1. Paper title
Grounded Conversation Generation as Guided Traverses in Commonsense Knowledge Graphs
2. link
https://www.aclweb.org/anthology/2020.acl-main.184.pdf
3. 摘要
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.
grounding to: 把…建立在…的基础上
4. 要解决什么问题
真人对话具有包含相关概念和多跳的特点。
5. 作者的主要贡献
ConceptFlow：利用常识知识图显式地构建对话流。
[?] 没看懂
把“潜在对话流”表示为“沿着常识关系遍历概念空间”
遍历由概念图中的图注意力引导，向概念空间中更有意义的方向移动，以生成更多的语义和信息响应。
6. 得到了什么结果
在Reddit conversations上优于knowledge-aware conversation models and GPT-2 based models
[?] knowledge-aware conversation models
参数减少70%。
[?] GPT-2 based models