Enhancing Aviation Safety: Natural Language Processing for Automated Classification of Aviation Occurrences

This study presents a natural language processing framework for automated classification of aviation safety occurrences. The proposed approach develops a domain-specific Bidirectional Encoder Representations from Transformers model trained on a large corpus of aviation safety reports to classify occurrences according to the European Union Aviation Safety Agency Key Risk Areas. The work addresses limitations of current manual safety assessment processes, which require weeks or months to analyze a single occurrence, delaying the identification of emerging risks. To overcome the 512-token limitation of transformer architectures, a sentence-based splitting and aggregation methodology is introduced. Long reports are divided into sentences, processed independently, and subsequently aggregated, preserving narrative context while avoiding information loss due to truncation. This approach improves classification performance by 6 percentage points in F1 score. Experimental results demonstrate that dataset size has a greater impact on performance than dataset balance. Training on the full dataset of 41,000 reports outperforms a balanced subset of 6400 reports by 10 percentage points in F1 score. The developed framework achieves an F1 score of 94%, exceeding human analyst consistency levels while reducing assessment time from months to seconds. The framework enables standardized, near-real-time aviation safety analysis, supporting proactive safety risk management.