Competency Questions as Executable Plans: a Controlled RAG Architecture for Cultural Heritage Storytelling
Summary: arXiv:2604.02545v1 Announce Type: new
Abstract
The preservation of intangible cultural heritage is a critical challenge as collective memory fades over time. While Large Language Models (LLMs) offer a promising avenue for generating engaging narratives, their propensity for factual inaccuracies or “hallucinations” makes them unreliable for heritage applications where veracity is a central requirement. To address this, we propose a novel neuro-symbolic architecture grounded in Knowledge Graphs (KGs) that establishes a transparent “plan-retrieve-generate” workflow for story generation.
Introduction
As societies evolve, the intangible aspects of cultural heritage, such as stories, traditions, and collective memories, often diminish. This erosion necessitates innovative approaches to preserve and disseminate cultural narratives accurately. Large Language Models (LLMs) have emerged as powerful tools for generating narrative content; however, their tendency to produce factually incorrect information poses significant challenges. In response, we explore a new neuro-symbolic architecture that leverages Knowledge Graphs (KGs) to ensure narrative accuracy and coherence.
Methodology
Our approach introduces a “plan-retrieve-generate” workflow that integrates competency questions (CQs) as executable narrative plans. This novel repurposing of CQs bridges the gap between high-level user personas and atomic knowledge retrieval. The architecture is designed to be evidence-closed and fully auditable, ensuring that generated narratives remain grounded in factual data.
Resource Utilization
We validate our architecture using a new resource: the Live Aid KG, a multimodal dataset that aligns 1985 concert data with the Music Meta Ontology while linking to external multimedia assets. This resource provides a rich foundation for testing our narrative generation strategies.
Comparative Evaluation
We conducted a systematic comparative evaluation of three distinct Retrieval-Augmented Generation (RAG) strategies over the constructed graph:
- KG-RAG: A purely symbolic retrieval approach focusing on factual precision.
- Hybrid-RAG: A text-enriched method that balances context and factual accuracy.
- Graph-RAG: A structure-aware strategy that emphasizes narrative coherence through graph traversal.
Findings
Our experiments reveal a quantifiable trade-off among the three strategies. The KG-RAG method excels in factual precision, while the Hybrid-RAG offers greater contextual richness. The Graph-RAG approach stands out for its narrative coherence, demonstrating the importance of structure in storytelling. These findings provide actionable insights for the design of personalized and controllable storytelling systems suited to cultural heritage applications.
Conclusion
Our proposed neuro-symbolic architecture represents a significant advancement in the field of narrative generation for cultural heritage storytelling. By transforming competency questions into executable plans, we ensure that narratives generated are both engaging and factually accurate. As we continue to refine these methodologies, we aim to enhance the preservation and dissemination of cultural heritage, ensuring that collective memories endure for future generations.