Recognising Emotions in Chatbot–Learner Dialogues in Educational Robotics

Emotions significantly influence learning processes, yet their detection and classification in authentic educational chatbot interactions remain underexplored, particularly for non-English contexts. This study investigates whether a BERT-based model using few-shot learning can reliably classify emotions according to Plutchik's eight-emotion framework in Italian educational robotics dialogues. We employed the Italian language model umberto-commoncrawl-cased-v1 with 120 manually created and expert-validated few-shot examples (15 per emotion category). Following proof-of-concept validation on MultiEmotions-IT (Top-1 accuracy: 54.6%, F1: 53.8%), we analyzed 958 authentic student-chatbot-tutor interactions from M5Stack/robotics programming contexts. Results revealed distinct role-based emotion patterns: student messages were dominated by fear (69.3%), indicating technical uncertainty and learning struggles, while tutor responses exhibited anticipation (65.3%), reflecting guidance-oriented scaffolding. Despite achieving 85.4% Top-3 accuracy on benchmarks, consistently low confidence scores (10-19%) in educational dialogues revealed a fundamental domain adaptation gap between social media-trained models and formal educational discourse. The findings align with Pekrun's Control-Value Theory and suggest practical implications for designing emotion-adaptive chatbots that respond to learner affective states. However, operational deployment remains constrained by confidence limitations, highlighting the need for domain-specific training data and multimodal emotion detection approaches.