A Semi-supervised Approach to Multi-label Classification of NOTAMs using BERT

Notices to Air Missions (NOTAMs) are key for the safe operation of commercial and civil aviation flights around the globe, as they provide up-to-date information on any disturbance on aerodromes and in the airspace. In this work, we propose a framework to train bidirectional encoder representations from transformers (BERT) to assign labels to NOTAM messages using a multi-label classification approach where each NOTAM can be assigned multiple labels. To deal with the scarcity of labeled data for this task, we propose a semi-supervised learning framework using the MixMatch algorithm to allow the leveraging of unlabeled NOTAMs, and reduce the need for expert labeled NOTAMs. We demonstrate that the MixMatch algorithm combined with focal loss improves the performance of BERT on the multi-label classification of NOTAMs, with the micro-averaged F1-score improving from 0.93 to 0.96 on a publicly available test dataset.