Notice-to-Airmen Text-Error-Correction Method Based on Ccf-MacBERT and Knowledge Graph

With the exponential increase in the number of the Notices to Airmen (NOTAMs), the inefficiencies and high error rates of manual processing have become evident. To address these issues, NOTAMs from the Intelligence Center between September 2020 and October 2023 were used as the base data. By employing the trained Casv-Graph Neural Network model and the tianzege-Convolutional Neural Network model, the phonetic and morphological feature similarities of Chinese characters were outputted, and a knowledge graph with phonetic and morphological features as nodes was constructed. Based on the similarities, a database of similar characters was compiled and used as a correction medium for the BERT model. Synonyms from the database were used as masking objects during the masked language phase of the MacBERT model to mask the original text, thereby constructing the Ccf-MacBERT model. This approach improves the model’s practical applicability and computational efficiency. When selecting synonyms with a similarity of 0.9 or higher and using the NOTAM-text data set, the training parameters of the model were superior to those of the existing NOTAM text-error-correction models.