Mitigating Negative Style Transfer in Hybrid Dialogue System

As the functionality of dialogue systems evolves, hybrid that accomplish user-specific goals and participate in open-topic chitchat with users are attracting growing attention. Existing research learns both tasks concurrently utilizing a multi-task fusion technique but ignores negative transfer phenomenon induced by unique textual style differences. Therefore, contrastive learning based on latent variable model is used to decouple various genres space. We devise supervised self-supervised positive sample constructions for diverse datasets. In addition, capitalize information contained decoupled variables, we employ prefix incorporates variables further control generation responses varying styles. performed extensive experiments three datasets, including dataset two task-oriented The experimental results demonstrate our method can mitigate issue achieves state-of-the-art performance multiple