A Semi-Supervised Learning Method for Event Extraction in Radiotelephony Communication Content

Event extraction in radiotelephony communication enables rapid and accurate extraction of event information, constructing standardized radiotelephony communication structures to provide real-time support for air traffic control while facilitating aviation safety management and accident investigation.However, existing event extraction methods for radiotelephony communication face challenges such as overlapping argument roles and scarce annotated data. To address these issues, this study proposes a semi-supervised learning-based event extraction model. The model employs BERT for word embedding and feature extraction, utilizes a bidirectional long short-term memory network (BiLSTM) to capture sequential sentence features, and adopts multiple binary classifiers for output to mitigate argument role overlap.Furthermore, the model incorporates rule-based template-generated unlabeled data for semi-supervised training and introduces a high-quality pseudo-label selection method to filter high-quality training samples. Experimental results on a self-constructed radiotelephony communication event extraction dataset demonstrate that the proposed model achieves F1-scores of 97.92% and 97.77% in trigger word extraction and argument extraction tasks, respectively. It effectively resolves argument role overlap issues and outperforms existing mainstream event extraction models.