Fairness in English classroom interaction is crucial for teaching quality and equity.Traditional assessment approaches are often limited.This study introduces a novel quantitative evaluation model using multimodal data distillation.It integrates heterogeneous data sources, constructs a joint representation space, and extracts key fairness indicators.Knowledge distillation transfers knowledge from a multimodal teacher model to a lightweight student model, achieving efficient compression and deployment.Validated on over 150 hours of real classroom data from 30 middle school English classes, the model achieves 92.3% accuracy in recognising teacher attention distribution.Its Gini coefficient error for student speaking opportunity is below 0.05, outperforming benchmarks.The compressed model retains only 40% of the original parameters, increases inference speed by 3.1 times, and maintains 94.7% core accuracy.
Zeng et al. (Thu,) studied this question.