Text Classification of Civil Aviation Safety Supervision Based on Multi-Model Fusion

This study proposed a novel method for text classification and co-occurrence network analysis based on 19,940 civil aviation regulatory records to improve regulatory methods and achieve precise and differentiated supervision. First, a hybrid deep-learning model combining BERT (bidirectional encoder representations from transformers) and TextCNN (text convolutional neural network) was developed to classify civil aviation regulatory records. Furthermore, a topic-based term frequency-inverse word frequency (TF-IWF) algorithm was used to extract key knowledge for each topic. Finally, a co-occurrence network graph of civil aviation safety supervision knowledge was drawn to analyze the intrinsic connections among various pieces of knowledge. The results showed that the BERT-TextCNN model outperformed 17 other classification models built using text representation methods, including the BERT, Word2vec, and Char models, and six neural networks, including the CNN, RCNN, DPCNN, Transformer, BiLSTM, and TextCNN models. The BERT-TextCNN model achieved macro precision, macro recall, and macro F1 score of 97.26%, 96.01%, and 96.61%, respectively. Furthermore, the topic-based civil aviation safety supervision knowledge-mining method effectively highlights key aspects of supervision work and presents a relationship network graph encompassing knowledge areas such as airports and equipment. This approach provides inspectors with a scientific basis for categorizing problem types and offers insights for an in-depth analysis of them. Additionally, it can facilitate the supervision and inspection of key regulatory matters, thereby improving regulatory efficiency.