Representational bias in expression and annotation of emotions in audiovisual databases

Emotion recognition models can be confounded by representation bias, where populations of certain gender, age or ethnoracial characteristics are not sufficiently represented in the training data. This may result in erroneous predictions with consequences of personal relevance in sensitive contexts. We systematically examined 130 emotion (audio, visual and audio-visual) datasets and found that age and ethnoracial background are the most affected dimensions, while gender is largely balanced in emotion datasets. The observed disparities between age and ethnoracial groups are compounded by scarce and inconsistent reports of demographic information. Finally, we observed a lack of information about the annotators of emotion datasets, another potential source of bias.