Overcoming Rigid and Monotonous: Enhancing Knowledge-Grounded Conversation Generation via Multi-granularity Knowledge

Xingsheng Zhang, Yifan Deng, Yue Hu, Yunpeng Li, Ping Guo

Abstract: Knowledge-grounded conversation (KGC) shows great potential in building an engaging and reliable chatbot, in which knowledge-aware generation is a key ingredient in it. Traditional methods, which usually generate responses based on sentence-level knowledge, are unable to locate the suitable knowledge pieces exactly, leading to insufficient utilization of knowledge and monotonous and rigid responses. In this paper, we propose a Multi-granularity Knowledge-aware Adaptive Generation (MKAG) model, which can adaptively select knowledge units from coarse to fine-grained based on various semantic segments in the response, as well as generating knowledgeable and informative responses. Specifically, an expandable knowledge supporter is devised to split knowledge units of various granularities according to semantics for positioning knowledge precisely. Besides, we design a segmentation-based decoding method, which is capable of generating multiple semantic segments sequentially, and each semantic segment can adaptively select the corresponding granularity of knowledge units for the generation. Experimental results show that MKAG significantly outperforms state-of-the-art methods in terms of both automatic and human evaluations, indicating that it can alleviate the issue of monotonous and rigid replies effectively.