Recognition of Cognitive Intervention Effects and Modeling of Learning Behavior Patterns in Second Language Acquisition Using a Multimodal Intervention-Aware Temporal Graph Network

To distinguish between “text improvement” and “cognitive gain” in generative-AI-assisted second language learning, this study proposes a Multimodal Intervention-Aware Temporal Graph Network (MITG-Net). The method integrates human-AI dialogue, revision trajectories, behavioral logs, and learner attributes. Through intervention-conditioned dynamic weighting and heterogeneous temporal graph modeling, it captures the relationships among prompts, feedback, revisions, and transfer, and jointly performs intervention-effect recognition and learning-behavior-pattern discovery. Experiments were conducted on 744 learning sessions collected from 186 undergraduates at three universities. MITG-Net achieved a Macro-F1 of 0.816 and a weighted Kappa of 0.854, outperforming the strongest baseline, HGT, by 3.4 percentage points in Macro-F1. Removing behavioral logs and revision trajectories reduced Macro-F1 by 0.054 and 0.045, respectively. The model identified five behavior patterns. Among them, the strategy-transfer application pattern achieved an average effect level of 2.78 and a transfer score of 0.82, substantially higher than the 1.36 and 0.31 obtained by the AI-dominated direct-adoption pattern. These results demonstrate that the proposed method can effectively recognize cognitive intervention effects and provide interpretable evidence for personalized support in second language learning.