Educational dialogue classification is a critical task for analyzing classroom interactions and fostering effective teaching strategies. However, the scarcity of annotated data and the high cost of manual labeling pose significant challenges, especially in low-resource educational contexts. This article presents the EduDCM framework for the first time, offering an original approach to addressing these challenges. EduDCM innovatively integrates distant supervision with the capabilities of Large Language Models (LLMs) to automate the construction of high-quality educational dialogue classification datasets. EduDCM reduces the noise typically associated with distant supervision by leveraging LLMs for context-aware label generation and incorporating heuristic alignment techniques. To validate the framework, we constructed the EduTalk dataset, encompassing diverse classroom dialogues labeled with pedagogical categories. Extensive experiments on EduTalk and publicly available datasets, combined with expert evaluations, confirm the superior quality of EduDCM-generated datasets. Models trained on EduDCM data achieved a performance comparable to that of manually annotated datasets. Expert evaluations using a 5-point Likert scale show that EduDCM outperforms Template-Based Generation and Few-Shot GPT in terms of annotation accuracy, category coverage, and consistency. These findings emphasize EduDCM’s novelty and its effectiveness in generating high-quality, scalable datasets for low-resource educational NLP tasks, thus reducing manual annotation efforts.
Knowledge tracing is a crucial task in personalized learning that models student mastery based on historical data to predict future performance. Currently, deep learning models in knowledge tracing predominantly use one-hot encodings of question, knowledge, and student IDs, showing promising results. However, they face a significant limitation: a vocabulary gap that impedes the processing of new IDs not seen during training. To address this, our paper introduces a novel method that incorporates aggregated features, termed ‘side information’, that captures essential attributes such as student ability, knowledge mastery, and question difficulty. Our approach utilizes side information to bridge the vocabulary gap caused by ID-based one-hot encoding in traditional models. This enables the model, once trained on one dataset, to generalize and make predictions on new datasets with unfamiliar students, knowledge, or questions without the need for retraining. This innovation effectively bridges the vocabulary gap, reduces the dependency on specific data representations, and improves the overall performance of the model. Experimental evaluations on five distinct datasets show that our proposed model consistently outperforms baseline models, using fewer parameters and demonstrating seamless adaptability to new contexts. Additionally, ablation studies highlight that including side information, especially regarding students and questions, significantly improves knowledge tracing effectiveness. In summary, our approach not only resolves the vocabulary gap challenge but also offers a more robust and superior solution across varied datasets.
The integration of large language models (LLMs) into intelligent tutoring systems offers transformative potential for personalized learning in higher education. However, most existing learning path planning approaches lack transparency, adaptability, and learner-centered explainability. To address these challenges, this study proposes a novel Multi-Agent Learning Path Planning (MALPP) framework that leverages a role- and rule-based collaboration mechanism among intelligent agents, each powered by LLMs. The framework includes three task-specific agents: a learner analytics agent, a path planning agent, and a reflection agent. These agents collaborate via structured prompts and predefined rules to analyze learning profiles, generate tailored learning paths, and iteratively refine them with interpretable feedback. Grounded in Cognitive Load Theory and Zone of Proximal Development, the system ensures that recommended paths are cognitively aligned and pedagogically meaningful. Experiments conducted on the MOOCCubeX dataset using seven LLMs show that MALPP significantly outperforms baseline models in path quality, knowledge sequence consistency, and cognitive load alignment. Ablation studies further validate the effectiveness of the collaborative mechanism and theoretical constraints. This research contributes to the development of trustworthy, explainable AI in education and demonstrates a scalable approach to learner-centered adaptive instruction powered by LLMs.