ACE: AI-assisted construction of educational knowledge graphs with prerequisite relations

Official URL: https://dx.doi.org/10.5281/zenodo.14250896

Knowledge graphs are effective tools for organizing information. In this work, we focus on a specialized type of Knowledge Graph called an Educational Knowledge Graph (EKG), with prerequisite relations forming paths that students can follow in their learning process. An EKG provides several features, including a comprehensive visual representation of the learning domain, and offers students alternative learning paths. The manual construction of EKGs is a time-consuming and labor-intensive task, requiring domain experts to evaluate each concept pair to identify prerequisite relations. To address this challenge, we propose a methodology that combines machine learning techniques and expert knowledge. We first introduce a prerequisite scoring mechanism for concept pairs based on semantic references captured through word embeddings. Concept pairs are then ranked with respect to their scores, and pairs with high scores are selected for expert evaluation, reducing the total number of pairs to be evaluated. The expert is iteratively presented with a concept pair, and an EKG is dynamically constructed in the background based on the expert’s label. As the graph evolves, some prerequisites can be inferred based on the existing ones, further reducing the expert’s task. We implemented our methodology in a web application, allowing experts to interact with the system and create their own graphs. Evaluations on real-life benchmark datasets show that our AI-assisted graph construction methodology forms accurate graphs and significantly reduces expert effort during the process. Further experiments conducted on a dataset from an educational platform demonstrate that students who study concept pairs in a prerequisite order determined by our methodology have a better overall success rate indicating that EKGs can improve learning outcomes in education. Interested readers can access additional material and the dataset at our Github repository.