New method maps causal chains in stories better than GPT-4o or Claude

Researchers have developed a hybrid approach to automatically extract causal relationships from narrative texts that outperforms leading large language models. The system combines linguistic analysis with machine learning to build visual maps showing how events in stories cause and connect to one another.

The Challenge#

Understanding causality in stories requires more than general language understanding. Traditional LLM-based approaches often struggle to precisely identify which events actually cause other events, versus merely describing what happens next. The researchers sought a more interpretable method that could reliably distinguish causal links from simple narrative sequences.

How the System Works#

The framework operates in stages. First, it uses LLM-based summarization to extract core, agent-centered events as vertices in the graph. The key innovation is an "Expert Index"—a set of seven linguistic features drawn from grammatical and syntactic patterns that reliably signal causality. These features feed into a classification model called STAC (Situation-Task-Action-Consequence), built on RoBERTa embeddings, which decides whether two events share a causal link. Finally, a five-step prompting process iteratively refines and connects the resulting causal graphs.

Results and Impact#

Testing on 100 narrative chapters and short stories showed the hybrid system consistently achieved higher precision in identifying causal relationships than pure LLM approaches, including GPT-4o and Claude 3.5. The method remains readable and interpretable—users can understand why the system identifies certain relationships—while maintaining efficiency. The team has released the tool as open source, offering researchers and practitioners a more reliable way to extract causal structure from unstructured narrative data.