arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2412.09248cs.CY
收藏

教育领域知识追踪与大语言模型的系统综述:机遇、问题与未来研究

A Systematic Review of Knowledge Tracing and Large Language Models in Education: Opportunities, Issues, and Future Research

Yongwan Cho, Rabia Emhamed AlMamlook, Tasnim Gharaibeh

AI总结:

本系统综述梳理教育领域知识追踪(KT)与大语言模型(LLMs)的结合方向,分析二者融合的三大功能,指出LLMs可提升KT性能但二者均受限于结构化数据集,为后续研究提供方向。

摘要:

知识追踪(Knowledge Tracing,KT)是旨在通过学习交互评估学生知识状态的研究领域,是智能教学系统(Intelligent Tutoring Systems,ITSs)的核心组成部分。尽管已取得显著进展,但目前尚无KT模型能同时在预测准确性和可解释性上均表现优异。与此同时,在海量自然语言数据集上预训练的大语言模型(Large Language Models,LLMs)已成为强大工具,在各类教育应用中具备巨大潜力。本系统综述探讨了教育场景下KT模型与LLMs结合的交叉点、机遇与挑战。综述首先调研了LLMs在教育中的应用,包括其对特定领域内容的适配能力,以及支持个性化学习的能力。随后梳理了KT模型从传统方法到前沿方法的发展与现状,旨在发掘LLMs可缓解的潜在挑战。综述核心聚焦LLMs与KT的融合,探讨了三大主要功能:解决KT领域的共性问题、克服特定KT模型的局限性、以及直接充当KT模型本身。研究发现,通过上下文学习、基于智能体的方法等定制化技术,可针对特定教育任务定制LLMs,有效处理复杂且不均衡的教育数据。这类模型能通过从题目数据中生成相关特征,提升现有KT模型的性能并解决冷启动问题。但当前两类模型都高度依赖结构化的有限数据集,未能利用多样化教育数据的价值,而这类数据本可提供对个体学习者的更深入洞察,支持各类教育场景。

英文摘要:

Knowledge Tracing (KT) is a research field that aims to estimate a student's knowledge state through learning interactions-a crucial component of Intelligent Tutoring Systems (ITSs). Despite significant advancements, no current KT models excel in both predictive accuracy and interpretability. Meanwhile, Large Language Models (LLMs), pre-trained on vast natural language datasets, have emerged as powerful tools with immense potential in various educational applications. This systematic review explores the intersections, opportunities, and challenges of combining KT models and LLMs in educational contexts. The review first investigates LLM applications in education, including their adaptability to domain-specific content and ability to support personalized learning. It then examines the development and current state of KT models, from traditional to advanced approaches, aiming to uncover potential challenges that LLMs could mitigate. The core of this review focuses on integrating LLMs with KT, exploring three primary functions: addressing general concerns in KT fields, overcoming specific KT model limitations, and performing as KT models themselves. Our findings reveal that LLMs can be customized for specific educational tasks through tailor-making techniques such as in-context learning and agent-based approaches, effectively managing complex and unbalanced educational data. These models can enhance existing KT models' performance and solve cold-start problems by generating relevant features from question data. However, both current models depend heavily on structured, limited datasets, missing opportunities to use diverse educational data that could offer deeper insights into individual learners and support various educational settings.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑