Application of machine learning to mapping primary causal factors in self reported safety narratives

Abstract

A new method for analysis of text-based reports in accident coding is suggested. This approach utilizes latent semantic analysis to infer higher-order structures between documents and provide an unbiased metric to the narrative analysis process. Results from this study on a small sample of aviation safety narratives demonstrates an unsupervised categorization accuracy of 44% for primary-cause within the existing taxonomy. If provided with a large sample set, the indication is that a significant increase in accuracy is possible along with the possibility of recoding between data sets. Demonstrated is the ability of LSA to capture contextual proximity of a narrative.