Rosalind Picard coined the term affective computing in 1995, whose definition is the possibility of the computer to be capable of recognizing and expressing emotion. Recognizing emotion means the ability of the artificial intelligence (AI) system to detect emotions, whereas expressing emotion is the capacity of the AI system to synthesize emotion artificially. The computing community has settled down some approaches to modeling emotions, as shown in Figure 2.
In general, the applications of modeling emotions used in artificial intelligence systems are from one of those types shown in Figure 2 — Appraisal, Bio-inspired, Categorical and Dimensional.
The categorical emotion model also known as discrete model labels the emotion according to emotions contained in a finite set of emotions. However, considering the human diversity when discussing about a finite set of emotions, according to a study carried out by Andrew Ortony and Terence J. Turner (1990), there is no consensus on the set of emotions.
Ekman’s model is the most known categorical model of emotion, which initial proposal was made by Paul Ekman and Wallace V. Frisen in 1971. This study showed that emotions can be grouped into 6 big ratings (6 big emotions — anger, fear, disgust, surprise, happiness, sadness). Afterwards, Ekman and Karl G. Heider included another emotion category called contempt.
Another type of emotion model is the appraisal model which classifies emotion origins based on the cognitive model association. This model was proposed by Ortonny, Clore and Collins (OCC) in The Cognitive Structure of Emotion article and was named OCC model. There are 22 cognitive processes associated with 22 emotions. OCC is classified as an appraisal model and not as a categorical model because it maps a cognitive process.
When the emotion experience is associated with multidimensional space for emotion, there is the possibility of having infinite emotions classification. The dimensional model frequently cited in literature is Pleasure, Arousal and Dominance (PAD). This model, presented by Russell and Mehrabian in 1977, consists of three independent variables to classify emotional states: Pleasure is related to pleasure and unpleasure; Arousal is related to the energy level -activation and deactivation; and Dominance is related to dominance and submission. The Self-Assessment Manikin (SAM) scale, developed by Margaret M. Bradley and Peter J. Lang in 1994, is commonly used for measuring the PAD’s variables.
Another approach of emotion model is bio-inspired and consists of principals found in the real life systems, e.g., in nature and in biology. Examples of bio-inspired algorithms used in emotion-AI are the artificial neural networks, these include the use of CNN (convolution neural networks) by the visual cortex. The emotion-augmented machine learning is an emerging domain that is under the bio-inspired emotion model.