Do Prompt-Based Models Really Understand the Meaning of their Prompts?