Topic-level Decomposition Of Knowledge Graphs For Multi-turn Dialogue Generation

The knowledge graph-based dialogue generation task aims to use knowledge to generate responses with better context coherence and diversity.The researchers’ exploration and discovery of this task has greatly improved the quality of dialogue generation and enhanced the sense of experience of humancomputer interaction.However,there are still two challenges to be solved in this task.One is that the huge amount of information in the knowledge graph(KG)makes it difficult for the model to obtain the knowledge information related to the topic of the dialogue.The other is that the long series of historical information contained in multiple turns of dialogue makes it difficult for the model to extract topic information related to the whole dialogue.The existing methods deal with the knowledge graph based on the global KG,without considering the semantic information of the local topic level of KG,and do not effectively integrate the long conversation context and KG information into the dialogue generation.Therefore,the dialogue responses generated by these models have the characteristics of generality,uncontrollability and incoherence.In order to solve the above problems,we propose a topic-level knowledgeaware dialogue generation model and improves the technical methods of information acquisition and fusion between KG and multi-turn dialogue.The main achievements and conclusions are as follows.(1)A topic-level knowledge-aware dialogue generation model is proposed.Specifically,the model first divides a given KG into a set of topic-level subgraphs,each of which captures a semantic component of the input KG.This cannot be achieved by the current method based on multi-hop segmentation.In addition,the model also designs a novel topic layer subgraph attention network to integrate the subgraph and the previous turns of dialogue,and decode the fused subgraph and the current turn of dialogue to get a reply.Through the above two algorithms,the model can pay attention to different topic components of KG,extracting knowledge information related to historical dialogue,and generate responses with consistent contextual topics and diverse information.(2)Experiments are carried out on the two datasets of Du Rec Dial and Kd Conv.And a large number of evaluation and analysis are carried out at the same time,including the comparative experiment and evaluation of the whole model and KG segmentation methods,the ablation experimental analysis of each module and case analysis.The results show that,compared with strong baselines,the model achieves the best results on the metrics of fluency,consistency,diversity and knowledge-relevance.This further proves that the topic-level segmentation of the KG helps the model to better understand and obtain the information of the KG,while separate modeling history dialogue and the current turn of dialogue can better fuse the topic information in the dialogue.Therefore,the model generates a response with consistent context and knowledge information.