Natural Affect Data: Collection and Annotation

Abstract

Emotions are crucial for healthy cognitive functioning and have direct relevance to learning and achievement. Not surprisingly then, affective diagnoses constitute a significant aspect of expert human mentoring. Consequently, computer-based learning environments seek to incorporate the social dynamics of such human teacher–learner interactions in order to make learning with computers more engaging and effective. Advances in the field of affective computing have opened the possibility of studying emotions from their nonverbal manifestations and have motivated several efforts towards realising automatic affect inference. However, development and validation of affect recognition systems requires representative data to serve as the ground-truth. For viable applications of affect-sensitive technology, and to ensure ecological validity, the use of context-relevant, naturalistic data is preferred. This chapter reports results from the collection and subsequent annotation of data obtained in a learning scenario. The conceptual and methodological issues encountered during data collection are discussed, and problems with labelling and annotation are identified. It provides an integrated account of the complexity and challenges associated with emotion assessment in naturalistic situations.