Can large language models understand uncommon meanings of common words?