Infer potential accidents from hazard reports: A causal hierarchical multi label classification approach

Abstract

Inferring categories of accidents caused by hidden hazards detected contributes to safety management and prevention of accidents. For the safety management of many enterprises in a large administrative area, it is necessary to rely on industry experts to review hazard reports produced by front-line employees and infer the categories of potential accidents, which is time-consuming and labor-intensive. In this study, a hierarchical multi-label classification model is proposed to learn a checklist reviewed by industry experts and realize the automatic inference of the accident categories based on hazard descriptions. We simultaneously use the causal effect estimation method designed according to the backdoor adjustment in causal theory to extract the causal part of the text that affects the inference and design a data augmentation method based on the discovered causal knowledge to make the model focus on the causal key information to improve the inference and generalization abilities of the models. From the perspective of theoretical and practical contributions, this study not only realizes the estimation of causal effect of hazard words and the automatic inference of accident categories, which provides support for further accident prevention and safety management. It also makes a successful attempt to apply causality theory combined with deep learning methods in the field of safety, providing a valuable reference for future research on the combination of causal theory and practical applications.