Teacher emotional support, question openness and student oral participation: The role of positive emotion across stages in AI-observed STEM and non-STEM classes

Background Student oral participation is essential for engagement and learning, yet participation tends to decline in large-class and STEM contexts where social risk and cognitive demands are higher. Existing studies often relied on self-reported or small-scale observations, which may introduce subjective bias and are labor-intensive, limiting large-scale evidence on how instructional practices influence participation across contexts. Aims This study examines how teacher emotional support and question openness influence student oral participation through student emotion, and whether these effects vary across educational stages and disciplinary contexts. Sample Data were collected from 1455 face-to-face classes in China. Methods Artificial intelligence-based techniques were used to capture fine-grained classroom interactions. Structural equation modeling and multigroup analyses were conducted to examine the mediating role of student positive emotion and to compare pathway patterns across educational stages and disciplinary contexts. Results Teacher emotional support was primarily associated with student oral participation via student positive emotion, whereas question openness showed a more direct association with participation. The indirect role of emotional support through positive emotion became stronger across educational stages, suggesting a compensatory function in secondary education where baseline participation tends to decline. Students in STEM classes were less emotionally responsive to teacher support but showed a stronger association between positive emotions and oral participation. Conclusions Emotional support is a context-sensitive intervention rather than a universally effective strategy. By combining large-scale AI-based observation with cross-stage and disciplinary comparisons, this study contributes methodological advances for classroom interaction research, although cross-cultural validation of AI-based measures is needed.