Automatic generation of sentimental texts via mixture adversarial networks

Abstract Automatic generation of texts with different sentiment labels has wide use in artificial intelligence applications such as conversational agents. It is an important problem to be addressed for achieving emotional intelligence. In this paper, we propose two novel models, SentiGAN and C-SentiGAN, which have multiple generators and one multi-class discriminator, to address this problem. In our models, multiple generators are trained simultaneously, aiming at generating texts of different sentiment labels without supervision. We propose a penalty-based objective in generators to force each of them to generate diversified examples of a specific sentiment label. Moreover, the use of multiple generators and one multi-class discriminator can make each generator focus on generating its own texts of a specific sentiment label accurately. Experimental results on a variety of datasets demonstrate that our SentiGAN model consistently outperforms several state-of-the-art text generation models in the sentiment accuracy and quality of generated texts. In addition, experiments on conditional text generation tasks show that our C-SentiGAN model has good prospects for specific text generation tasks.