Including semantic information in prompts improves the performance of pre-trained vision-language models.