Application and Effectiveness Evaluation of Deep Learning-Based Multimodal Sentiment Analysis in Ideological and Political Classrooms

In ideological and political classrooms, teachers often struggle to identify students’ emotional states accurately and adjust teaching strategies in time, which can reduce teaching effectiveness. This study proposes a deep-learning-based multimodal sentiment analysis method for classroom application. A multimodal acquisition system collects facial expression, speech intonation, and text semantic data from 320 students across 36 class sessions. A parallel neural architecture combining ResNet, BiLSTM, CNN-LSTM, BERT, and an attention-based fusion mechanism is constructed to extract and integrate visual, audio, and textual features. Transfer learning, data augmentation, and focal loss are used to improve model robustness, while an edge-computing architecture protects privacy by processing biometric data locally. Experimental results show that the proposed multimodal attention model achieves 92.7% accuracy, outperforming the best single-modal and bimodal models. In teaching application, classroom participation increases by 28.6%, knowledge acquisition by 21.4%, and satisfaction from 67.3% to 89.8%. The system supports intelligent teaching and is compatible with wireless classroom sensing and electromagnetic-safe data acquisition environments.