NAGNet: A Convolutional Neural Network for Real-Time Sentiment Analysis of Students

Abstract

Students' emotional state during class can reflect their understanding of classroom knowledge. Analysing the emotional data of students is helpful for ensuring that teachers understand students' learning effect and improve their teaching quality. With the rapid development of information education, student emotion analysis has been widely studied in the field of educational technology. However, due to the complexity and inconsistency of educational scenes, traditional facial expression recognition algorithms have low accuracy, losing information while extracting facial expression features, and student emotion analysis faces several technical issues. Therefore, based on the theory of deep learning, this paper proposes a novel network model NAGNet for recognizing and analysing students' expressions in classroom environments, to assist teachers in carrying out reasonable teaching effect evaluations. In this model, Res2Net50 is used as the backbone network to enhance the feature extraction ability of the model. Moreover, non-local attention module is added to integrate the global facial expression feature information to achieve fine-grained sentiment analysis. In addition, we replace the original pooling algorithm with generalized mean (GeM) pooling to pool the features of the last layer. The experimental results on the public affection dataset FERPlus, which includes eight emotions, show that the NAGNet model achieves an accuracy of 89.3%. Thus, the proposed model has higher accuracy than existing methods. In a real-world application scenario, the NAGNet model can analyse and detect students' emotions in real time. What's more, it has good robustness.