<![CDATA[ Concept maps are effective learning tools for improving students’ deep understanding and critical thinking skills. However, manual assessment of concept map propositions is time-consuming, subjective, and difficult to perform consistently. Although transformer-based automated scoring approaches have shown strong performance, most still operate as black-box models, limiting transparency in educational contexts. This study proposes an ordinal DeBERTa-based automatic scoring model integrated with Explainable Artificial Intelligence for evaluating concept map propositions in database learning. A dataset of propositions scored on an ordinal scale of 0–3 by human raters was used as ground truth. DeBERTa, MiniDeBERTa, and DeBERTaV3 were trained using the CORAL ordinal classification framework. Model performance was evaluated using RMSE, Macro F1-score, and Quadratic Weighted Kappa. SHapley Additive Explanations and Local Interpretable Model-agnostic Explanations were applied to test data for token-level explanations, while faithfulness was evaluated using deletion–insertion AUC, comprehensiveness, and sufficiency. Experimental results showed that the ordinal DeBERTa model achieved an RMSE of 0.523, a Macro F1-score of 0.574, and a QWK of 0.814, indicating low prediction error and strong agreement with human raters. SHAP produced better comprehensiveness (0.228) and lower sufficiency (0.241) than LIME comprehensiveness (-0.145) and sufficiency (0.726), indicating more faithful and semantically consistent explanations. The findings also show that model predictions are influenced not only by individual tokens but also by contextual and relational proposition structures. However, this study is limited to token-level and local contextual explanations. Future work may explore semantically aware perturbations and global explanation analysis to better capture relational semantics and contextual dependencies. ]]>
Copyrights © 2026
