Abstract
Traditional virtual agents often evaluate their environment through a purely egoistic lens, focusing on personal goal attainment. This research introduces a computational model that integrates social norms and group identity into the emotional appraisal process. By utilizing a normative architecture and the concept of Cognitive Unit Strength, the model allows agents to experience complex social emotions such as shame and reproach. Experimental results from a simulated social scenario demonstrate that this model significantly improves the believability of virtual agents by aligning their emotional reactions with human social expectations.
Introduction: Beyond Egoistic Agents
In current gaming and simulation environments, virtual agents typically operate as "egoists." Their internal logic dictates that an event is "good" if it facilitates a personal goal and "bad" if it hinders one. However, human social interaction is far more nuanced, governed by a complex web of social norms and group dynamics. We do not merely react to goal obstruction; we respond to the violation of social obligations.
Crucially, our emotional response to a rule-breaker depends on our relationship with them. We might feel "reproach" toward a stranger who violates a norm, but "shame" when a close friend—a member of our in-group—does the same. This paper addresses the lack of formal representation for these social bonds in affective computing.
The Normative Architecture
The researchers proposed a two-part architecture designed to function as an agent's "moral compass." The first part is the Normative Component, which manages the detection and evaluation of social rules.
As shown in the architecture diagram, the system relies on several key modules:
- Norm and Role Libraries: These define the rules (e.g., "no smoking") and the specific roles to which they apply (e.g., "customer").
- Norm Instantiator: This module monitors the environment to determine which norms are currently active based on the context.
- Norm Evaluator: This tracks whether an active norm has been fulfilled or violated, passing this information to the emotional appraisal system.
Emotional Appraisal and Social Identity
The second part of the architecture is the Emotional Component, based on the classic OCC (Ortony, Clore, and Collins) model. The core innovation here is the introduction of Cognitive Unit Strength ($U$). This variable quantifies the social bond between the observer and the actor based on similarity (shared attributes) and affiliation (liking).
As illustrated in the logic table above, the model uses $U$ to differentiate between types of social emotions. When a norm is violated (negative praiseworthiness), the agent's reaction depends on whether the actor is part of their "cognitive unit" (in-group). If the bond is weak ($U < \alpha$), the agent feels reproach toward the other. If the bond is strong ($U > \alpha$), the agent internalizes the action and feels shame.
Experimental Evaluation: The Bar Scenario
To test the model, the researchers developed a "Bar Scenario" where a virtual agent observes a character smoking in a non-smoking area. The experiment manipulated two variables: Norm Salience (how visible the rule was) and Group Identity (whether the smoker was a friend or a stranger).
Data from 301 participants confirmed the model's efficacy. Participants perceived significantly higher levels of shame in agents when a friend violated the norm, whereas they perceived higher levels of reproach when a stranger was the offender. Furthermore, a strong positive correlation ($p < 0.001$) was found between the emotions participants expected to feel and the emotions the agents displayed, suggesting the model successfully mimics human social appraisal.
Conclusion and Future Work
By integrating social identity into emotional appraisal, this model allows virtual agents to react to the social world with greater realism. While the current model provides a robust framework for feeling social emotions, future work will focus on cultural variations in norm intensity and the implementation of "enforcement" behaviors, such as agents actively punishing or correcting norm-breakers.
Summary updated