Emotion dual-space network based on common and discriminative features for multimodal teacher emotion recognition

This study addresses the challenges in teacher emotion recognition (TER), namely the lack of high-quality multimodal datasets and insufficient modeling of common and discriminative emotional features across modalities. To this end, the authors construct a novel multimodal TER dataset comprising 102 real classroom lessons and 2,170 video segments spanning multiple educational stages and subjects.

The dataset is uniquely annotated with interaction-oriented emotion labels—happiness, neutrality, satisfaction, and questioning—that reflect teacher–student dynamics, while excluding negative emotions inconsistent with professional teaching contexts. To effectively model multimodal emotional expression, the paper proposes the Emotion Dual-Space Network (EDSN), which integrates an Emotion Commonality Space Construction (ECSC) module to capture shared emotional information across modalities using central moment differences, and an Emotion Discrimination Space Construction (EDSC) module that employs a gradient reversal layer and orthogonal projection to extract modality-specific emotional features and eliminate redundancy.

Extensive experiments demonstrate that EDSN achieves state-of-the-art performance on the proposed TER dataset with an accuracy of 77.0% and a weighted F1-score of 0.769, significantly outperforming baseline models. Furthermore, the model shows strong generalization on public benchmarks (CMU-MOSI, IEMOCAP), validating its robustness. This work makes significant contributions by providing a realistic, multimodal educational dataset and introducing a novel dual-space framework that effectively balances emotion commonality and discrimination in multimodal fusion.

The work titled “Emotion Dual-Space Network Based on Common and Discriminative Features for Multimodal Teacher Emotion Recognition”, was published on Frontiers of Digital Education (published on July 4, 2025).

Reference: 

Ting Cai, Shengsong Wang, Jing Wang, Yu Xiong, Long Liu. Emotion Dual-Space Network Based on Common and Discriminative Features for Multimodal Teacher Emotion Recognition. Frontiers of Digital Education, 2025, 2(3): 25 

https://doi.org/10.1007/s44366-025-0063-x