Schema-Guided Culture-Aware Complex Event Simulation with Multi-Agent Role-Play

Abstract: Complex news events, such as natural disasters and socio-political conflicts, require swift responses from the government and society. Relying on historical events to project the future is insufficient as such events are sparse and do not cover all possible conditions and nuanced situations. Simulation of these complex events can help better prepare and reduce the negative impact. We develop a controllable complex news event simulator guided by both the event schema representing domain knowledge about the scenario and user-provided assumptions representing case-specific conditions. As event dynamics depend on the fine-grained social and cultural context, we further introduce a geo-diverse commonsense and cultural norm-aware knowledge enhancement component. To enhance the coherence of the simulation, apart from the global timeline of events, we take an agent-based approach to simulate the individual character states, plans, and actions. By incorporating the schema and cultural norms, our generated simulations achieve much higher coherence and appropriateness and are received favorably by participants from a humanitarian assistance organization.

• The paper introduces Miriam, a simulation framework that employs dual-layer agents and schema-guided role-play to simulate complex news events.
• It utilizes a graph-based event schema to structure simulations with temporal dependencies and logical conditions, enhancing overall realism.
• The framework dynamically integrates cultural norms to tailor simulations to geo-diverse socio-political contexts, improving crisis response accuracy.

Schema-Guided Culture-Aware Complex Event Simulation with Multi-Agent Role-Play

This paper introduces a sophisticated simulation framework, termed "Miriam," designed for the simulation of complex news events such as natural disasters and socio-political conflicts. The paper addresses a nuanced problem of simulating rare events where historical data is often insufficient. The proposed system leverages schemas to enhance domain knowledge and incorporates user-defined assumptions to contextualize specific scenarios. The authors significantly advance the field by embedding geo-diverse cultural norms into the simulation process, hence enhancing the realism and appropriateness of the generated simulations.

Key Contributions

1. Simulation Framework:

Miriam employs a dual-layered simulation approach that includes a global controller and character-based actors termed "agents." The global controller orchestrates event progression by setting timelines and actions, whereas character agents simulate individual behaviors based on predefined profiles. Each agent plans and reacts to events, guided by the event schema and context-specific cultural norms.

2. Event Schema:

The system's reliance on event schemas allows for the structuring of simulations in a way that embodies domain-specific knowledge. The schema is represented as a graph with events as nodes, enabling the system to simulate events with temporal dependencies and logical conditions (AND, OR, XOR).

3. Cultural Norms Integration:

A distinctive feature of the system is its ability to integrate cultural norms. The cultural enrichment process dynamically modifies simulations to reflect social behaviors pertinent to different geographic or ethnic contexts. This aspect of the simulation is crucial for generating realistic and culturally sensitive narratives.

Numerical Results and Claims

The system's efficacy is quantified through an automatic evaluation comparing variations of the simulation framework with and without schema guidance and cultural norms. The results reveal superior performance when both elements are integrated, with improvements in coherence, entailment, realism, and cultural appropriateness, as assessed by GPT-4o.

Implications and Future Work

The framework's practical implications are substantial, particularly in enhancing crisis response and training logistic models. By simulating events with fine-grained cultural and social understandings, policymakers can better prepare for potential crises. Theoretically, this work pushes the boundaries of how event schemas and cultural norms can be codified within AI systems, presenting new avenues for research in multi-agent systems and AI-driven societal simulations.

Future developments could extend to increasing the sophistication of agent interactions and introducing more detailed environmental modeling. Additionally, integrating real-time data streams for dynamic scenario updates could further enhance the realism and utility of the simulations.

In conclusion, Miriam stands out for its methodical integration of event schemas and cultural insights into a coherent simulation framework. This work lays a foundation for improved simulation capabilities in AI, offering significant advancements in both theoretical understanding and practical applications of complex event modeling.