Human Rating of Emotional Expressions - Scales vs. Preferences

A bstract: Human ratings of emotional expressions are the foundation for building and training automatic affect recognition systems. We compare two rating schemes for labeling emotional expressions: likert scales and pair-wise preferences. A statistical analysis shows that while there is a strong correlation between the two schemes, there are also frequent mismatches. Our findings indicate that the schemes perform differently well per affect label. We discuss reasons for this and outline planned future work based on the findings.