An LLM–BERT and Complex Network Framework for Construction Accident Causation Analysis

Construction accident reports contain rich causal evidence; however, their unstructured narratives make systematic analysis difficult. Recent advances in large language models (LLMs) have created new opportunities to leverage such information at scale. This study develops an integrated LLM–BERT–network framework for analyzing construction accident causation. Based on 347 official construction accident investigation reports, a DeepSeek-based pipeline with human-in-the-loop quality control was used to extract causal keywords describing direct and indirect causes, yielding 2572 keywords. A BERT-based semantic normalization procedure then consolidated synonymous expressions, reducing 811 deduplicated keywords to 104 normalized terms (an 87.2% reduction in vocabulary size). A manual sample-based evaluation further supported the reliability of the LLM-based extraction and BERT-based normalization procedures. The normalized keywords were further organized into a hierarchical taxonomy and used to construct a directed keyword-association network linking indirect and direct causes for structured relational analysis. To strengthen methodological rigor, additional validation and analytical experiments were conducted, including manual sample-based evaluation of keyword extraction, sensitivity analysis of normalization settings, and examination of representative failure cases. The results support the reliability and robustness of the proposed framework. The analysis indicates that behavior-related factors and management deficiencies occupy structurally important positions in the directed network. Overall, the findings suggest that construction accidents arise from the interaction of human, managerial, environmental, material, and technical factors rather than isolated single causes. Effective prevention therefore requires system-oriented interventions that strengthen worker competence, supervision, training, accountability, and hazard identification.