NOTAM Smartification

Notice to airmen (NOTAM) are text messages targeted to pilots to make them aware of short term events and obstacles. Typical examples are military actions in airspaces or closed runways. Pilots in Switzerland are obliged to conduct a NOTAM briefing before take-off. Thousands of NOTAMs are broadcasted every day and poorly filtered concerning a pilot's envisaged route and contain a lot of irrelevant information. Smart NOTAM is a service offered by Skyguide that filters irrelevant messages for a specific route or period. This service also includes a smartification process that shortens the messages by omitting meaningless words and phrases. It also converts the messages into a standardised phraseology.

This project provides an overview of current natural language processing literature in general and machine translation, as well as text summarization in the domain of NOTAM messages. Having access to thousands of NOTAMs with their smartified version, a machine learning model that performs the smartification process was implemented and evaluated.

A data quality assessment showed that from 1.1 million messages, almost 50% of the raw NOTAM messages are already smart and thus do not need to be smartified at all. Furthermore, raw messages leave the smartification process either unchanged or with substantial changes. Based on these findings, an additional model was implemented which classifies whether a NOTAM needs to be smartified or not.

For the classification task, several text classification models based on either recurrent neural networks, convolutional neural networks (CNNs) or transformers were evaluated. The text classification model based on CNNs proved to be the most accurate and achieved an F1 score of 93.66% on unseen test data. To perform the smartification process, a sequence to sequence transformer model was trained and evaluated. NOTAMs contain a high number of factual details such as coordinates or frequencies, making the smartification process vulnerable to the inaccurate generation of such details. To improve the reproduction of factual details, the transformer model was extended by a pointer-network that allows it to copy words directly from the source text. This way, the evaluated model achieved a BLEU score of 86.12% and a ROUGE score of 90.23% on unseen test data. Yet there are some cases where the model fails to smartify the messages: if some factual details are added from another source or if the messages are written in a language other than using English abbreviations.