Towards few-shot mixed-type dialogue generation

Building an agent capable of conducting both open-domain and task-oriented dialogues, known as mixed-type dialogues, has been an enduring challenge for the AI community. Previous approaches have focused on constructing large-scale human-annotated datasets for training models. However, annotating these datasets is expensive and hinders the practical application of these models. This paper identifies a novel challenge, few-shot mixed-type dialogue generation. To address this challenge, we first present a pre-trained dialogue generation framework with modular-based architecture and prompttuning component. Additionally, we collect a mixed-type dialogue dataset that combines persona-chat with conversational recommendation or task-oriented dialogues within a single dialogue session. Specifically, the modular-based architecture allows us to easily incorporate more supervised signals and human-annotated information, thereby facilitating the learning of sessionlevel dialogue logic. We pre-train this dialogue generation framework using multiple external datasets and then fine-tune it on the mixed-type dialogue dataset we collected. Experimental results demonstrate that the three key designs — modular-based architecture, prompt-tuning component, and model pre-training — significantly enhance the performance of this framework compared to state-of-the-art baselines.