Natural affect data - Collection & annotation in a learning context

Automatic inference of affect relies on representative data. For viable applications of such technology the use of naturalistic over posed data has been increasingly emphasised. Creating a repository of naturalistic data is however a massively challenging task. We report results from a data collection exercise in one of the most significant application areas of affective computing, namely computer-based learning environments. The conceptual and methodological issues encountered during the process are discussed, and problems with labelling and annotation are identified. A comparison of the compiled database with some standard databases is also presented.