Ontology-Guided Pedagogically Meaningful Knowledge Component Extraction from Code

Abstract

Accurate extraction of Knowledge Components (KCs) is critical for fine-grained learner modeling in programming education. Yet existing approaches remain limited: manual Q-matrices ignore solution variability; Abstract Syntax Tree (AST) based methods may not produce pedagogically meaningful KCs; and Large Language Model (LLM)-based methods lack a standardized KC taxonomy, leading to inconsistent concept names across problems. Hybrid approaches that combine expert-defined ontologies with automated KC extraction remain underexplored, particularly for modeling pedagogically meaningful concepts and their relationships. We propose a unified ontology-guided framework for automatic KC extraction that integrates pedagogical grounding, relation-aware learning, and multi-dimensional evaluation. The framework comprises a programming ontology of 61 KCs aligned with ACM/IEEE curricular guidelines and expert-validated; a Relational Graph Convolutional Network (R-GCN) that learns relation-specific propagation of AST-derived seed signals over multi-relational ontology edges; and a three-axis evaluation framework spanning KC extraction performance, cognitive validity of extracted KCs, and AFM-based predictive validity. On the CSEDM dataset, R-GCN demonstrates consistent improvements in KC extraction (Jaccard = 0.905; Hamming loss = 0.037) and yields approximately 58% "Good" KCs in the learning-curve analysis, compared with 11-26% for baselines. These results indicate that integrating pedagogical structure with relation-aware learning leads to more accurate, interpretable, and educationally effective KC extraction.