Action Sequencing in Construction Accident Reports using Probabilistic Language Model

Construction remains among the most hazardous workplaces, thus a significant amount of time and effort in reporting and investigating the accident occurrences has been done in the past decades by government agencies. In light of construction safety, analyzing textual information in construction accident records may assist in our comprehension of past data and be used to minimize future risks. Many attempts have been made in previous studies to identify causes and related entities but yet consider worker activities and behaviors. This study presents a framework that adopts a Probabilistic Language Model to sequence actions taken by workers that depict construction scenarios from unstructured accident narrative reports. The proposed approach achieved outstanding performances with the highest sequence accuracy and pairwise sequence accuracy of 84.81 % and 89.12%, respectively. Moreover, an action sequence database that can explain the relationship between workers' actions was created. This research is anticipated to contribute to enhancing understanding and establishing safety management systems to actively forecast and prevent accidents.