Automatic repetition instruction generation for air traffic control training using multi-task learning with an improved copy network

To eliminate the need for human pseudo-pilots in air traffic controller training, a multi-task framework with a copy mechanism is proposed to automatically generate a repetition instruction, which can greatly reduce human and device resources. The proposed framework is implemented using a sequence-to-sequence architecture and is optimized through multi-task learning, including text instruction understanding (TIU) and repetition instruction generation (RIG). In the proposed framework, a shared encoder is designed to learn the representations from the input sequence. Two task-specific attention modules are proposed to extract task-specific features, and different decoders are applied to generate the TIU and RIG outputs. The TIU module provides both word- and sentence-level implicit semantic representations through slot filling and intent detection tasks. In the decoding procedure of the RIG task, the slot context vector is regarded as an additional input used to solve the problem of ambivalent words. Simultaneously, an improved copy network based on multi-task learning is proposed to consider the correlations between the input instruction and the output repetition in the RIG module. To avoid a loss imbalance for the TIU and RIG tasks, the gradient normalization algorithm is applied to learn the loss weights automatically by adjusting the loss gradient. Finally, the proposed framework was trained and evaluated using a real-world air traffic control corpus. The experiment results demonstrate that the proposed framework significantly outperforms other state-of-the-art methods for the RIG task, achieving a novel integrated metric of 97.19%.