Multi-label classification of contributing causal factors in self-reported safety narratives

Abstract

Three methods are demonstrated for automated classification of aviation safety narratives within an existing complex taxonomy. Utilizing latent semantic analysis trained against 4497 narratives at the sentence level, primary problem and contributing factor labels were assessed. Results from a sample of 2987 narratives provided a mean unsupervised categorization precision of 0.35% and recall of 0.78% for contributing-factors within the taxonomy. Categorization of the primary problem at the sentence level resulted in a modal accuracy of 0.46%. Overall, the results suggested that the demonstrated approaches were viable in bringing additional tools and insights to safety researchers.