Frequency-Semantic Enhanced Variational Autoencoder for Zero-Shot Skeleton-based Action Recognition