AI enhanced adaptive tutoring system by generating individualized questions and answers based on cognitive diagnostic assessment

今年我继续进行学习者知识评估（LKA）的工作。我进一步探索了细粒度评估和可解释性的研究。在我之前的工作[BESC ' 20]的基础上，我提出了一种新的模型，它不仅可以输出学习者的细粒度知识状态，还可以输出项目特征，使其具有可解释性。从六个角度对五个真实数据集进行了广泛的模型分析，验证了其优越性。这项研究已经发表在顶级期刊《神经计算》上。另一项工作解决了数据稀疏和信息丢失的基本问题，同时提高了模型的性能。本文探讨了将知识结构（KS）纳入LKA以潜在地解决上述问题。该工作从学习日志中自动生成KS，并提出了一种新的带有注意机制的图模型。大量的实验证明了该方法的有效性。这项工作已发表在顶级期刊[IJIS]上。上述工作激发了多模态学习分析的新思路。我发表了一篇关于使用多模态分析为智慧教育提供见解的经验证据的综述论文。我还参与了一篇发表在[ICCE ' 21]上的文章，其中提出了一种基于图的LKA方法。我也完成了我的博士论文，在这篇论文中我总结了我的博士工作。在此基础上，结合学习者建模和领域建模，提出了动态LKA的通用框架。基于这一框架，本文提出了三种方法，每种方法都针对现有研究中的一个具体问题。

This year I continue the work on learner's knowledge assessment (LKA). I have further explored the research of fine-grained assessment and interpretability. Improved on my previous work [BESC’20], I propose a novel model that can not only output the learners’ fine-grained knowledge states but also the item characteristics, enabling the interpretability. Extensive model analyses conducted from six perspectives on five real-world datasets validate its superiority. This work has been published in a top journal [Neurocomputing].Another work solves the fundamental issues of data sparseness and information loss while improving the model performance. It has explored to incorporate the knowledge structure (KS) into the LKA to potentially resolve the above issues. This work automatically generates the KS from the learning logs and proposes a novel graph model with the attention mechanism. Extensive experiments show the effectiveness. This work has been published in a top journal [IJIS].The above work stimulates a new idea of multimodal learning analysis. I have published a review paper about the empirical evidence on the usage of multimodal analysis to provide insights for smarter education. I also participated in a work published in [ICCE’21], in which a graph-based method is proposed for LKA.I also finished my doctoral thesis, in which I summarize my PhD works. Overall, it proposes a general framework for dynamic LKA by integrating both learner and domain modeling. Based on this framework, it proposes three approaches, each addressing one specific issue in existing studies.