Beyond LLMs A Linguistic Approach to Causal Graph Generation from Narrative Texts

We propose a novel framework to generate causal graphs from narrative texts, bridging the gap between high-level causality and finergrained event-specific relationships.Our approach first extracts concise, agent-centered "vertices" using an LLM-based summarization strategy.We then introduce an Expert Index-seven linguistically grounded features-and incorporate them into a STAC (Situation, Task, Action, Consequence) classification model.This hybrid system (RoBERTa embeddings + Expert Index) achieves superior precision in identifying causal links compared to LLM-only baselines.Finally, we apply a structured, five-iteration prompting process to refine and construct a connected causal graph.Experiments on 100 chapters and short stories show that our method consistently outperforms GPT-4o and Claude 3.5 across key dimensions of causal graph quality, while maintaining comparable readability.The resulting open-source tool offers an interpretable and efficient solution for capturing nuanced causal chains within narrative texts.