Deep Learning–Based Named Entity Recognition and Resolution of Referential Ambiguities for Enhanced Information Extraction from Construction Safety Regulations.

Construction safety regulations and standards contain a massive number of fall protection requirements with respect to different equipment, facilities, and operations. Automated field compliance checking aims to detect field violations of construction safety regulations for improved compliance and safety. Recent research efforts focused on automated tracking of labor and equipment toward improved violation detection and safety compliance. However, extracting and modeling safety requirements for supporting automated violation detection or safety alert systems remains highly manual. Toward addressing this gap, information extraction provides an opportunity to automatically extract requirements from construction safety regulations for comparisons with field information to detect violations (or predict and prevent violations before they occur). However, existing information extraction methods are limited in terms of their scalability and/or accuracy. To address this need, this paper proposes a deep learning–based information extraction method for automatically extracting named entities describing fall protection requirements [e.g., scaffold, horizontal direction, or 1.82 m (6 ft)] from construction safety regulations and resolving referential ambiguities. The proposed information extraction method consists of three main submethods: (1) a deep learning–based method to recognize entities from the regulations, (2) a deep learning–based method to recognize referential ambiguities in the extracted entities, and (3) a named entity normalization method to resolve these ambiguities. The proposed method was implemented and tested on 20 selected Occupational Safety and Health Administration (OSHA) sections related to fall protection. An overall information extraction precision, recall, and F-1 measure of 93.2%, 89.6%, and 91.1% were obtained, which indicates good information extraction performance. [ABSTRACT FROM AUTHOR]