ATCBERT: Few-shot Intent Recognition for Air Traffic Control Instruction Understanding

Intent recognition in Air Traffic Control (ATC) aims to identify the pilot-controller vocal communication intents, which play a critical role in maintaining the safety and efficiency of air traffic. However, the recognition performance of intents with few samples is limited by the number of corresponding instructions in the scenario of an AI-assisted ATC operation. In this paper, ATCBERT is purposed as a novel framework for few-shot intent recognition to achieve high recognition accuracy for some intents in air traffic. Firstly, a joint pre-trained module is designed. By using the large corpus data labeled in the industry, Bidirectional Encoder Representation from Transformers (BERT) is fine-tuned to obtain the feature representation of the corpus in the domain of ATC. Besides, the masked language model is introduced to capture contextual information so as to enhance generalization ability. Secondly, the few-shot classification module is proposed. The classifier is trained and updated by a support set selected randomly from a dataset with few samples, and it is used to predict the intent of a new corpus. Finally, the proposed model is validated in the real-world ATC dataset derived from pilot-controller vocal communication in Beijing, China. It achieves a classification accuracy of 93.6% for few-shot intent recognition, outperforming the performance of the baselines. The findings of this research provide insights for intent recognition in ATC.