Grounded Conversation Generation as Guided Traverses in Commonsense Knowledge Graphs
This addresses the challenge of coherent and meaningful dialogue generation for AI systems, representing an incremental improvement with a novel method for a known bottleneck.
The paper tackles the problem of generating more semantic and informative responses in conversations by explicitly modeling conversation flows using commonsense knowledge graphs, resulting in ConceptFlow outperforming previous knowledge-aware and GPT-2 based models while using 70% fewer parameters.
Human conversations naturally evolve around related concepts and scatter to multi-hop concepts. This paper presents a new conversation generation model, ConceptFlow, which leverages commonsense knowledge graphs to explicitly model conversation flows. By grounding conversations to the concept space, ConceptFlow represents the potential conversation flow as traverses in the concept space along commonsense relations. The traverse is guided by graph attentions in the concept graph, moving towards more meaningful directions in the concept space, in order to generate more semantic and informative responses. Experiments on Reddit conversations demonstrate ConceptFlow's effectiveness over previous knowledge-aware conversation models and GPT-2 based models while using 70% fewer parameters, confirming the advantage of explicit modeling conversation structures. All source codes of this work are available at https://github.com/thunlp/ConceptFlow.