the two generated questions q1 and q2, the PINC score is computed.…”

“…However, the above studies on QG only investigate the generation of a question given a relevant passage, which cannot better meet the needs of a specific domain. Therefore, researchers [9][10][11][12][13] proposed diversity question generation (DQG) to improve the quality using external, internal knowledge or by improving the decoding end. But the existing studies still have the problem of low answerability of generated questions due to the repetition of generated question words as well as the existing studies are one-sided and only study the paragraph-level diversity question generation (PL-DQG).…”

Reading bots: The implication of deep learning on guided reading

“…A much greater effort was placed into generating diverse and hard questions. High-quality questions were generated based on the pre-trained language models (Wang et al, 2018 ; Kumar et al, 2019 ; Pan et al, 2020 ; Cheng et al, 2021 ) and, in particular, for educational purposes (Stasaski et al, 2021 ; Rathod et al, 2022 ; Zou et al, 2022 ). In addition, for multiple-choice questions, distractor generation also received due attention (Liu et al, 2005 ; Susanti et al, 2018 ; Gao et al, 2019 ; Qiu et al, 2021 ; Ren and Zhu, 2021 ; Zhang and VanLehn, 2021 ).…”