CLAIR: Evaluating Image Captions with Large Language Models