Neuro-Symbolic Causal Chain Extraction from Industrial Maintenance Narratives with Hard Constraint Enforcement

Abstract

Extracting multi-hop causal chains from industrial maintenance narratives is essential for root cause analysis and predictive maintenance, yet existing information extraction methods struggle with long-range dependencies and frequently generate physically impossible causal transitions. We present a neuro-symbolic framework that integrates fine-tuned language models with symbolic constraint enforcement to extract structurally valid causal chains from unstructured maintenance text. Our approach introduces three novel components: (1) a domain-specific ontology grounded in ISO 14224 that formalizes eight node types and eleven forbidden causal transitions derived from engineering first principles; (2) a deterministic corpus generation methodology that produces fifty training narratives from twenty-five failure paths validated against published fault tree analysis standards; and (3) a constrained beam search algorithm that reranks neural model outputs using an Illegal Link Table to guarantee zero constraint violations. Experimental evaluation demonstrates that our method achieves 100% constraint compliance and 80% entity-level recall on held-out test cases, representing a 26 percentage point improvement over unconstrained generation baselines. We publicly release our ontology, corpus, and implementation to support reproducible research in industrial natural language processing.