Pre-Trained Vision-Language Models as Partial Annotators