Event Extraction for aviation accident reports through attention-based multi-label classification

In the aviation domain, historical aviation accidents are recorded to prevent catastrophe in the future. The national transportation safety board (NTSB) provides a large volume of accident reports and designs aviation event taxonomy to summarize what happens during the accidents. By utilizing the event taxonomy, we are able to design tools to analyze the accident reports, such as information retrieval systems, fault tree analysis, and causal chain analysis. However, one of the challenges is how to appropriately identify the accident reports with those event taxonomies. Traditionally, domain experts will label those events manually, which is time-consuming and subjective. In this paper, we propose the formulate the event labeling task as a sequence generation task. By adopting an advanced sequence generation framework, our experiment results show that we can achieve promising results on those frequent events.