Exploiting Persona Perception for Diverse Generation from Limited Personalized Data

Persona-based dialogue generation aims to produce more various and consistent conversations based on correlated personas. Existing personalized dialogue systems still have some issues, involving the responses generated from the model are boring and inconsistent with pre-defined persons. In a personalized dialogue system, preserving a consistent persona is fundamental in engaging users. In this paper, we devise an innovative structure Persona Interactivate Generator (PIG) to model the correlations between query, persona, and responses for persona perception. A detector is then employed for consistent understanding, encouraging the dialogue model to produce consistent responses. Additionally, the extra explicit persona is hard to obtain in a practical scenario, and we proposed an architecture Persona Generative Adversarial Network (Per-GAN) for persona enhancing. After that, we regard personalized dialogue generation as a reinforcement learning problem and gather the various rewards from the Per-GAN and consistent detector to lead the PIG to respond better. Last but not least, comprehensive experiments and analyses on persona chat dataset ConvAI2 from varying metrics demonstrate our method’s effectiveness.