A BERT-Based Model for Question Answering on Construction Incident Reports

Construction sites are among the most hazardous work-places. To reduce accidents, it is required to identify risky situations beforehand, and to describe which countermeasures to put in place. In this paper, we investigate possible techniques to support the identification of risky activities and potential hazards associated with those activities. More precisely, we propose a method for classifying injury narratives based on different attributes, such as work activity, injury type, and injury severity. We formulate our problem as a Question Answering (QA) task by fine-tuning BERT sentence-pair classification model, and we achieve state-of-the-art results on a dataset obtained from the Occupational Safety and Health Administration (OSHA). In addition, we propose a method for identifying potential hazardous items using a model-agnostic technique.