Chemical or food risk is relative to situations in which chemical products are dangerous for human or animal health and consumption, and for environment. The experts that guarantee the control and management of such substances face large amount of scientific literature, that have to be analyzed to support the decision making process. We propose an automatic assistance for the analysis of this literature. We tackle the task as the categorization problem: we want to categorize the sentences from corpora into classes of substance-related risk. We use two approaches: supervised machine learning and information retrieval. The results obtained with supervised machine learning (all classes together, F-measure around 0.8 for food risk, between 0.61 and 0.64 for chemical risk) are better than those obtained with information retrieval (all classes together, F-measure between 0.18 and 0.226 for food risk, between 0.20 and 0.32 for chemical risk). Recall is competitive with the two approaches.