Human-machine dialogue is the hot spot of current research. In this paper, we proposed an end-to-end dialogue system based on external knowledge, which realized active guidance and topic transfer in multiple rounds of dialogue. Our system is built on the pointer generator model so that the output token in the response can be generated or copied from the conversation history or background knowledge according to a trainable action probability distribution. At the same time, with the data processing and optimization of the model structure, the designed system is capable of generating high quality responses. In the 2019 NLP Language and Intelligence Challenge, our proposed dialogue system ranked third in the automatic evaluation, and ranked fifth in the manual evaluation.