Emotion classification in educational AI systems requires reliable detection of domain-specificaffective states, particularly frustration, which is consistently associated with disengagement andreduced learning outcomes. Prior work introduced an IndoBERT-based classifier for Indonesianeducational chatbot interactions, but reported poor multiclass performance (accuracy 0.31, MacroF1 0.22), attributed to a small, imbalanced training corpus and the absence of efficient fine-tuningtechniques. This paper presents a technical follow-up addressing these limitations through twocontributions. First, we apply Low-Rank Adaptation (LoRA) fine-tuning to the baseline modelusing a large-scale public Indonesian emotion dataset (73,672 entries), with a theoreticallygrounded label remapping strategy informed by Lazarus's Cognitive Appraisal Theory and theFrustration-Aggression Hypothesis. Second, we propose an Appraisal-Based FrustrationComposite Scoring mechanism that aggregates multiclass emotion probabilities usingpsychologically motivated weights to produce a binary frustrated/non-frustrated classification atan empirically optimized threshold. The fine-tuned model achieves Macro F1 of 0.84 — a 3.8×improvement over baseline — with binary frustration detection reaching F1 0.86 and accuracy0.91. Error analysis reveals systematic confusion at the frustrated–angry boundary, consistent withtheoretical accounts of frustration escalation, and identifies domain mismatch as a secondarylimitation motivating future domain-specific fine-tuning. The model is publicly available athttps://huggingface.co/ZenyxS/indobert-emotion-v2-lora
M. Fabian Prasetyo (Tue,) studied this question.