Information Integration of Regulation Texts and Tables for Automated Construction Safety Knowledge Mapping

The explicit safety knowledge contained in regulations in the form of texts and tables is crucial for construction safety management. However, the presence of rich semantic content within texts and the intricate layout of complex tables makes domain information extraction challenging. Therefore, this research proposed a hybrid approach to map safety knowledge graphs by automatically extracting information from both texts and tables in a scenario-oriented manner, combining rules and deep learning methods to achieve a balance between scene applicability and method flexibility. Furthermore, metrics from social network analysis (SNA) were applied to evaluate and verify the quality of the constructed knowledge graph. For extracting semantic information from text, the proposed approach supplemented the semantics information of the sentence and balanced the granularity of knowledge by combining the BERT-BiLSTM-CRF-based named entity recognition (NER) model and semantic role labeling (SRL)-based information extraction model. For irregular tables, a unified automatic extraction method was developed to process nested tables without preprocessing. The experiment constructed a comprehensive and scenario-oriented knowledge graph with 907 nodes, and showed high precision and recall for texts (89.37%, 85.42%) and tables (97.11%, 85.22%) on the test data. SNA results showed the proposed method ensured information richness and structural complexity.