Leveraging Humans to Detect and Fix Representation Misalignment