Navigating Massive Text Reports: An Automated Approach to Aviation Safety Reporting System Safety Event Detection

Abstract

In response to the increasing report intake and the need for more efficient and effective safety event detection and monitoring in the national airspace system (NAS), this work proposes an automated and sustainable workflow that collectively applies natural language processing (NLP) techniques to the Aviation Safety Reporting System (ASRS) report narratives. The proposed workflow uses latent Dirichlet allocation (LDA), a probabilistic model for uncovering hidden topics within text documents, to perform topic modeling and constructs topic trajectories based on Document Frequency – Inverse Document Frequency $(\mathrm{DF}-\mathrm{IDF})$ . By applying spectral analysis to the constructed trajectories, we identify multiple periodic and aperiodic topics in unfiltered ASRS report narratives from 2011 to 2021 with minimal human input. Through validation, we confirm that the periodic and aperiodic topics detected can be linked to actual safety events trending in the aviation community and are valuable for safety improvement. Beyond existing topic modeling applications on ASRS reports, this work could deliver more relevant and practical information to safety management experts, not only filling the gap between the topic model output and practical applications but also providing safety management experts with a holistic view of the entire report intake to oversee commonalities. The proposed workflow, with appropriate modifications, has the potential to be adapted to other safety reporting systems.