Empathy is a core capability of human-like chatbots, aiming to perceive the speaker’s situation and emotion in a conversation and respond appropriately. It encompasses both affective and cognitive dimensions. Existing methods often rely on the speaker’s coarse-grained emotion label of dialogue to guide response generation, overlooking the fact that emotions of both participants in a conversation are dynamically evolving, and influenced by their own emotional state and interaction with others. Furthermore, they lack the ability to selectively incorporate external knowledge, leading to the noise injection and thereby impairing the model’s empathetic ability. In this paper, we propose the EEKR model to address these issues. In terms of affection, we adopt the method that combines independent and interactive encoding of the emotional information from both participants in a conversation to capture the fine-grained emotional changes between them, thereby more accurately predicting the emotion category of the response. Then, the predicted emotion label for response is used to guide response generation. In terms of cognition, we leverages the information content of utterance for content filtering, and effectively select commonsense knowledge through the interaction between utterance and external knowledge. Extensive experiments on the benchmark dataset EMPATHETICDIALOGUES demonstrate that our model outperforms current state-of-the-art empathetic dialogue models in both automatic and human evaluations.