Text this: Learning behavior aware features across spaces for improved 3D human motion prediction