DOI

10.24348/coria.2021.court_20

Recently, transformer-based models have beaten humans in Natural Language Understanding (NLU) tasks such as text classification, and have been used in specialized fields such as healthcare. In this context, our general aim is to explore how such models could help support analysts working in safety in aviation, in particular when they are used on incident reports. In this article, we work with the Aviation Safety Reporting System (ASRS) data set. It is made up of incident reports in English, as well as supporting metadata. Such reports are characterized by the heavy use of specialized language, abbreviations, and domain-specific vocabulary, as opposed to general day-to-day English. We explore the idea that analyst work can be re-framed as a set of NLU tasks. We then propose an experimental procedure to try and use transformer-based models on one of these tasks.