A dual adversarial structure of generative adversarial network for nature language generation

Purpose Recently, due to the practicability in several domains, generative adversarial network (GAN) has successfully been adopted in the field of natural language generation (NLG). The purpose of this paper focuses on improving the quality of text and generating sequences similar to human writing for several real applications. Design/methodology/approach A novel model, GAN 2 , is developed based on a GAN with dual adversarial architecture. We train the generator by an internal discriminator with a beam search technique to improve the quality of generated sequences. Then, we enhance the generator with an external discriminator to optimize and strengthen the learning process of sequence generation. Findings The proposed GAN 2 model could be utilized in widespread applications, such as chatbots, machine translation and image description. By the proposed dual adversarial structure, we significantly improve the quality of the generated text. The average and top-1 metrics, such as NLL, BLEU and ROUGE, are used to measure the generated sentences from the GAN 2 model over all baselines. Several experiments are conducted to demonstrate the performance and superiority of the proposed model compared with the state-of-the-art methods on numerous evaluation metrics. Originality/value Generally, reward sparsity and mode collapse are two main challenging issues when adopt GAN to real NLG applications. In this study, GAN 2 exploits a dual adversarial architecture which facilitates the learning process in the early training stage for solving the problem of reward sparsity. The occurrence of mode collapse also could be reduced in the later training stage with the introduced comparative discriminator by avoiding high rewards for training in a specific mode. Furthermore, the proposed model is applied to several synthetic and real datasets to show the practicability and exhibit great generalization with all discussed metrics.