Didik Dwi Prasetya, Muhammad Zaki Wiryawan, Triyanna Widiyaningtyas, Nurul Rismayanti, Azlan Mohd Zain, Heni Vidia Sari
Concept maps are widely used to represent learners' conceptual understanding through interconnected propositions. However, assessing the semantic alignment between teacher and student concept maps remains challenging, especially for Indonesian-language contexts. This study proposes the use of IndoBERT, a transformer-based language model pretrained on 220M Indonesian tokens, to analyze semantic similarity between teacher and student propositions in open-ended concept maps. The research follows the CRISP-IM framework, encompassing stages of data collection, preprocessing, modeling, and evaluation. Using cosine similarity on IndoBERT embeddings, the system computes semantic overlap across 27 student maps and a teacher reference map. Semantic alignment labels were derived from expert judgment based on conceptual equivalence between student and teacher propositions. Experimental results demonstrate a high level of agreement with expert annotations, achieving accuracy of 96.3%, precision of 96.2%, recall of 100%, and F1-score of 98% at a similarity threshold of 0.7. These findings indicate that IndoBERT effectively captures semantic nuances in educational concept-map assessment. The study contributes to the growing body of work on AI-driven formative assessment, providing a scalable approach for evaluating conceptual understanding in Indonesian learning contexts. © 2025 IEEE.
State University of Malang, Department of Electrical Engineering and Informatics, Malang, Indonesia; Universiti Teknologi Malaysia, Faculty of Computing, Johor Bahru, Malaysia
