Improvement on Knowledge-grounded Dialogue Generation with Continuous Prompt Tuning

In this paper, a two-stage knowledge-grounded dialogue generation framework is proposed, which addresses the problem of generic and uninformative dialogue generation. An integrated system of continuous prompt tuning and response selection is introduced, which effectively filters out invalid results and improves the quality of the generated answers. The number of training parameters is significantly reduced by our training method, making it more efficient compared to existing models. The effectiveness of our approach is demonstrated on a recently proposed dialogue dataset using various metrics.