Neural Semantic Parsing of Building Regulations for Compliance Checking

Abstract

Computerising building regulations to allow reasoning is one of the main challenges in automated compliance checking in the built environment. While there has been a long history of translating regulations manually, in recent years, natural language processing (NLP) has been used to support or automate this task. While rule- and ontology-based information extraction and transformation approaches have achieved accurate translations for narrow domains and specific regulation types, machine learning (ML) promises increased scalability and adaptability to new regulation styles. Since ML usually requires many annotated examples as training data, we take advantage of the long history of building code computerisation and use a corpus of manually translated regulations to train a transformer-based encoder-decoder model. Given a relatively small corpus, the model learns to predict the logical structure and extracts entities and relations reasonably well. While the translation quality is not adequate to fully automate the process, the model shows the potential to serve as an auto-completion system and to identify manually translated regulations that need to be reviewed.

BRISE-Plandok: a German legal corpus of building regulations

“…The case study they describe involves applying the system to automated compliance checking of a single building, we do not have knowledge of any additional evaluation of the approach. Finally, another recent approach to the automatic parsing of building regulations is that of Fuchs, Witbrock, Dimyadi, and Amor (2022), who train an encoder-decoder model using the Transformer architecture (Vaswani et al, 2017) in an attempt to perform end-to-end parsing of natural language sentences into logic formulae. Their experiments are based on a dataset created by Dimyadi, Fernando, Davies, and Amor (2020) and containing a sample of the New Zealand Building Code (NZBC) mapped to the XML-based Legal-RuleML (LRML) format (Athan et al, 2013).…”