CODE-ACCORD: A Corpus of building regulatory data for rule generation towards automatic compliance checking

Area of Science:

• Computer Science
• Artificial Intelligence
• Natural Language Processing

Background:

• Automatic Compliance Checking (ACC) in Architecture, Engineering, and Construction (AEC) requires automated interpretation of building regulations.
• Challenges include natural language complexity and limited resources for Machine Learning (ML) to convert textual rules into machine-readable formats.

Purpose of the Study:

• To address the challenges in automating building regulation interpretation for ACC.
• To introduce CODE-ACCORD, a manually annotated dataset for machine-readable rule generation.

Main Methods:

• Collected 862 self-contained sentences from English and Finnish building regulations.
• Manually annotated entities and relations by 12 annotators, followed by curation.
• Developed a dataset with 4,297 entities and 4,329 relations as ground truth.

Main Results:

• The CODE-ACCORD dataset provides a robust ground truth for machine learning tasks.
• The dataset contains 4,297 entities and 4,329 relations across various categories.
• Facilitates ML and Natural Language Processing (NLP) applications like text classification, entity recognition, and relation extraction.

Conclusions:

• CODE-ACCORD enables the application of advanced ML techniques, including deep neural networks and large language models, to ACC.
• The dataset is crucial for advancing research and development in automated compliance checking within the AEC industry.