关键词: 融合空域, 指令解析, 深度学习, 意图识别, 槽填充, 联合模型

Abstract:

With the rapid development of UAV technology， the number of UAVs and airspace demand has increased significantly， which make it a trend for UAVs to enter integrated airspace. To modify the traditional ATC management pattern of “man in loop” and increase the control efficiency of fusion airspace， deep learning method is applied to air traffic control instruction parsing. According to the characteristics of ATC terms， this paper sorted， expanded， classified and annotated the instructions to write an ATC instructions dataset which can be used for learning of natural language understanding model. BiGRU-CRF model is used as the infrastructure of deep learning， and attention mechanism and intention feedback mechanism are added to construct a joint model to obtain instruction intention and instruction parameters. The evaluation results on ATC instruction dataset and ATIS dataset show that the model is nearly 1.5% better than the infrastructure in the task of intent identification and slot filling. The method presented in this paper is practical and effective， which provides strong support for the development of UAV natural language control technology.

Key words: fusion airspace, instruction parse, deep learning, intention recognition, slot filling, joint model