Accurate Extraction of Construction Safety Requirements Tailored to Project Characteristics: Integrating Multivariate NLP Techniques

作者：Zhijiang Wu, Jianyao Jia, Liang Xiao, Yuanyuan Zhang, Guofeng Ma · 发表于：Journal of Construction Engineering and Management · 年份：2025 · DOI：10.1061/jcemd4.coeng-16265 · 被引用次数：4 · 研究领域：Occupational Health and Safety Research、Risk and Safety Analysis、Safety Warnings and Signage

Construction safety requirements (SRs) that document a wealth of safety information (e.g., codes, standards, preferences) are an important basis for project managers to develop safety strategies. The extensive volume of SRs embedded within project documents presents a significant challenge for information extraction, as they often lack explicit association with specific projects, and ambiguous expressions also introduce bias in the extraction process. To address this limitation, this study proposes a two-stage integration framework for adaptive recommendation and information extraction of SRs, wherein the first stage recommends appropriate requirement types for specific projects, and the second stage extracts information content by incorporating element characteristics. This framework introduces a latent Dirichlet allocation topic model for clustering safety targets and designs a target-type (TT) association model to select appropriate requirement types. Meanwhile, the multivariate technique in natural language processing, integrating element characteristics, is employed to extract element information and incorporate it into SRs. The result shows that the TT correlation model effectively recommends four appropriate requirement types for civil and industrial buildings. The improved term frequency-inverse document frequency algorithm achieves nearly a 30% improvement in precision and recall rates for object element extraction compared to traditional methods. Additionally, the m...