As more and more attention is paid to the digital protection of traditional culture, Labanotation, as a way to protect traditional culture, has also attracted the attention of scholars. Especially in the field of automatically generating Labanotation, some methods have been proposed. However, existing Labanotation automatic generation methods ignore the high-order kinematic dependencies between the performer’s body joints. Furthermore, the static modeling of the channel skeleton topology loses the unique correlation of each channel. Therefore, these methods cannot accurately represent complex dance movements. We propose a Hypergraph Self-Attention (HSA) and Channel Topology Specialization (CTS) network (HSA-CTS) for automatic generation of Labanotation. HSA-CTS includes a global context attention feature extraction module and a local channel topology specialized feature extraction module. First, it models the human body’s high-order kinematic dependence on the global spatiotemporal relationship of the skeleton, then dynamically learns the topology in different channels and effectively aggregates joint features in different channels for human action recognition. Experimental results show that our proposed method outperforms state-of-the-art methods on Laban16 and Laban48 (common datasets for Labanotation studies).