Towards Coherent and Captivating Topic Transitions in Knowledge-Grounded Conversations

Anonymous

Abstract: Knowledge-grounded conversations require skillful usage of knowledge to generate suitably diverse responses to keep user captivated while maintaining coherence to the dialogue context. However, current approaches that directly match knowledge with dialog context can result in capturing spurious correlations between knowledge and context, leading to either incoherent or mundane topic transitions in the generated dialogs that fail to engage.In this work, we introduce the Coherent and Captivating Topic Transition (C2T2) method to select the appropriate knowledge to be used in next response, resulting in topic transitions that are coherent to the ongoing conversations while providing adequate topic development for an engaging dialog.Our C2T2 employs transition-aware features designed to consider both historical contextual coherence as well as sequential topic development under a knowledge shifting constraint to select the next knowledge, thereby generating the response for an engaging conversation.We also designed a pointer network-based knowledge inference module to take into consideration of the relations among knowledge candidates during knowledge inference. Extensive experiments on two public benchmarks demonstrated the superiority of C2T2 on knowledge selection. Analysis on fine-grained knowledge selection accuracy also showed that C2T2 could better balance the topic adhesion and knowledge diversity in dialogs than existing approaches.

Paper Type: long