Using Machine Learning and Natural Language Processing to Classify Safety Incident Narratives. Millions of safety incidents occur each year and are captured by government agencies and safety organizations. Many of these records include narrative descriptions of the incident details that are important to develop interventions to improve safety. Underwriters Laboratories has developed an approach using natural language processing and machine learning techniques to significantly reduce the effort associated with large scale analysis. Using variations of Bidirectional Encoder Representations from Transformers (BERT), narratives were transformed to dense vector representations suitable for use with machine learning algorithms. Both supervised and unsupervised machine learning approaches have been used to classify the narratives based on particular analysis goals. For an analysis for furniture tipover risk, a supervised model was used, while an unsupervised model was used to uncover previously unknown characteristics of a broader set of safety incidents. The unsupervised approach used K-Means clustering; post-processing and analysis included keyword analysis, metadata correlation and expert insight. The results of this project validated previously assigned metadata elements and added to the understanding of precursor events, incident characteristics and demographics of victims involved in the incidents. The use of these methods significantly reduced the time associated with the analysis, uncovered previously hidden insights and show strong potential to replace manual coding of metadata elements.