Comparing Valence-Arousal and Positive-Negative Affect Models of Affect: A Nonlinear Analysis of Continuously Annotated Emotion Ratings

The Affective Computing community commonly uses dimensional models of emotions to rate conscious emotional perceptions in emotion elicitation tasks. Although several structures of affect have been introduced in the literature, the valence and arousal dimensions have had the most impact. In this study, we compared Watson and Tellegen’s Positive-Negative Affect model to Russell’s Valence-Arousal plane. We used the publicly available Continuously Annotated Signals of Emotions (CASE) dataset, which provides ratings along the valence and arousal dimensions continuously annotated while watching video clips eliciting four emotions: scariness, amusement, relaxation, and boredom. We derived the Positive and Negative Affect time series from the valence and arousal time series through a 45° rotation of Russell’s plane. We calculated the median values and Fuzzy Entropy for each time series and video clip to investigate their linear and nonlinear dynamics. Our analysis showed that Watson and Tellegen’s model had fewer statistically significant differences between emotions than Russell’s model when considering the median values. However, when investigating the dynamic evolution of perceptions, the Positive Affect dimension showed the highest discriminative power, identifying the time series traced during the boring stimuli as the most regular and statistically different from all others. Our findings suggest that further acquisitions of continuously annotated ratings in several experimental settings, and the investigation of the nonlinear coupling between more dimensions, could significantly improve real-time emotion recognition.