In many industry fields, accident reports are required by law, and must be provided even for minor incidents. These documents describe the context and events that lead to an accident. Often, descriptions are given in text format by the victim themselves and/or the occupational safety technician. Over time, accident report archives may grow into a large amount of data for manual analysis. Moreover, information is not always available in organized and systematic formats, with datasets populated by multiple people, leading to problems such as inconsistencies, typographical errors and redundant data. Therefore, it is worth developing an automated form of data analysis. In recent years, text mining and natural language processing techniques have been developed and employed in diverse fields, allowing for obtainment of useful information from text datasets. This paper proposes a framework for accident reports classification based on their description, with the objective of determine whether or not an injury leave would be expected. We employ Bidirectional Encoder Representations from Transformers (BERT), a stateof- art natural language processing method, to tackle the aforementioned problem. We also present an example applied to a real accident report database from a hydropower company. The proposed method achieves an accuracy of 0.44 and a Matthews Correlation of Coefficient (MCC) of 0.79, which represent promising results, considering the characteristics of the dataset.