AI-assisted construction of educational knowledge graphs

Knowledge graphs are effective tools for organizing information. In this dissertation,we introduce a specialized graph called the Educational Knowledge Graph (EKG).This graph visualizes the concepts within a domain by circles and indicates theirprerequisite relations with arrows. Such a visualization provides a comprehensiverepresentation of the learning domain and shows students the appropriate order inwhich to learn these concepts. EKGs can be further enriched by textual informationdefining what each concept means and why it forms a prerequisite relation withthe other concepts in the graph. Manual construction of EKGs is a challengingand time consuming task for three main reasons. First, it requires assigning precisedefinitions to each concept within the domain. Second, the domain experts needto evaluate each concept pair for identifying possible prerequisite relations. Third,the identified prerequisite relations must be justified. To address the first two challenges,we propose a methodology that combines machine learning techniques withexpert knowledge. Given a domain name, our approach automatically generatesspecific descriptions for a predetermined number of concepts related to that domainand then assigns a prerequisite probability score to each concept pair using the generateddescriptions. The high scored pairs are then asked to a human expert forvalidation in an iterative manner. With each round of expert feedback, the EKG inthe background is updated dynamically and the final EKG is constructed once theexpert decides to finish the interaction. In order to address the third challenge, we describe a fine-tuning procedure for Large Language Models (LLMs) to teach themto identify and explain the prerequisite relations between the concepts. The resultsshow that LLMs, fine-tuned according to our described procedure are effective modelsfor prerequisite detection and can generate satisfactory explanations when askedto clarify the reasoning behind these relations. Finally, we present all the describedmethodologies in this dissertation in a web application. By using this application, weprovide instructors with the ability to create their own AI-assisted EKGs and offerthem to their students during their courses as supplementary learning materials.