Read Then Respond: Multi-granularity Grounding Prediction for Knowledge-Grounded Dialogue Generation

Yiyang Du, Shi-Wei Zhang, Xianjie Wu, Zhao Yan, Yunbo Cao, Zhoujun Li

Abstract: Retrieval-augmented generative models have shown promising results in knowledge-grounding dialogue systems. However, identifying and utilizing exact knowledge from multiple passages based on dialogue context remains challenging due to the semantic dependency of the dialogue context. Existing research has observed that increasing the number of retrieved passages promotes the recall of relevant knowledge, but the performance of response generation improvement becomes marginal or even worse when the number reaches a certain threshold. In this paper, we present a multi-grained knowledge grounding identification method, in which the coarse-grained selects the most relevant knowledge from each retrieval passage separately, and the fine-grained refines the coarse-grained and identifies final knowledge as grounding in generation stage. To further guide the response generation with predicted grounding, we introduce a grounding-augmented copy mechanism in the decoding stage of dialogue generation. Empirical results on MultiDoc2Dial and WoW benchmarks show that our method outperforms state-of-the-art methods.