Article Preview
TopIntroduction
Emotion and affect are intrinsic to our cognition, and any attempt at a detailed understanding of the human mind will require attention to this domain (Sloman, 1981; Lisetti & Gmytrasiewicz, 2002). While models of emotion have immediate applications, for instance in human computer interaction and user modeling, their main significance might lie deeper: the question of how it is possible that a mind, a biologically implemented information processing machine, is able to feel, to undergo emotional episodes, to turn into a self-reflecting and social agent has remained a dazzling issue to the philosophy of mind and cognitive science in general, and an adequate functional model of emotions will be an important part of the answer.
Computational modeling of emotion and affect has seen a wide variety of different approaches, which I will not review here. (Authoritative summaries on the state of the art may be found elsewhere, for instance in Gratch, Marsella, & Petta, 2011; for a look at its history consult Hudlicka & Fellous, 1996; Gratch & Marsella, 2005.) The nature of these approaches has been largely determined by applications, for instance for behavior modeling, for supporting communication with artificial systems, and for social simulations. Such applications favor externalist, descriptive models of emotional agents, for instance in belief/desire/intention frameworks (for the original notion of BDI, see Bratman, 1987). Conversely, if the goal is an understanding of cognitive behaviors, self-assessment of agents, the mechanisms of filtering and biasing in memory access, perception and action control, and the relationship between emotion and motivation, we require internalist, functional models. Such models will not treat the system as a black box, but expose its internal processing.
Externalist models arguably dominate today’s research in synthetic emotions, with a focus on the very successful family of appraisal theories of emotion (Roseman, 1991; Lazarus, 1991; Ellsworth & Scherer, 2003). Appraisals reflect assessments of external and internal stimuli of an agent, and they give rise behavioral and dispositional consequences. The intensity and range of affects and emotions is subject to individual variance (Russel, 1995), and their directedness is the result of adaptive learning, but the dimensionality, general expression and cognitive structure of emotions is largely invariant (Ekman & Friesen, 1971; Izard, 1994). For example, while a person might learn in what situations fear is appropriate or inappropriate, the ability to perceive fear itself is not acquired, rather, it stems from the way its organism is equipped to react to certain external or internal stimuli. Thus, it makes sense to develop general taxonomies of emotional states. The well-known Ortony-Clore-Collins model (OCC) (Ortony, Clore, & Collins, 1988) represents a high-level classification of these assessments: It treats emotions as valenced reactions to the consequences of events, to the actions of agents, or to aspects of objects, by distinguishing whether those situations and actions are desirable or undesirable, happen to oneself or another agent, are manifest or projected and so on (Figure 1). The OCC model elegantly captures the difference between social emotions (the appraisal of actions for oneself and others) and event-based emotions like hope or relief. Even though it is not exhaustive (in its original form, it does not account for all high-level emotions like jealousy or envy), it scales easily by adding additional appraisal conditions.
Figure 1. Taxonomy of higher-level emotions (adapted from Ortony, Clore, & Collins, 1988, p. 19)