PersonaGAN: Personalized Response Generation via Generative Adversarial Networks

Current personalized dialogue systems do not thoroughly model the context to capture richer information, and still tend to generate short, incoherent and boring responses. To tackle these problems, in this paper we propose a generative adversarial network model PersonaGAN for personalized dialogue generation. In addition to hierarchical modeling of context, we introduce a speaker-aware encoder in the generator to capture richer context information. Besides, we apply adversarial training to personalized dialogue generation task via using a transformer-based matching model as discriminator. The discriminator could give higher rewards for the responses which look like human written and lower rewards for machine generated responses. Such training strategy encourages the generator to generate responses which are grammatically fluent, informative and logically coherent with context. We evaluate the proposed model on a public available dataset and yield promising results on both automatic and human evaluation, which show that our model can generate more coherent and personalized responses while ensuring fluency.