Using Social Relationships to Control Narrative Generation

Julie Porteous, Fred Charles, Marc CavazzaAAAI 2015 · DOI

Abstract

Interactive Storytelling (IS) requires mechanisms that allow authors to steer narrative generation toward dramatic and diverse outcomes. This paper introduces the NETWORKING system, which utilizes social relationships as a declarative control mechanism. By defining character ties through a visual interface, users can influence the trajectory of AI-driven stories in a medical drama setting. Experiments demonstrate that small changes in the social network lead to significant narrative shifts, providing an intuitive and efficient means of narrative control.

1. Introduction

Traditional AI planning typically seeks the most efficient path to a goal. In the context of narrative generation, however, "optimality" is secondary to dramatic tension and variety. Authors need tools to control the "shape" of a story—its conflicts, resolutions, and character arcs—without becoming mired in low-level logic. This research explores whether social relationships can serve as a high-level "control knob" for narrative diversity, particularly in genres like soap operas where interpersonal drama drives the plot.

2. The NETWORKING System

The NETWORKING system provides a three-tiered approach to narrative control:

  • Social Network Specification: Users interact with a visual interface to assign one of 12 relationship categories (e.g., professional-rival, secretly-dating, friend) to a cast of characters.
  • Narrative Generation Logic: The system employs a plan-based decomposition approach, using narrative situations as "pseudo-landmarks" to guide the planner toward specific dramatic beats.
  • Relationship-Driven Action Selection: Social ties function as filters for action selection. For example, an "extreme-antagonist" relationship enables actions like SPREAD-MALICIOUS-GOSSIP, while a "friend" relationship enables SHOW-APPRECIATION-WORKLOAD-SUPPORT.
Overview of the NETWORKING system: (1) specification of the social network via a UI, (2) narrative generation logic incorporating social preconditions, and (3) resulting narrative diversity based on initial social states.

3. Experimental Evaluation

The system was evaluated using a custom Medical Drama Domain Model consisting of 18 characters (doctors, nurses, and patients) and 100 narrative actions, which expand into approximately 25,000 ground instances. The authors measured "inter-narrative distance" using Levenshtein distance to quantify how changes in the social network affected story outcomes.

Results indicate that the system is highly sensitive to social configuration. Changing only four relationships resulted in a 60% difference in story content for 90% of generated narratives. Furthermore, the system maintains real-time performance, allowing for immediate visualization of narrative consequences in a 3D environment. This suggests that social networks are a highly efficient "low-bandwidth, high-impact" interface for narrative control.

4. Conclusion and Future Work

Social relationships provide a powerful and intuitive interface for controlling narrative generation. By shifting the focus from explicit story branching to initial social states, the NETWORKING system enables the creation of diverse, high-impact narratives. While currently focused on the medical drama genre, the methodology is generalizable. Future work will explore applying these techniques to other genres and conducting formal user studies to assess the system's intuitiveness for non-expert authors.

Summary updated

BibTeX
@inproceedings{porteous2015using,
  author = {Julie Porteous and Fred Charles and Marc Cavazza},
  title = {{Using Social Relationships to Control Narrative Generation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2015},
  doi = {10.1609/aaai.v29i1.9781},
}
Figure