Evidence remains limited on how unsafe behaviors connect and progress directionally within construction accident sequences. This study investigates these relationships using 329 written fall-from-height (FFH) accident investigation reports from Jiangsu Province, China.
Text mining was integrated with complex network analysis. Unsafe behavior nodes were identified through coding and topic modeling, while a convolutional neural network was used to extract progressive and parallel relations from accident narratives. The resulting directed network comprised 71 nodes and 217 edges and was analyzed using degree, node strength, betweenness centrality, and shortest-path measures.
The resulting network exhibits a heterogeneous structure, with a limited number of unsafe behaviors occupying prominent positions across multiple structural measures. The analysis further identifies recurrent behavioral pathways linking unsafe acts during FFH accident development, indicating that documented accident processes involve interconnected rather than isolated behaviors. The identified directional relations represent reported behavioral progression in accident narratives.
The findings can support safety interventions that move beyond isolated unsafe acts by targeting structurally important behaviors together with their associated behavioral pathways. Application of the identified priorities to other settings should consider differences in local construction and accident-reporting contexts.
This study develops an evidence-linked approach for constructing unsafe behavior networks directly from accident narratives. By extracting directional behavioral relations from documented accident sequences rather than defining network connections solely through co-occurrence or expert judgment, the approach provides a reproducible basis for examining the organization and progression of unsafe behaviors in construction accidents.
