Sentiment-aware NLP in reflective writing: toward fair and adaptive dashboards for Indonesian project-based learning

Reflective writing in project-based learning (PjBL) supports students‘ cognitive and affective development, yet instructors often cannot review all reflections, risking missed emotional cues. This study explores how sentiment-aware NLP can make students emotions visible at scale in Indonesian higher education through interactive dashboards. We analyzed 71 first-year informatics students′ reflections using an Indonesian-trained RoBERTa sentiment classifier. Sentiment distribution was 77.5% positive, 11.3% neutral, and 11.3% neutral, and 11.3% negative. Manual validation of 20% of texts confirmed high model reliability (86.7% accuracy, Cohen′s κ = 0.82). Statistical tests showed no significant differences in performance across sentiment groups, though male students more frequently expressed negative sentiment. Qualitative analysis revealed that frustation often co-occurred with persistence, indicating negative sentiment does not necessarily signal disengagement. Critically, findings reflect Indonesian cultural norms of harmony and indirect expression, highlighting risks of bias when dashboards rely on Western-trained models. The study demonstrates that sentiment-aware NLP can deliver meaningful affective insights in low-resource languages. Integrated into dashboards, this approach enables timely, fair, and culturally responsive instructor interventions in PjBL. Despite the modest dataset, the method is scalable and points toward adaptive learning environments that holistically support cognitive and affective dimensions of student learning.