Multi-task hierarchical network for semantic understanding of air traffic controller-pilot communication

Flight situational awareness in civil aviation relies on the semantic understanding of both the key details and the full picture from the Air Traffic Controller (ATCo) and pilot communication. This paper proposes a novel end-to-end Multi-Task Hierarchical Network (MTHN) for automatically understanding ATCo-pilot communication, handling slot filling, role detection, and intent recognition at different levels while adaptively integrating them. Specifically, we introduce a word-based knowledge-masked slot distillation module that constructs an ATC knowledge base to dynamically mask keywords during teacher-student distillation. Considering the distinct intent differences between ATCos and pilots, we design a sentence-based role-aware intent attention module that extracts role label space vectors as context to enrich intent representations. To exploit the complementarity across different semantic levels in ATCo-pilot communication, we explicitly develop an adaptive bi-interaction flow module that dynamically explores semantic dependencies among tasks. Extensive experiments on real-world datasets collected in China show the superior performance of MTHN, compared to state-of-the-art baselines in both general natural language understanding and ATC-specific text processing. Our results highlight that MTHN achieves 99.26 %, 97.25 %, and 96.22 % accuracy across key slots, as well as 96.59 % accuracy in speaker role classification. Moreover, it can perceive multi-label deep intents behind sentences. These analytical findings demonstrate the potential to reduce human errors in high-concurrency ATCo-pilot interactions under dense operational conditions.