Automatic Semantic Verification of BiLSTM model for Readbacks in Radiotelephony Communication

To ensure the safety of air transportation, the pilot must read back the instructions sent by the air traffic controller, and the controller has to further confirm it. However, due to fatigue, tension, negligence and other reasons, the controller often failed to find out the readback error in time, which is a huge hidden danger in the safety of civil aviation transportation. In order to solve this problem, an automatic semantic verification method based on bidirectional Long Short-Term Memory (BiLSTM) is proposed. Firstly, two parallel BiLSTM are used to extract the semantic features of controller’s instructions and the pilot's readback respectively; then a matching matrix is generated by matching the output of the two BiLSTMs at each time step; finally, a k-Max pooling layer is added after matching matrix to pick out the top-k matching features and input it into a Multi-Layer Perceptron (MLP) to obtain a final matching score to determine whether the readback is consistent in semantics with the instruction sent by controllers. The experiment shows that the method is effective in solving the automatic verification task of the aviation radiotelephony communication, and the average test accuracy is 90.53%.