Oil and gas refineries convert the crude oil into valuable by-products. To that end, a series of mass and heat transfer processes take place, which involve different hazardous materials under extreme operational conditions. The uncontrollable release of these materials may lead to catastrophic accidents. In order to avoid these undesirable events, related to the operation of hazardous facilities, international and national standards demand the execution of Risk Analysis (RA). RA is a systematic study of accidental hypotheses and starts with the identification of the potential hazards. This step often involves the evaluation of different documents, such as plant design, equipment and/or material specifications, and other sources of information. Firstly, risk analysts extract information to characterize the system to be analyzed. Next, they evaluate the data, which are frequently in the form of texts, in order to postulate possible accidents and its consequences. In this context, text mining and Natural Language Processing (NLP) can be useful for exploratory text analysis. In this paper, we propose a practical methodology based on Bidirectional Encoder Representations from Transformers (BERT) to identify the potential consequences of accidents in an oil refinery, and then assist the identification of potential hazards related to the system. Thus, it will be possible to reduce the required efforts in completing the RA early stages. The developed model was able to predict the 11 possible consequences identified in the documents. The highest score to predict the occurrence was obtained to burn injury, with a precision of 97.5% and recall of 95.12%. The model achieved an accuracy of 96.87% on test data.