A BERT-based Intent Recognition and Slot Filling Joint Model for Air Traffic Control Instruction Understanding

Air Traffic Control (ATC) relies on vocal communication between controllers and pilots to ensure safety and improve control efficiency. However, as the volume of flights grows and the frequency of vocal communication becomes higher, the work load on controllers becomes heavier, which may lead to unsafe accidents. It is crucial to extract information from control speech and understand it. But the highly specialized terminology and the varied presentation of ATC instructions make it challenging to develop a comprehensive semantic parsing model for extracting key information Intents and Specific Parameters (ISPs) of ATC instructions. To address this issue, this paper proposes a BERT-based Intent Recognition and Slot Filling joint model (BERT-IRSF) for ATC instruction understanding. The proposed model employs an end-to-end approach to extract sentence representations with context-dependent relationships. Firstly, the text-rich embedding vectors of ATC instructions are fused. Secondly, the semantic extraction of control commands is performed using a transformer encoder, in which a federated model is also integrated that explicitly establishes the connection between intents and slots, processing both tasks simultaneously by sharing semantic features. Finally, the labels are predicted using softmax and the Conditional Random Field (CRF) layer to get the intent and slot results. The proposed model is evaluated on a real operational ATC semantic understanding dataset, and experimental results show significant improvement in multiple metrics compared to other state-of-the-art deep learning methods, which demonstrates the potential of the proposed method to providing decision support for ATC automation system.