PROTOTYPE-TO-STYLE: Dialogue Generation With Style-Aware Editing on Retrieval Memory

The ability of dialogue systems to express pre-specified style during conversations has a direct, positive impact on their usability and user satisfaction. While it attracted much research interest, existing methods often generate stylistic responses at the cost content quality. In this work, we introduce prototype-to-style (PS) framework tackle challenge generation. proposed first exploits an Information Retrieval (IR) system extracts response prototype from retrieved response. A generator then takes desired as input produce high-quality To effectively train model imitate real testing environment, new style-aware learning objective denoising strategy. Results three benchmark datasets (gender, emotion, sentiment) two languages demonstrate that approach significantly outperforms baselines both in terms in-domain cross-domain evaluations.