Sequence image layout generation for construction accident simulation using domain-tuned NER by ZSL-PLM and scene graph learning

Safety in the construction industry remains a critical concern, with accident analysis often relying on time-intensive, error-prone manual processes. The unstructured nature of accident reporting text complicates the extraction and interpretation of sequence information essential for understanding accident dynamics. This study addresses these challenges by proposing a domain-specific AI framework that automates the extraction, structuring, predicting, and visualization of sequence information to generate interpretable 2D image layouts. A key achievement of the framework is its ability to infer accident sequences and generate layouts without relying on large-scale or ground truth image datasets for training. By integrating domain-tuned named entity recognition (NER), scene graph learning, and graph convolutional networks (GCNs), the framework achieves a highly robust entity diversity and demonstrates accurate entity recognition. These metrics, alongside notable improvements in spatial alignment (+88.89 %) and temporal consistency (+68.75 %) over the text-based model using DALL-E API, laying a foundation for robust framework advancing domain-specific deep learning applications and enhancing sequential analysis of construction accidents.