Automating building code interpretation using NLP and transfer learning for enhanced code compliance

Building codes are essential for ensuring safety, quality, and regulatory compliance in construction projects. Currently, extracting information from regulatory texts is manual, time-consuming, and prone to errors, particularly because it involves interpreting natural language. This paper focuses on leveraging deep learning to train a model capable of predicting essential entities from regulatory texts, thereby enabling the interpretation and understanding of the natural language used in regulations. We fine-tuned a pre-trained BERT (Bidirectional Encoder Representations from Transformers) model to perform entity recognition for CODE-ACCORD dataset, which contains English and Finnish regulatory clauses annotated with BIO (Begin, Inside, Outside) tags across four essential categories: Object, Property, Quality, and Value. The trained model achieved high precision in entity recognition, effectively generalizing predictions to previously unseen regulatory clauses.. This paper contributes primarily by automating the semantic interpretation of regulatory texts, thereby establishing a foundation for more efficient and accurate automated code compliance checking processes.