Combining BERT with numerical variables to classify injury leave based on accident description

The occurrence of work accidents may threaten the workers’ health and lead to consequences for the organizations as well, such as restructuring of work and direct/indirect costs with the absence of the worker. In this context, accident investigation reports contain information that can support companies to propose preventive and mitigative measures and identify causes and consequences of injury events. However, this information is frequently complex, redundant, and/or incomplete. Additionally, a complete human review of the entire database is arduous, considering numerous reports produced by a company. Indeed, Natural Language Processing (NLP)-based techniques are suitable for analyzing a massive amount of textual information. In this paper, we adopted NLP techniques to determine whether an injury leave would be expected from a given accident report. The methodology was applied to accident reports collected from an actual hydroelectric power company using Bidirectional Encoder Representations from Transformers (BERT), a state-of-art NLP method. The text representations provided by BERT model were combined with numerical and binary variables extracted from the accident reports. These combined variables are input to a Multilayer Perceptron (MLP) that predicts the occurrence of the accident leave for a given accident. After cross-validation, the results showed a median accuracy of 73.5%. Additionally, we discuss several reports that presented high and low proportions of correct classifications by the models tested and discussed the possible reasons. Indeed, accident investigation reports provide useful knowledge to support decisions in the safety context.