Identification of unsafe behaviors and coupling relationship based on enhanced text preprocessing and association analysis

Purpose Identifying unsafe behaviors among construction workers on building sites is critical for improving construction safety. However, existing methods often face challenges related to data noise, ineffective phrase extraction, and inadequate feature extraction during text preprocessing. This study aims to present an improved method for identifying unsafe behaviors by enhancing the text preprocessing stage. Design/methodology/approach To address common issues such as missing data, inconsistencies, and redundancies in accident reports, a multi-stage data cleaning process was developed. This process incrementally cleans and optimizes the data to improve the quality and completeness of behavior descriptions. For phrase extraction, Pointwise Mutual Information (PMI) was used to capture high-frequency correlated phrases, while Conditional Random Fields (CRF) were applied to accurately determine phrase boundaries related to unsafe behaviors, improving the analysis of complex text structures. Additionally, association rule analysis was used to identify latent relationships between unsafe behaviors, providing a scientific basis for targeted intervention strategies. Findings The method developed in this study successfully identifies and categorizes unsafe behaviors, refining an initial list into five categories containing sixteen critical unsafe behaviors. The approach provides robust support for timely improvements in construction safety management practices. Originality/value This study introduces an enhanced text preprocessing method to identify unsafe behaviors among construction workers, improving the accuracy and efficiency of safety risk management. The proposed framework provides significant value for improving the safety management practices in construction settings by offering a more systematic and reliable approach to identifying hazardous behaviors.