Narrative Planning: Balancing Plot and Character

Mark O. Riedl, Robert Michael YoungJAIR 2010 · DOI

TL;DR

This paper introduces the Intent-Driven Partial Order Causal Link (IPOCL) planner, a novel algorithm for generating believable narratives by explicitly modeling character intentions alongside plot progression. It addresses the limitations of traditional planning systems by ensuring that character actions appear motivated and understandable to an audience, leading to more engaging and coherent stories.

What Problem Does This Paper Address?

Computational narrative generation, the automated creation of meaningful event sequences, faces a significant challenge in producing stories that are both logically coherent and characteristically believable. Traditional AI planning systems excel at generating sequences of actions to achieve a goal but often overlook the crucial aspect of character intentionality. This can result in narratives where characters act in ways that seem arbitrary or unmotivated to an audience, breaking the suspension of disbelief. The authors argue that for a narrative to be truly understandable and engaging, it must exhibit both a logical causal progression of plot and character believability, where characters are perceived as intentional agents. Existing narrative generation systems either rely on pre-scripted stories, limiting adaptability, or use simulation-based approaches that may struggle with global plot coherence. This paper aims to bridge this gap by developing a deliberative planning approach that explicitly reasons about character goals and motivations.

Methods

The core contribution of this paper is the Intent-Driven Partial Order Causal Link (IPOCL) planner. This algorithm extends traditional Partial Order Causal Link (POCL) planners by incorporating a richer plan representation and a novel search process. Key aspects of IPOCL include:

  • Expanded Plan Representation: IPOCL plans include "frames of commitment" which explicitly record a character's internal goals and the sequence of actions undertaken to achieve them.
  • Decoupling Author and Character Goals: Unlike standard planners, IPOCL distinguishes between the author's overall goal for the story and the individual goals of the characters within the narrative.
  • Intentionality Reasoning: The planner actively searches for plausible character intentions that explain actions, ensuring that characters appear to be intentional agents. This is achieved through three new types of flaws: open motivation flaws, intent flaws, and intentional threat flaws.
  • Least-Commitment Search: IPOCL employs an iterative, least-commitment search process to identify and repair these flaws, building a complete and believable narrative plan.
  • Opportunistic Discovery of Intentions: Character intentions are not pre-defined but are discovered opportunistically during the planning process, allowing for more dynamic and emergent narrative structures.

Datasets

The paper describes an evaluation using a generated narrative based on the "Aladdin" story. The specific domain and problem specification, including initial states, goal situations, and operator schemata, are detailed in Appendix A. This domain was used to generate narratives for comparison between the IPOCL planner and a conventional POCL planner. No external datasets were used for training.

Key Insights and Ideas

  • Character believability is paramount: The authors convincingly argue that for a narrative to be effective, characters must be perceived as intentional agents. This goes beyond simply achieving plot goals; it's about why characters act.
  • Planning can generate believable narratives: By extending planning paradigms to explicitly model character intentions, AI can move beyond generating mere sequences of events to crafting stories with psychological depth.
  • Decoupling author and character goals is crucial: Treating characters as independent agents with their own motivations, which may not always align with the author's ultimate goal, leads to more nuanced and realistic narratives.
  • Intentionality is not just an effect, but a process: IPOCL's focus on how intentions are formed and acted upon, rather than just stating an intention as an outcome, is key to its success.
  • The audience's perspective matters: The planner reasons about character intentionality from the audience's viewpoint, ensuring that the generated narrative supports human comprehension and inference.

Key Results

The paper presents an empirical evaluation comparing narratives generated by IPOCL and a conventional POCL planner. Participants were asked to read stories and answer "why" questions about character actions.

  • Hypothesis 1 Supported: Participants reading IPOCL-generated narratives showed a statistically significant tendency to rate "good" question-answer pairs more highly than those reading POCL-generated narratives. This suggests IPOCL narratives better support human understanding of character intentions.
  • Hypothesis 2 Supported: Participants reading IPOCL-generated narratives rated "poor" question-answer pairs significantly lower than those reading POCL-generated narratives. This indicates that IPOCL narratives are less likely to contain actions that appear unmotivated or confusing.
  • Improved Narrative Comprehension: The results strongly suggest that IPOCL-generated fabulas lead to better reader comprehension of character goals, intentions, and motivations compared to those generated by conventional planning algorithms.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Inability to Model Character Failure: IPOCL, like most planners, is designed to find successful plans. It cannot currently generate narratives where characters fail to achieve their goals, a common element in comedy and tragedy.
  • Limited Handling of Conflicting Intentions: While IPOCL can order conflicting frames of commitment, its mechanisms for detecting or preventing contradictory intentions are weak. Better heuristics and semantic analysis are needed.
  • Distinct Fabula and Sjuzet Reasoning: The current approach assumes fabula (what the story is about) and sjuzet (how it's told) can be reasoned about separately. For sophisticated effects like suspense, integrating these might be necessary.
  • Heuristic Development: The performance of IPOCL is heavily reliant on effective heuristic functions to guide the search. Developing more sophisticated domain-dependent and domain-independent heuristics is crucial for generating longer and more complex narratives efficiently.
  • Handling of Orphans: While IPOCL attempts to repair orphaned actions, cases that remain unrepaired require backtracking, which can be computationally expensive.
  • Narrative Length Control: The evaluation did not control for narrative length, which could potentially influence results. Future work should address this.

Summary updated

BibTeX
@article{riedl2010narrative,
  author = {Mark O. Riedl and Robert Michael Young},
  title = {{Narrative Planning: Balancing Plot and Character}},
  journal = {JAIR},
  year = {2010},
  doi = {10.1613/jair.2989},
}
Figure