Exploring the Neural Basis and Validity of Ordinal Emotion Representation Through EEG

Accurately measuring emotion is a major challenge in advancing the understanding of human emotion and developing emotional artificial intelligence. In many existing studies, participants’ emotional ratings in interval scales are considered the true reflection of their emotional experiences. However, recent research suggests that ordinal annotations of emotions can more accurately capture the emotional expression process, providing a potential method for more precise emotion measurement. However, our understanding of the characteristics and validity of this new form of emotion representation is still relatively lacking. In particular, there is a lack of research using neural signals to explore the validity and neural basis of ordinal emotion representation. In this study, we used a video-elicited electroencephalogram (EEG) dataset (n = 123) to identify the neural basis of ordinal emotion representation and demonstrate its validity from a neural perspective. Furthermore, we explored various characteristics of ordinal emotion representation, showing how it is superior to the interval form. First, we conducted inter-situation representational similarity analysis (RSA) and inter-subject RSA to test the degree to which ordinal representation captures both group commonalities and individual differences of emotion. Next, we investigated the characteristics of ordinal representation under different combinations of emotion items, including uni-variate and multivariate emotions, positive and negative emotions. Our results show that both group commonalities and inter-subject variations in EEG features are better explained by ordinal emotion representations than by interval ones. Multivariate ordinal representations showed better inter-subject reliability and higher representational similarity with EEG features compared to uni-variate counterparts, highlighting the co-occurrence nature of human emotions. Compared to negative emotions, ordinal representation showed greater improvements for positive emotions, suggesting that the complexity of positive emotions is well captured by ordinal representations. Taken together, these findings demonstrate that multivariate ordinal emotion ratings provide a more accurate measure of real emotional experience, which is crucial for enabling machines to precisely understand and express human emotions.