An ideal emotional dialogue system should have emotion perception capability in conversation across various new scenarios, thereby extracting user needs to guide dialogue generation. However, due to the lack of sufficient training data in new scenarios, improving the model’s zero-shot emotion perception capability in conversation has become a new challenge. Moreover, current research mostly focuses on single tasks, which cannot comprehensively reflect the model’s emotion perception ability. In this paper, we propose an Internal-to-External Chain-of-Thought method (IoECoT) for the emotion perception in conversation. First, the personality information of target user are extracted from the dialogue history as internal factors, while the polarity of utterances serves as external factors. Guided by internal factors, the model perceives emotions based on external factors. By evaluating the model’s performance on both Emotion Recognition in Conversation (ERC) and Emotion Inference in Conversation (EIC), we comprehensively assess its emotion perception capabilities. We conduct extensive experiments, and IoECoT outperforms other baselines on multiple Large Language Models (LLMs) and four datasets, demonstrating that our approach enhances LLMs’ emotion perception performance in zero-shot setting. Code is released at: https://github.com/betterfly123/IoECoT .