The Study On Target-guided Open-domain Conversation Based On GCN

With the continuous improvement of computing power,the neural network based conversational system using large-scale data training has made great progress,but the data-driven approach also brings many problems.Since the dialogue of the chitchat dataset usually come from different users,the dialogue topics in the dataset are not coherent,and after the end of one topic,they jump to other unrelated topics,resulting in generating irrelevant response of the conversational system.To solve these problems,Target-Guided open-domain Conversation(TOC)integrates goal realization and conversation strategy into Open-domain Conversational system by controlling topic consistency.First,the target of the conversational system is defined.The dialog policy then take part in the dialog generation process.This thesis uses deep learning technology to study TOC tasks and analyzes the limitations of existing methods.The main research contents are as follows:(1)For the problems that the current TOC method only uses most recent historical conversation and lacks effective information,it is not enough for model to fully understand the current topic and the accuracy of topic prediction is low.A topic-graph enhanced target-guided method(TGT)based on undirected graph is proposed.Text graphs constructed by adjacent topics are introduced,and the information of neighbor nodes on the text graphs is used to supplement the semantic information of current topics.The graph convolutional neural network is used to encode topics to enhance topic semantics,so that the model can learn the deep correlation between topics.After each prediction task,the attention weight of the topic related to the current topic is calculated to update its representation,so as to obtain the topic closer to the target.(2)In order to further strengthen the semantic of important information in conversation and enable the model to screen useful information,a target-guided conversational model(SF-TOC)based on semantic enhancement and feature fusion is proposed.The dialogue content is divided into different granularity,and the distance weight is introduced into the coarse-grained discourse modeling to integrate the long-distance discourse information.In the fine-grained word modeling,a subgraph of related topics is constructed to mitigate the influence of irrelevant topics.In addition,the framework of sequence matching and topic matching is used for response retrieval.In experimental validation,we demonstrated on Conv AI2 and Reddit datasets that TGT and SF-TOC can construct conversation sequences that are target guided and achieve target topic with a high achievement rate.