We propose to combine the merits of template-based and corpus-based dialogue response generation by introducing a prototype-based, paraphrasing neural network, called P2-Net, which aims to enhance quality of the responses in terms of both precision and diversity.