Story Planning with Vignettes: Toward Overcoming the Content Production Bottleneck

Mark O. Riedl, Neha SugandhIcids 2008 · DOI

TL;DR

This paper introduces the Vignette-Based Partial Order Causal Link (VB-POCL) planning algorithm, a novel approach to automated story generation that addresses the bottleneck of content production. By incorporating pre-defined "vignettes" (story fragments representing good situations), VB-POCL expands the search space of possible narratives, enabling the generation of stories with desirable structural properties that were previously difficult to achieve.

What Problem Does This Paper Address?

The paper tackles the growing challenge of content production outpacing consumption, particularly in entertainment, education, and training. In an "on-demand" society, the demand for new stories, quests, and scenarios is high, but human creators struggle to keep up. While automated story generation offers a solution, existing planning-based methods often lack the creativity to produce novel and valuable narratives. The authors aim to enhance the creativity of these systems by increasing the variety of narratives they can explore, thereby alleviating the production bottleneck.

Methods

The core contribution is the Vignette-Based Partial Order Causal Link (VB-POCL) algorithm. This is a modification of standard partial order planners. The key innovation is the integration of a vignette library. Vignettes are defined as "good" examples of specific story situations or contexts, stored as plan fragments.

When a standard planner encounters an "open condition flaw" (a precondition of an action that isn't met), VB-POCL introduces a third strategy: vignette reuse. This involves retrieving a vignette from the library that contains an action whose effect can satisfy the unsatisfied precondition. This process doesn't immediately fix the flaw but instead annotates the plan with a "fit flaw," indicating that the vignette needs to be integrated. The algorithm then iteratively "fits" the vignette into the plan, adding actions, causal links, and temporal constraints. This iterative fitting allows for interleaving with other flaw repairs and helps avoid redundant actions.

The paper also briefly touches upon "far transfer" for vignettes, a process to adapt vignettes from different domains to the target story's domain, though the detailed implementation is deferred to other work.

Datasets

Not discussed in the paper. The paper focuses on the algorithm and its theoretical underpinnings, not on empirical evaluation with specific datasets.

Key Insights and Ideas

  • Vignettes as "Good" Story Fragments: The idea of pre-defining and reusing "good" story fragments (vignettes) is a clever way to inject domain-specific narrative knowledge and aesthetic qualities into automated generation without requiring the system to "reinvent the wheel" for every situation.
  • Expanding the Narrative Search Space: By allowing the planner to incorporate these pre-existing, well-formed narrative chunks, VB-POCL can explore a much wider range of potential story structures, including those with actions that might not be strictly causally necessary for the final outcome but contribute to mimetic qualities or narrative richness.
  • Addressing the "Bottleneck": The core motivation – overcoming the production bottleneck – is a significant real-world problem. This approach offers a computational solution by increasing the efficiency and creativity of story generation systems.
  • Mimetic Requirements: The paper suggests that "dead-ends" (actions not directly contributing to the final outcome) might be important for mimetic requirements and reader engagement, a concept that traditional planners might overlook. VB-POCL's ability to explore such paths is a key differentiator.

Key Results

The paper does not present quantitative results from empirical evaluations. Instead, it focuses on the theoretical contribution of the VB-POCL algorithm and its potential to generate narratives with favorable structural properties that were difficult to achieve with previous planning-based methods. The example provided (a simplified Middle-earth story) illustrates how the algorithm can construct a plan incorporating a vignette.

Limitations and Future Work

  • Lack of Quantitative Evaluation: The paper acknowledges that an evaluation of VB-POCL has not yet been performed.
  • Heuristics for Vignette Selection: The paper notes that while VB-POCL can explore a larger space, it lacks understanding of how multiple vignettes interact or contribute to overall story quality. Future work needs to develop heuristics for controlling search space exploration and ranking solution "goodness."
  • Controlling Search Space: The authors suggest that future work requires strategies for controlling the search space exploration, including heuristics for ranking solution "goodness."
  • Vignette Transformation Complexity: While "far transfer" is mentioned, the computational complexity of transforming vignettes from arbitrary domains is noted as NP-complete, suggesting that offline processing is crucial.
  • Dead-End Recognition: The paper highlights the need for more research into creating intelligent heuristics to recognize when "dead-ends" are favorable, tolerable, or damaging from a narrative perspective.
  • Vignette Source and Context: The paper mentions that vignettes can come from various sources, potentially without abstract context information, which could pose challenges.
  • Interleaving Vignettes: The flexibility of interleaving multiple vignettes can lead to new issues and backtracking, indicating a need for better heuristics to discriminate between vignettes.

Summary updated

BibTeX
@inproceedings{riedl2008story,
  author = {Mark O. Riedl and Neha Sugandh},
  title = {{Story Planning with Vignettes: Toward Overcoming the Content Production Bottleneck}},
  booktitle = {Icids},
  year = {2008},
  doi = {10.1007/978-3-540-89454-4\_23},
}

Loading related content...

Figure