Automated Sifting of Stories from Simulated Storyworlds

Wilkins Leong, Julie Porteous, John ThangarajahIJCAI 2022 · DOI

TL;DR

This paper introduces ARC SIFT, a novel system that allows users to intuitively sift through complex simulated storyworlds by drawing "story arcs" representing desired narrative patterns. By combining dynamic character networks with dynamic time warping, ARC SIFT automates the tedious process of story query creation, making it accessible to non-technical users and enabling more efficient exploration of emergent narratives.

What Problem Does This Paper Address?

Interactive narrative systems, from games to educational tools, often generate complex stories through the bottom-up interactions of virtual characters. While this emergent narrative approach offers rich storytelling potential, it creates a significant challenge: how to find specific, desired stories or patterns within the vast sea of generated events. Current methods for "story sifting" (identifying these stories) typically require users to write complex, technical search queries, posing a barrier for non-expert users. This paper addresses this gap by proposing an intuitive, visual method for specifying story criteria and an automated system to match these criteria to simulated narratives.

Methods

The core of ARC SIFT lies in two main components:

  • Dynamic Character Networks (DCNs): These networks track the evolving social affinity or "fortunes" between pairs of characters over time within a simulation. Interactions between characters are assigned positive or negative "consequence scores," which are aggregated into a matrix representing the character network at each time step. This sequence of networks forms a Dynamic Character Network.
  • Dynamic Time Warping (DTW): This algorithm is used to compare the similarity between two time-series sequences, even if they are not perfectly aligned in time. In ARC SIFT, DTW compares user-drawn "query story arcs" (representing desired narrative shapes) with the "simulation story arcs" generated from the DCNs. The system identifies the simulation story arc with the lowest DTW cost score, indicating the best match.

The system then returns the Abstract Interaction Log (AIL) events corresponding to the best-matching simulation story arc, effectively presenting the "sifted" story.

Datasets

The paper uses a simulated storyworld called "Prom Week" as its testbed. This scenario, based on a previously developed JavaScript library (AgentMaps), models students at a school preparing for prom. The simulation generates Abstract Interaction Logs (AIL) which are used as input. No new datasets were introduced; rather, the authors leverage a simulated environment to generate their own data.

Key Insights and Ideas

  • Visualizing Narrative Intent: The paper highlights the power of visual representations, specifically "story arcs," as an intuitive way for users to express complex narrative desires without needing to code. This democratizes story sifting.
  • Bridging Emergent Narrative and User Control: ARC SIFT offers a crucial bridge between the unpredictable nature of emergent narrative systems and the user's need to find specific, meaningful stories within them.
  • DTW as a Narrative Alignment Tool: The application of Dynamic Time Warping, commonly used in signal processing, to align narrative shapes is a novel and effective approach for matching user-defined story arcs to simulated events.
  • Character Fortunes as Narrative Indicators: The concept of tracking character "fortunes" (positive/negative consequences of interactions) provides a quantifiable and interpretable metric for defining narrative trajectories.

Key Results

A user study with 32 non-technical participants evaluated ARC SIFT's intuitiveness and accuracy.

  • Intuitiveness: In Task 1, where participants drew story arcs to match provided story summaries, a majority of participants (ranging from 50% to 84.4% depending on the story shape) drew arcs that accurately matched the original query arcs used by ARC SIFT.
  • Accuracy: In Task 2, where participants selected story summaries that best matched a given story arc, ARC SIFT's sifted story was the same as the most commonly chosen story by participants in all tested cases (ranging from 28.1% to 87.5% agreement).
  • Confidence: Participants reported generally high confidence in their responses for both tasks, with a median rating of 4 out of 5.

These results suggest that ARC SIFT is both an intuitive and accurate tool for story sifting.

Limitations and Future Work

The authors acknowledge several limitations and propose future directions:

  • Limited Scope: The current system primarily focuses on the relationship between two characters at a time. Future work aims to extend this to sift stories involving more than two characters.
  • Domain Specificity: While the Prom Week scenario is used, the authors plan to apply ARC SIFT to a wider range of domains and larger search spaces.
  • Granularity of Sifting: The current approach sifts entire stories. Future work could involve sifting based on granular segments of simulation story arcs.
  • User Interpretation: The dynamic character network's interpretation of events could be tailored to individual user preferences.
  • Output Modalities: While the current output is text-based, future work could explore mapping sifted stories to other media like 2D or 3D visualizations.

Summary updated

BibTeX
@inproceedings{leong2022automated,
  author = {Wilkins Leong and Julie Porteous and John Thangarajah},
  title = {{Automated Sifting of Stories from Simulated Storyworlds}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.24963/ijcai.2022/686},
}
Figure