Abstract: A common problem in the generation of empathic dialogues is that the dialogue model tends to generate general responses, such as " I don’tknow " , which are common but meaningless responses in the corpus. A generic response can reply to any conversation above. In order to alleviate this problem, a conditional variational autoencoding framework is introduced into the decoder to expect the generated sentences to have text diversity. In order to better understand the speaker’s emotion and semantics, in the encoder module, common-sense reasoning generation module COMET and the sentiment dictionary were used to enhance the semantic information and sentiment information in the dialogue. Therefore, the VT-CEM model was proposed by using the COMET encoder and the variational decoder jointly. Experiments on the EmpatheticDialogues dataset show that the VT-CEM model can produce higher fluency and richer text diversity relative to multiple baselines.