Development of Multilabel Text Classification with Contrastive Learning for Safety Hazard Analysis in Large-Scale Construction Projects

Text analysis of safety hazards is particularly important to grasp construction safety risks. However, hazard descriptions and locations are usually weakly directional and ambiguous, leading to a significant one-to-many mapping relationship. In addition, there are many hazard occurrence locations existing in large-scale projects, and the differences of spatial information set between descriptions are small. Common models often struggle in accurately achieving multilabel text classification (MLTC) tasks. Therefore, this paper proposes an MLTC method with contrastive learning to assist in the analysis of safety hazards in large-scale construction projects. In this process, A Lite Bidirectional Encoder Representations from Transformers (ALBERT) and the Text Convolutional Neural Network (TextCNN) are utilized for text embedding and feature enhancement. A contrastive learning module fusing self-supervised and supervised learning is proposed to enhance text representation capability. Combining two learning methods, the module increases the sparsity of samples with more diverse labels in vector space while fully learning manually information. Then, hazard information is analyzed to further know key hazard points at each location. To validate effectiveness, a hazard location multilabel data set is semiautomatically constructed for the first time, to the best of our knowledge. The F1 score of our model is 0.8728, which surpasses common models. When category labels are excessive (18 in total), the contrastive learning module leads to more remarkable performance. Then the hazard points are fed back to locations through graphical information, aiding managers in gaining a visual understanding of hazard distribution. This study provides a promising intelligent method to quickly grasp the spatial distribution of hazards in large-scale projects.