Mega project safety hazards entities extraction and knowledge mining method based on UIE and improved Apriori Algorithm

Abstract

Mega projects generate a vast amount of safety hazard inspection records, which contain valuable knowledge on the relationships between various hazard elements and are essential for safety management. However, manually extracting safety hazard information and uncovering their internal correlations is time-consuming and inefficient, making it difficult to provide timely feedback for on-site safety management. An intelligent extraction and knowledge mining method was proposed for hazard source entities based on the Universal Information Extraction(UIE) framework and an improved Apriori algorithm. First, a safety hazard entity recognition model is constructed using the UIE framework, with specific entity extraction prompts defined. The model is fine-tuned with few-shot learning to achieve efficient and accurate automatic extraction of safety hazard entities. Then, an improved Apriori algorithm is introduced, considering the constraints of hazard data types, to perform multi-factor association rule mining and visualization. Case analysis shows that the proposed safety hazard entity extraction model achieved F1 scores of 0.892 and 0.886 on the validation and test datasets respectively, significantly outperforming the baseline model′s scores of 0.253 and 0.307, and the overall entity recognition rate improves 36.66%. Additionally, the extracted multi-factor strong association rules are visualized using Sankey diagrams and association network graphs, demonstrating good interpretability. Research findings provides an efficient and intelligent method for mining knowledge from the vast amount of safety hazard text data generated in mega construction projects, offering data-driven support for the development of targeted safety management measures on construction sites.

Keywords

mega project / safety hazards / universal information extraction(UIE) / knowledge mining / natural language processing(NLP)