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

2025-04-10 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

A novel framework is proposed for generating causal graphs directly from narrative texts, effectively bridging high-level causality with detailed event relationships. This method first employs large language model (LLM)-based summarization to extract concise, agent-centered vertices. It then integrates an "Expert Index," comprising seven linguistically informed features, into a Situation-Task-Action-Consequence (STAC) classification model. This hybrid system, combining RoBERTa embeddings with the Expert Index, achieves superior precision in identifying causal links compared to pure LLM-based approaches like GPT-4o and Claude 3.5. A structured five-iteration prompting process further refines and constructs connected causal graphs. Experiments on 100 narrative chapters and short stories confirm its enhanced quality and readability. The approach is available as an open-source tool.

Key takeaway

For NLP engineers and AI scientists developing systems for narrative understanding, this linguistic-LLM hybrid framework offers a more precise and interpretable method for causal graph generation. You should consider integrating linguistically informed features, such as the "Expert Index," and structured iterative prompting into your causal extraction pipelines. This approach demonstrably outperforms pure LLM solutions like GPT-4o and Claude 3.5, providing higher quality and more readable causal chains from complex texts.

Key insights

A hybrid linguistic-LLM framework significantly improves causal graph generation from narrative texts by integrating expert linguistic features.

Principles

• Linguistic features enhance LLM-based causal identification.
• Iterative prompting refines graph construction.
• Agent-centered vertices improve graph clarity.

Method

LLM-based summarization extracts agent-centered vertices. A STAC classification model, combining RoBERTa embeddings with a 7-feature "Expert Index," identifies causal links. A five-iteration prompting process refines the graph.

In practice

• Integrate linguistic features for causal link precision.
• Employ iterative prompting for graph refinement.
• Use agent-centered summarization for clear vertices.

Topics

• Causal Graph Generation
• Narrative Understanding
• Large Language Models
• Linguistic Features
• RoBERTa Embeddings
• STAC Classification

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Hybrid Human-LLM Corpus Construction and LLM Evaluation for the Caused-Motion Construction — LLMs struggle with complex linguistic constructions like CMC, necessitating novel corpus creation methods.
• Context Graphs: Free Association, Consistency, Narratives & Truth — LLMs function via free association on narrative graphs, trading consistency for broader, stochastically generated outputs.
• Conceptual Networks for Cross-Linguistic Idiomatic Expressions:A Feature-Based Graph Approach — Idioms organize by conceptual schema, not language, offering interpretable, cross-linguistically stable meaning representations.
• CascadeMind at SemEval-2026 Task 4: A Hybrid Neuro-Symbolic Cascade for Narrative Similarity — LLM vote agreement tracks difficulty, enabling confidence-aware routing for narrative similarity.
• Beyond Surface Forms: A Comprehensive, Mechanism-Oriented Taxonomy of Indirect Linguistic Encoding for LLM-Based Coded Language Detection — A mechanism-oriented taxonomy of indirect linguistic encoding significantly enhances LLM-based coded language detection.
• From Prompts to Tokens: Internalizing Causal Supervision in Vision-Language Model for Multi-Image Causal Reasoning — BridgeVLM internalizes visual causal reasoning in VLMs by converting causal graphs into executable tokens for improved control.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.