Text this: Latent space improved masked reconstruction model for human skeleton-based action recognition