Causal-Constrained GraphRAG: Enhancing Explainability and Reliability in Safety-Critical Parameter Extraction

ACL ARR 2026 January Submission8486 Authors

Keywords: Causal-Constrained GraphRAG, Causal NLP, Retrieval-Augmented Generation, Knowledge Graph Reasoning, Safety-Critical AI

Abstract: Causal-Constrained GraphRAG is a neurosymbolic framework designed to improve the explainability and reliability of parameter extraction from long, structured, and causally rich regulatory corpora used in safety-critical engineering domains such as fire safety and evacuation modeling. The framework parses heterogeneous technical manuals into a topologically grounded knowledge graph, extracts explicit cause-effect-action chains through weak supervision combined with an active-learning annotation loop, and injects validated causal structures into a GraphRAG reasoning stage to constrain large language model inference. Instantiation on a curated Evacuation and Fire Safety question--answer dataset (approximately 24k pairs) and a Neo4j-based causal knowledge graph enables evaluation across multi-hop regulatory reasoning, numerical and sign consistency, and end-to-end parameter generation tasks. Experimental results demonstrate substantial improvements in factual retrieval accuracy and causal coherence, including notable F1 gains in occupant-load interpretation and high mean opinion scores from domain experts, alongside improved response stability under repeated technical queries. The framework supports safe and auditable automation of engineering parameter extraction while explicitly preserving human-in-the-loop validation for cases involving ambiguous, incomplete, or conflicting regulatory guidance.

Paper Type: Long

Research Area: Retrieval-Augmented Language Models

Research Area Keywords: retrieval-augmented generation, neurosymbolic reasoning, interpretability, information extraction, safety and alignment

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Publicly available software and/or pre-trained models, Data analysis

Languages Studied: English

Submission Number: 8486