A Multimodal Myanmar Emotion Dataset for Emotion Recognition

Effective emotion recognition is vital for human interaction and has an impact on several fields such as psychology, social sciences, human-computer interaction, and emotional artificial intelligence. This study centers on the innovative contribution of a novel Myanmar emotion dataset to enhance emotion recognition technology in diverse cultural contexts. Our unique dataset is derived from a carefully designed emotion elicitation paradigm, using 15 video clips per session for three emotions (positive, neutral, and negative), with five clips per emotion. We collected electroencephalogram (EEG) signals and eye-tracking data from 20 subjects, and each subject took three sessions spaced over several days. Notably, all video clips used in experiments have been well rated by Myanmar citizens through the Self-Assessment Manikin scale. We validated the proposed dataset's uniqueness using three baseline unimodal classification methods, alongside two traditional multimodal approaches and a deep multimodal approach (DCCA-AM) under subject-dependent and subject-independent settings. Unimodal classification achieved accuracies ranging from 62.57% to 77.05%, while multimodal fusion techniques achieved accuracies ranging from 75.43% to 87.91%. These results underscore the effectiveness of the models, and highlighting the value of our unique dataset for cross-cultural applications.