Construction of multimodal perception and quantitative assessment model of interactive participation in English classroom

AI enhances classroom behaviour analysis, yet participation assessment remains subjective, single-dimensional, and lacks real-time quantitative support.This study proposes a multimodal framework integrating video, audio, eye tracking, and positioning to capture facial expressions, body posture, speech emotion, and semantic content.A dual-stream ResNet, TCN, and pretrained transformer are employed to extract visual, acoustic, and textual features, with cross-modal alignment achieved via timestamp synchronisation and learnable positional encoding.A multi-head self-attention-based fusion module and multi-task evaluation head quantify participation frequency, interaction depth, emotional engagement, and knowledge feedback.Experiments on 192 students across 90 classes achieve 91.2% accuracy, over 90% F1-score, and over 93% consistency with teacher ratings, significantly outperforming baseline methods.