A Neuro-Symbolic AI Pipeline for Generating SHACL-Based Digital Building Regulations

Abstract: Automating regulatory compliance in the Architecture, Engineering, and Construction industry remains a challenge due to the semantic gap between natural language regulations and machine-interpretable code. Large Language Models (LLMs) are promising to bridge this gap, however, they often suffer from the lack of domain knowledge to correctly interpret and digitize a regulatory statement. This paper proposes a neuro-symbolic AI pipeline that utilizes GraphRAG (Retrieval-Augmented Generation) to ground LLMs in domain ontologies and SHACL templates. A multi-stage prompting technique is applied, that first maps regulatory terms to their equivalents in the ontology, then introduces First-Order-Logic as an intermediate representation to ensure isomorphic translation from text to code, and finally converts the text to a SHACL shape. The results demonstrate that this method reduces hallucinations and improves the precision of digital rule generation, contributing to a scalable methodology for transforming complex building regulations into machine-interpretable code.