Classification of NOTAMs for General and Scenario-Specific Applications Using Fine-Tuned BERT

The management of NOTAMs on the flight deck remains a challenge for flight crews, who must process large volumes of dense textual information to identify the notices relevant to their mission. Despite ongoing modernisation efforts, the NOTAM system still relies on a legacy format with limited support for automated filtering and prioritisation, which increases pilot workload and motivates the use of Natural Language Processing (NLP) techniques for information management. This paper presents therefore an NLP framework based on a shared pre-trained distilled BERT (Bidirectional Encoder Representations from Transformers) encoder with two task-specific prediction heads, designed to support NOTAM analysis at different stages of the flight. The first task classifies each NOTAM into one of 20 operational categories to support pre-flight briefing. The second task estimates the operational severity of a NOTAM under three emergency diversion scenarios to support in-flight decision-making within cockpit decision-support systems. To address the lack of annotated data and the imbalance across categories, a dataset of 3000 NOTAMs was built through a human-supervised Large Language Model (LLM) generation workflow seeded with real messages. All samples are annotated for both tasks according to written annotation rules. The proposed framework is based on a DistilBERT encoder which receives fine-tuned through Low-Rank Adapters (LoRA), achieves a macro F1-score of 0.967 on classification and R2; values between 0.82 and 0.87 across the three severity scenarios.