Adaptive and Personalized Exercise Generation for Online Language Learning

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 10184–10198 July 9-14, 2023 ©2023 Association for Computational Linguistics Adaptive and Personalized Exercise Generation for Online Language Learning Peng Cui Mrinmaya Sachan Department of Computer Science, ETH Zürich {peng.cui, mrinmaya.sachan}@inf.ethz.ch Abstract Adaptive learning aims to provide customized educational activities (e.g., exercises) to ad- dress individual learning needs. However, man- ual construction and delivery of such activities is a laborious process. Thus, in this paper, we study a novel task of adaptive and personalized exercise generation for online language learn- ing. To this end, we combine a knowledge trac- ing model that estimates each student’s evolv- ing knowledge states from their learning history and a controlled text generation model that gen- erates exercise sentences based on the student’s current estimated knowledge state and instruc- tor requirements of desired properties (e.g., do- main knowledge and difficulty). We train and evaluate our model on real-world learner in- teraction data from Duolingo and demonstrate that LMs guided by student states can gener- ate superior exercises. Then, we discuss the potential use of our model in educational ap- plications using various simulations. These simulations show that our model can adapt to students’ individual abilities and can facilitate their learning efficiency by personalizing learn- ing sequences.1 1 Introduction Adaptive learning technologies which continuously monitor student progress to dynamically adjust the level or type of learning materials based on the in- dividual’s abilities are quite popular (Becker et al., 2018). Empirical studies have shown various bene- fits of adaptive learning, such as improved student learning outcomes (Bailey et al., 2018; Holthaus et al., 2019), lower dropout rates (Daines et al., 2016), and increased instructor satisfaction (Yarnall et al., 2016). Despite their effectiveness, designing adaptive systems is challenging as it usually in- volves planning a series of exercises that is person- alized and adaptive to each student, which requires 1Our implementation is available at https://github.com/ nlpcui/AdaptiveQG. Figure 1: We first assess student knowledge states from their learning history and then generate exercises based on estimated states and instructor control of desired properties including domain knowledge (vocabulary) and difficulty levels (expected error numbers). diverse exercise planning as well as an understand- ing of the student learning process. On the other hand, powered by advances in neu- ral NLP, works have been done for automatically generating text-based exercises or questions for educational purposes in second language learn- ing (Heck and Meurers, 2022; Perez and Cuadros, 2017), mathematics (Polozov et al., 2015; Zhou and Huang, 2019; Wang et al., 2021), and com- puter science (Susanti et al., 2017). Nevertheless, how to apply these approaches in adaptive systems remains an open question. First, existing meth- ods largely rely on pre-defined question templates or specified information sources (e.g., a passage), thereby resulting in limited knowledge coverage and low question difficulty control, and as a conse- quence, do not meet each student’s individual and nuanced learning needs. Besides, they are usually designed to generate standalone exercises, whereas …

Content truncated to 500 words