Text Generation Based on a Japanese Data Set Using Generative Adversarial Networks

Automatic text generation using generative adversarial networks （GANs） such as SeqGAN, TestGAN, and LeakGAN has attracted substantial attention. There are numerous studies regarding the automatic generation of English and Chinese texts using Gans or related algorithms. Moreover, methods to generate sentences automatically based on the speeches delivered by the former US president Barack Obama, captions of images, or Chinese poems have been realized. However, there are only a few reports regarding the generation of text from Japanese sentences. In this study, to investigate the differences in generating text in different languages, we generate sentences form the Japanese novel Botchan," which was written by Soseki Natsume, using maximum likelihood estimation, SeqGAN, TextGAN, and LeakGAN. Furthermore, we evaluate the generated text using the metrics of negative log-likelihood loss, 2-gram BLEU, and embedding similarity. Consequently, it is concluded that LeakGAN can generate the most natural text and that, for unknown reasons, TextGAN does not perform well with regard to automatic text generation."journal articl