Large Language Models as Annotators: Enhancing Generalization of NLP Models at Minimal Cost