MultiSense-L2Net: AI-mediated co-regulation of emotion and motivation in second language writing

Language acquisition in second language (L2) writing requires learners to constantly regulate their emotions and motivation as they plan, draft, and revise text. Although previous research has established the salience of these affective processes, most investigations are based on retrospective self-reports that mask their temporal dynamics. This paper proposes MultiSense-L2Net, a novel AI-mediated multimodal architecture capable of capturing task-proximalemotion and motivation trajectories during L2 writing using physiological signals, keystroke dynamics, and reflective text. The framework fuses physiological, behavioral, and textual modalities with deep neural encoders that are interleaved with cross-modal attention and fed into a temporal graph neural network (TGNN) to model phase-aware affective- and motivational-transitions during planning, drafting, and revision. Transfer learning from publicly available affective datasets (WESAD and DEAP) allows for robust pre-training before adapting to an L2 writing targeted dataset. The experimental results show that MultiSense-L2Net can model task-related affective states (macro-F1 = 0.89) and predict intra-task motivational change (RMSE = 0.07) with high accuracy, outperforming unimodal and static fusion baselines. The framework was further validated through a pilot multimodal study with 10 L2-learners, demonstrating stable emotion classification and motivation modeling performance under subject-independent evaluation. Notably, the framework is framed as a process-analytic modeling technique rather than a psychometric indicator of latent traits. By providing an example of how multimodal AI can monitor emotion-motivation fluctuations with high temporal granularity, this work enables methodologically informed investigations of AI-supported, affect-aware writing tools and co-regulatory learning processes in the future.