Utilizing text descriptions from large language models to guide skeleton feature learning in a global-local-global manner.