Grounded Conversation Generation as Guided Traverses in Commonsense Knowledge Graphs

Abstract

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.

References in corpus (11)

• Neural Machine Translation by Jointly Learning to Align and Translate
• Sequence to Sequence Learning with Neural Networks
• A Neural Conversational Model
• The Curious Case of Neural Text Degeneration
• Unified Language Model Pre-training for Natural Language Understanding and Generation
• ConceptNet 5.5: An Open Multilingual Graph of General Knowledge
• A Survey on Dialogue Systems: Recent Advances and New Frontiers
• Topic Aware Neural Response Generation
• Neural Text Generation with Unlikelihood Training
• Generating Informative and Diverse Conversational Responses via Adversarial Information Maximization
• Flexible End-to-End Dialogue System for Knowledge Grounded Conversation