Affective modelling in ERIC

Affect Generation in ERIC occurs by first using a set of cause-effect and belief relations to assign appraisals of emotion-eliciting conditions (EECs) to facts in the world model based on a set of goals and desires. These appraisals are used by ALMA [Gebhard, 2005] to generate an affective state (medium-term moods and emotion events) according to the OCC cognitive model of emotions [Ortony et al, 1988]. The generated affective state is then expressed by the ERIC agent in his lexical selection, his gestures, and his facial expressions; in addition, expression through his prosody is planned as future work.

Appraisal of Emotion-eliciting Conditions

ALMA generates emotions using appraisals of seven emotion-eliciting conditions: desirability, probability, realization and liking of events, praiseworthiness and agency of actions, and appealingness of objects. These appraisals are made by comparing events, actions and objects observed in the world against the agent's goals and desires.

If an event or action we desire occurs, this is appraised positively; conversely, if an event or action we desire to avoid occurs, this is appraised negatively. Also, when an object we like occurs in the discourse state (because it has just been mentioned in the commentary), it is appraised positively; conversely, when an object we dislike occurs in the discourse state, it is appraised negatively. Appraisals of agency and realisation are made by observing the world model. Agency is appraised by identifying the agent of each action. An event is appraised as realised if it has occurred in the world. As well as these basic appraisals, cause-effect relations allow us to appraise events, actions and objects that are not specified in the goals and desires.

Goals and Desires

To be able to compare events, actions and objects against goals and desires, we need to specify events or actions we would like to see occur (goals), events or actions we would not like to see occur (antigoals), and objects we like or dislike. Additionally, the agent maintains a set of beliefs about other actors' goals and desires, to enable him to judge the desirability of events for the players in the scenario.

The goals and desires are specified in a separate configuration file to the main affect module, since they are dependent on the world model and thus domain-specific. Thus they can easily be altered to match both a change in domain and a change in intended audience.

Cause-Effect Relations

Some of these appraisals are simply specified in the goals: for example, liking and appealingness of objects. Similarly, agency of actions can be easily observed from the world model when the actions are observed; and realisation of events is similarly observed from the world model. However, in order to reason about events and actions that are not directly part of our goals, but still somehow related to (or influencing) our goals, we need to codify the relations between non-goal events, actions and objects and goals. This is done by means of the four relations: leadsto, hinders, supports and contradicts. The first two relations model causality:

• a leadsto b if a increases the likelihood of b, and
• a hinders b if a decreases the likelihood of b.

The second two relations model logical deduction or belief:

• a supports b if knowledge of a supports belief in b, and
• a contradicts b if knowledge of a contradicts belief in b.

These relations are used to propagate appraisals of desirability, praiseworthiness, and likelihood in the following way:

• a leadsto b
  • a inherits some of b's desirability/praiseworthiness, and
  • the likelihood of b is increased when a occurs.
• a hinders b
  • a inherits the opposite of b's desirability/praiseworthiness, and
  • the likelihood of b is decreased when a occurs.
• a supports b
  • if a is observed (realization = true), then b also has realization = true
• a hinders b
  • if a is observed (realization = true), then b has realization = false

In these calculations, desirability and praiseworthiness are interchangeable: they appraise almost the same quality, except that desirability describes an event and praiseworthiness describes an action. Thus, if a is an event and b is an action, a's praiseworthiness will increase proportional to b's desirablity.

The generated emotions are depicted in the character's facial expression, and he may also comment on them in his discourse. His mood state is reflected in the gestures that accompany his speech and waiting, and also influences his lexical selection during natural language generation.