LLM-based Structured Intermediate Representation for building regulation encoding

Abstract

Automating the encoding of building codes into formal representations is a critical step for compliance checking, yet it is often hindered by the linguistic complexity of regulations. This paper develops a two-stage pipeline that utilises Large Language Models (LLMs) to generate LegalRuleML (LRML) through a Structured Intermediate Representation (SIR). Using a dataset derived from the New Zealand Building Code (NZBC), the paper demonstrates that the SIR-based approach outperforms direct translation, with four fine-tuned models achieving F1 scores of 73.29%–77.23%, representing improvements of up to 2.40% over the current state-of-the-art method. Despite these gains, a large gap remains compared to expert-crafted SIRs (80%+). Analysis reveals that LLMs still struggle with logical structural integrity and that standard cross-entropy loss poorly aligns with formalisation goals. These findings highlight the potential of LLM-generated SIRs while underscoring the necessity of task classification and logic-driven verification in automated rule encoding systems.