GNR-Net: An Explainable Generative–Graph Neural Framework for Narrative Representation and Valuation

The increasing volume of creative narratives, such as screenplays and scripts, presents both opportunities and challenges for computational understanding and valuation. Traditional expert-driven assessments are subjective and limited in scalability, while existing text-based models struggle to capture the causal, emotional, and structural dimensions of storytelling. To address these challenges, this study proposes GNR-Net, an explainable Generative AI (GenAI) and Graph Neural Networks (GNNs) hybrid framework that integrates long-context large language modeling (LLM) with Knowledge Graph (KG) reasoning for interpretable narrative understanding and valuation. The framework comprises three components: a semantic graph construction module that transforms narrative texts into multi-relational KG structures; a generative narrative encoder based on LLaMA-3 that models long-range coherence and emotional causality and regularizes Narrative Coherence Score (NCS); and a narrative value estimator that combines relational graph convolution (R-GCN) reasoning with multi-head regression and adversarial calibration to quantify creativity, coherence, and market potential. Experiments on four screenplay corpora (IMSDb, ScriptBase, Cornell Movie Dialogs, and WGA Unproduced) demonstrate that GNR-Net outperforms nine baselines, achieving up to 13.2% MAE reduction and 0.82 Pearson correlation with expert ratings. Ablation and visualization analyses further show that jointly leveraging generative semantics and graph reasoning yields more faithful explanations linking emotional arcs to valuation dimensions. The proposed framework provides a practical pathway toward explainable computational creativity by bridging GenAI, GNN-based KG reasoning, and cultural data intelligence.