The identification of the severity of patient safety events promotes prioritized safety analysis and intervention. The Harm Scale developed by the Agency for Healthcare Research and Quality is widely used in the US hospitals. However, recent studies have indicated a moderate to poor inter-rater reliability of the Harm Scale across a number of US hospitals. Although the reasons are multi-folded, biased human judgments are recognized as a prominent factor. We proposed that key information to identify and refine the severity of harm is contained in the narrative data in patient safety reports. Using automated text classification to categorize harm scores is intended to provide reduced subjective judgments and much improved efficiency. We evaluated different types of classification algorithms using a corpus of patient safety reports from a US health care system. The results demonstrate the effectiveness and efficiency of the proposed methods. Accordingly, human biases on the application of harm scores are expected to be largely reduced. Our finding holds promise to serve as a semi-supervised tool during the process of manually reviewing and analyzing patient safety events.