Developing an empathetic dialogue system has been facing challenges for years. Although most of the responses generated from the existing empathetic dialogue systems are to some extent grammatical and emotional, they are still far from human standards. The system-generated responses are usually uncorrelated to the user's emotional contexts. In this paper, we propose an approach to automatic extraction of the emotional keyphrases from the context, then the emotional keyphrases are fed into the Conditional Transformer language model as the conditional information. We also design a loss function called Keyphrase Correlation (KPC) to guide the model to generate responses with relevant emotional keyphrases. As a result, the diversity of the generated responses is highly enhanced. The experimental results also verified the effectiveness in terms of empathy capability and emotional correlation in the user-system conversation.