Using structural topic modeling to identify latent topics and trends in aviation incident reports

The Aviation Safety Reporting System includes over a million confidential reports describing aviation safety incidents. Natural language processing techniques allow for relatively rapid and largely automated analysis of large collections of text data. Interpretation of the results and further investigations by subject matter experts can produce meaningful results. This explains the many commercial and academic applications of natural language processing to aviation safety reports. Relatively few published articles have, however, employed topic modeling, an approach that can identify latent structure within a corpus of documents. Topic modeling is more flexible and relies less on subject matter experts than alternative document categorization and clustering methods. It can, for example, uncover any number of topics hidden in a set of incident reports that have been, or would be, assigned to the same category when using labels and methods applied in earlier research. This article describes the application of structural topic modeling to Aviation Safety Reporting System data. The application identifies known issues. The method also reveals previously unreported connections. Sample results reported here highlight fuel pump, tank, and landing gear issues and the relative insignificance of smoke and fire issues for private aircraft. The results also reveal the prominence of the Quiet Bridge Visual and Tip Toe Visual approach paths at San Francisco International Airport in safety incident reports. These results would, ideally, be verified by subject matter experts before being used to set priorities when planning future safety studies.