Generate Modern Chinese Poems from News Based on Text Style Transfer Using GAN

Guan-Fu Peng, Yi-Shian Yang, Chen-Yu Tsai, Von-Wun Soo

Abstract: In this paper, we investigate techniques that can transfer a news story into a poem. We train cycle-GAN that can conduct text style transfer from news style to poem style even lack of parallel corpus. We compare teacher forcing and free-running modes of training as well as different attention mechanisms in the GAN and cycle-GAN architectures. We found that there is a trade-off between degree of style transfer and content preserving that can be controlled by the ratio of reconstruction and transfer using different training modes of the discriminator and the generator. We show that both GAN and cycle-GAN can be trained to convert news into poems to some extent using non-parallel corpus.