Contributeurs:
État de publication: Publiée (2025 )
Titre du livre: Artificial Intelligence in Education
Éditeur: Springer, Cham
Volume:
Intervalle de pages: 319–332
ISBN: 978-3-031-98420-4
Résumé: Fully data-driven knowledge tracing approaches require the acquisition of large amounts of data. However, given the cold-start problem in learning systems, it is crucial to accurately model a learner’s knowledge even with little or no available data. Traditional approaches rely on expert knowledge to design preliminary models, such as Bayesian Networks (BNs), to estimate learners’ knowledge. However, obtaining expert input is often costly, time-consuming, and subject to variability due to the implicit nature of human expertise. In this study, we explore an alternative approach: leveraging Large Language Models (LLMs) to generate a BN for knowledge tracing. Given a predefined set of domain-specific knowledge elements (skills), we prompt an LLM to construct a BN that represents the relationships between these elements. To evaluate the effectiveness of this approach, we integrate the LLM-generated BN with Deep Knowledge Tracing (DKT) and test it on three datasets: a private dataset with a BN fully designed by domain experts, as well as two publicly available datasets, Bridge to Algebra and ASSISTments. The results demonstrate that (1) the LLM-generated BN is comparable to the expert-designed BN, and (2) DKT combined with the generated BN outperforms the standard DKT model, highlighting the potential of LLMs to enhance traditional knowledge tracing methods. This research contributes to the ongoing discussion on the integration of generative AI in intelligent tutoring systems and demonstrates its capability to automate the creation of knowledge-tracing models, reducing reliance on expert intervention and extensive datasets while improving predictive performance in adaptive learning environments.
Mots clé(s):
Dimension(s):
Théorie de l'activité:
Appartenance: