Exploring the Dialogue Comprehension Ability of Large Language Models