ATSIU: A large-scale dataset for spoken instruction understanding in air traffic control

Abstract

Spoken instruction communication between air traffic controllers and pilots is a crucial and fundamental process of air traffic control (ATC). Automated understanding of these instructions can significantly enhance the safety and efficiency of air traffic, making this field a prominent area of current research. However, thorough scrutiny of instruction semantics and associated benchmarks for instruction understanding in the ATC domain have remained unexplored. In this paper, we aim to build a large-scale specialized air traffic spoken instruction understanding (ATSIU) dataset to bridge this gap. The proposed dataset features a tailored hierarchical intent taxonomy, encompassing 9 coarse-grained intents and 26 fine-grained intents, together with 78 customized slots. It was developed by transcribing over 200 hours of raw ATC audio into 19.8k texts, each meticulously annotated with golden intent and slot labels by industry professionals. Moreover, we present an air traffic spoken instruction understanding network (ATSIU-Net) as a baseline method for ATC spoken instruction understanding, which employs a pre-trained language model and a joint learning mechanism to facilitate collaborative ATC intent detection and slot filling. Extensive experiment results demonstrate that ATSIU-Net establishes promising performance benchmarks while revealing key challenges in intent granularity, flight phases, multi-task learning, and low-data scenarios. It is believed that this work not only showcases the potential of advanced algorithms in ATC-specific domain, but also provides diverse research topics for the common natural language processing community.