Generative knowledge-guided review system for construction disclosure documents

Construction disclosure documents are crucial for the safe and orderly execution of construction projects, making their effective review indispensable. However, accurately retrieving and applying regulatory compliance information from unstructured repositories of construction knowledge remains a significant challenge in this review process. To address this issue, this paper proposes a generative knowledge-guided review system that incorporates a Dynamic Semantic Knowledge Chunking (DSKC) strategy, designed to enhance the semantic association of knowledge within construction text knowledge bases. Building upon this foundation, a Generative Knowledge-Guided Retrieval (GKGR) framework is introduced to improve the accuracy of knowledge retrieval during the review process, thereby enhancing the overall reliability of document review. Experiments conducted on four newly established datasets and benchmarks demonstrate substantial improvements of 21.5 % in MRR@3 and 10.9 % in F1-Score compared to the baseline method, and further outperform state-of-the-art retrieval techniques.