(Georgia Institute of Technology, 2023-06)
Chandra, Chetan; Jing, Xiao; Bendarkar, Mayank; Sawant, Kshitij; Elias, Lidya R.; Kirby, Michelle; Mavris, Dimitri N.
Data-driven methods form the frontier of reactive aviation safety analysis. While analysis of quantitative data from flight operations is common, text narratives of accidents and incidents have not been sufficiently mined. Among the many use cases of aviation text-data mining, automatically extracting safety concepts is probably the most important. Bidirectional EncoderRepresentations from Transformers (BERT) is a transformer-based large language model that is openly available and has been adapted to numerous domain-specific tasks. The present work provides a comprehensive methodology to develop domain-specific BERT model starting from the base model. A preliminary aviation domain-specific BERT model is developed in this work. This Aviation-BERT model is pre-trained from the BERT-Base model using accident and incident text narratives from the National Transportation Safety Board (NTSB) and AviationSafety Reporting System (ASRS) using mixed-domain pre-training. Aviation-BERT is shown to outperform BERT when it comes to text-mining tasks on aviation text datasets. It is also expected to be of tremendous value in numerous downstream tasks in the analysis of aviation text corpora.
(Georgia Institute of Technology, 2023-06)
Jing, Xiao; Chennakesavan, Akul; Chandra, Chetan; Bendarkar, Mayank; Kirby, Michelle; Mavris, Dimitri N.
The advent of transformer-based models pre-trained on large-scale text corpora has revolutionized Natural Language Processing (NLP) in recent years. Models such as BERT (Bidirectional Encoder Representations from Transformers) offer powerful tools for understanding contextual information and have achieved impressive results in numerous language understanding tasks. However, their application in the aviation domain remains relatively unexplored. This study discusses the challenges of applying multi-label classification problems on aviation text data. A custom aviation domain specific BERT model (Aviation-BERT) is compared against BERT-base-uncased for anomaly event classification in the Aviation Safety Reporting System (ASRS) data. Aviation-BERT is shown to have superior performance based on multiple metrics. By focusing on the potential of NLP in advancing complex aviation safety report analysis, the present work offers a comprehensive evaluation of BERT on aviation domain datasets and discusses its strengths and weaknesses. This research highlights the significance of domain-specific NLP models in improving the accuracy and efficiency of safety report classification and analysis in the aviation industry.