Construction standards across countries provide technical guidelines to ensure the quality and safety of buildings and infrastructure, and they are periodically revised to reflect advancements in construction technology. However, due to their broad scope and interdependent nature, these standards often contain redundant or conflicting provisions, making the revision process complex and inefficient. Traditional expert-driven, manual approaches are time-consuming, costly, and prone to errors.
To address these challenges, we propose a large language model–based framework (COSLLM) for classifying redundancy and conflicts in construction standard sentences.
COSLLM leverages a construction domain–trained large language model to perform semantic comparisons between standard sentences. It employs a two-stage reasoning process with adaptive Chain-of-Thought (CoT) prompting to conduct in-depth analysis of sentence pairs that are likely to be redundant or conflicting. This approach ensures both high classification accuracy and computational efficiency.
The framework achieves an accuracy of 97.9% and a macro F1 score of 0.907 in classifying sentence pairs—extracted from Korean construction standards—into redundant, conflicting, or neutral categories. Furthermore, we developed a real-time web-based system based on COSLLM to support the efficient development and revision of construction standards.