Introduction
Analogy-based story generation (ASG) creates new narratives by mapping a source story (a known "case") onto a target situation. While promising, these systems are often brittle and prone to semantic "hallucinations"—for example, mapping a "bird" to "food" and then concluding that the food "died." Historically, researchers have added domain knowledge to these systems in an ad-hoc manner. This paper seeks to move toward a principled engineering approach by identifying exactly where domain knowledge lives within an ASG architecture and measuring the impact of strengthening specific areas.
The Knowledge Container Framework
The authors base their theoretical framework on Richter’s Knowledge Containers, originally designed for Case-Based Reasoning (CBR). They identify five primary containers where domain knowledge can be localized to improve the generation process:
- Vocabulary: The data structures and primitive notions (e.g., what constitutes a "character").
- Similarity Measure: How the system identifies which source story is relevant to the target.
- Story Repository (Case Base): The collection of pre-authored high-quality stories.
- Solution Transformation: The logic used to map elements from source to target (the "analogical transfer").
- Solution Evaluation: A final check to ensure the generated story is semantically coherent.
To test this framework, the researchers used Riu, an interactive narrative system utilizing the Structure Mapping Engine (SME). They systematically "upgraded" these containers using external knowledge sources like WordNet for semantic relations and Force Dynamics, a cognitive linguistics framework for representing physical and social forces between characters.
Key Insights: Representation and the Similarity Paradox
One of the paper's most significant findings is that the Vocabulary container is the most influential. By moving from simple labels to "Force Dynamics" annotations—which represent who is pushing, resisting, or yielding—the system found better analogical mappings in nearly 73% of cases.
Interestingly, the study revealed a "Similarity Paradox." While one might expect a sophisticated WordNet-based similarity measure to drastically improve results, it only marginally outperformed simple keyword matching. In story generation, a "less similar" source might actually lead to a more creative or surprising story, whereas in traditional problem-solving, it would simply be a "wrong" answer.
Results and Evaluation
- Vocabulary Impact: Force Dynamics (FD) led to superior mappings in 72.7% of the 55 generated story combinations compared to a baseline representation.
- Evaluation Success: The WordNet-enhanced evaluation routine improved output in 5.45% of cases by preventing specific semantic errors (like the "food died" example). While the percentage is low, it represents a highly cost-effective way to boost system reliability.
- Quantity vs. Quality: Increasing the number of stories in the repository significantly increased narrative variety, but did not necessarily improve the structural quality of the analogies themselves.
Conclusion
The research demonstrates that the location of knowledge—whether in the vocabulary, the similarity measure, or the evaluation phase—significantly dictates the quality and variety of the resulting narrative. Future work aims to refine evaluation metrics to better distinguish between "valid but boring" stories and those that are truly creative, and to explore how to author larger story repositories more efficiently, as content creation remains a major bottleneck for ASG systems.
Summary updated