SPAN: Structure-Preserving Analogical Narratives from Procedural Scientific Texts

Scientific explanations often encode complex entities, interactions, dependencies, and causal relationships that can be difficult for learners to internalize directly. Analogical narratives offer a promising way to make such knowledge more intuitive and memorable, yet large language models often generate stories that are fluent but structurally incomplete, overly literal, mechanically retold, or dependent on superficial metaphorical substitutions. We introduce SPAN (Structure-Preserving Analogical Narratives), a multi-stage framework for transforming scientific text into self-contained fictional narratives while preserving the underlying relational and causal structure. SPAN decomposes generation into five explicit stages: causally explicit scientific rewriting, causal and constraint-aware knowledge-graph construction, functional role and affordance grounding, coherent causal story-world synthesis, and causally audited narrative realization. The framework is designed so that important events in the target narrative emerge from established fictional properties, constraints, interactions, and state changes rather than from arbitrary narrative convenience. We evaluate SPAN against direct LLM generation, ParallelPARC-inspired relational generation, local-to-global analogy construction, role-mapping approaches, and structure-constrained prompting. Evaluation separates three complementary dimensions: generic analogical structure, science-aware structural fidelity, and science-blind narrative quality, using pairwise model judgments, benchmark-based validation, and human assessment. Stage-wise ablations further examine the contribution of each intermediate representation. By treating analogy generation as the construction of a coherent fictional world whose behavior mirrors the structure of the source explanation, SPAN aims to move beyond surface metaphorization toward more faithful, interpretable, and narratively effective scientific analogies.