Emotion Representations and Modelling for HCI Systems
Two invited talks will be held at ERM4HCI 2013 in Sydney, Australia, introducing key aspects of emotion representations and modelling for HCI systems as well as applications and implications. For further details regarding themes and aspects please refer to the speaker's talk abstract below (Scherer, Schuller).
Klaus R. Scherer
Director, Swiss Center for Affective Sciences, University of Geneva
Actuarial studies show that strong emotions such as fear, rage, or despair are quite rare in everyday life. Most of the time we experience rather subtle affective states such as astonishment, irritation, interest, regret, enjoyment, or amusement. Unfortunately these fleeting affects have been rarely studied in the affective sciences, partly because of the lack of appropriate models and partly because of the difficulty of measuring the respective response patterns. Yet, to fully realize the potential of affective computing and intelligent affective interaction in applied settings, it is essential to be able to detect and communicate subtle emotional messages of the type copiously used in everyday interactions. In this talk I will demonstrate that a component process model of emotion, based on high-resolution appraisal processes, constitutes a promising basis for the production and perception of these types of subtle emotion. Recent empirical evidence from our laboratory will be reported that show promising ways in which appraisals can be inferred from facial cues (as measured in natural settings) to generate inferences on subtle emotional states through a semantic rule structure.
Representation and Modelling of Human Emotion for Computers: Psychological Foundations vs. Machine Learning
Björn W. Schuller
Head of the Machine Intelligence and Signal Processing Group at the Institute for Human-Machine Communication, Technical University Munich
The representation and modelling of emotion in the context of human-computer interaction has led to an astonishingly rich variety of approaches. These reach from discrete classes with highly varying inventory of labels and “complex” (blended) emotions to “soft emotion profiles” similar to a tagging approach but attaching probabilities to each label in a closed inventory and continuous modelling based on emotion “primitives” such as arousal or valence. In this talk I want to highlight technical implications of such representations that are often rooted in psychological foundations for machines learning from (interaction) data. Particular challenges arise when “translating” between data annotated in different such schemes or when dealing with unlabelled or multimodal data where labels often have to be merged at different time levels. In particular, recent approaches in transfer learning, as well as active, semi-supervised, and unsupervised learning will be discussed in this context. In addition, the establishment of the emotion “gold standard” based on multiple raters and its impact on machine performance will be featured. This includes chances arising from the knowledge and exploitation of individual labeller tracks, e.g., for confidence level computation. Further, multi-task learning will be considered as a chance to exploit correlation across emotion primitives. In a concluding statement, practical recommendations will be given alongside avenues for future efforts.