Research On Improved Intelligent Generative Dialogue Algorithm Based On Knowledge Graph

Dialogue system is a research hotspot in the field of human-computer interaction.It is widely used in online customer service,question and answer and other fields.Due to the lack of massive knowledge support and constraints on the search scope of question and answer solution space,the traditional generative dialogue system still has the following problems: The diversity index of generated response content is not high.The need s of users cannot be met adequately.Furthermore,there is a relatively large gap with the real response.To solve these problems,this thesis makes the following research:(1)By comparing the characteristics,advantages and disadvantages of closed domain and open domain dialogue systems,it is found that open domain dialogue system has wider application prospects.By comparing the retrieval and generative dialogue models,it is concluded that the generative dialogue system is more flexible.Therefore,this thesis chooses open domain generative dialogue model as our research target.(2)Firstly,we design three generative dialogue algorithms: Seq2 Seq based on GRU(Gated Recurrent Unit),based on GL?G and GL?G based on knowledge graph.Then we verify the promotion effect of Bi-directional GRU,dialogue target sequence and knowledge graph on the generative dialogue model by using these three generative dialogue algorithms respectively.In the first place,we propose to replace the ordinary GRU unit in the original Seq2 Seq baseline model with Bi-directional GRU unit.The model can generate coherent and contextual response content.The improved Seq2 Seq model can avoid generating incoherent response to a certain extent.Secondly,the dialogue target sequence module is introduced to enable the model to intelligently guide the direction of the dialogue according to the target sequence,which solves the problems of generating more general replies and falling into the dead cycle of the dialogue.Finally,aiming at the empty,single and lack of diversity of the generated response content,the knowledge graph is introduced into the dialogue model,which can provide more key knowledge and more key words in the process of generating the responding dialogue.The experimental result shows that the dialogue model with knowledge graph is better than GL?G model.The diversity index DIST-2 is improved by 4.0% with richer content generated by the model.Through the comparative analysis of the above three algorithms,it is concluded that GL?G algorithm based on knowledge graph has the best performance in various evaluation indexes.But there are still some problems,such as lack of constraints on the search scope of question and answer solution space and difficulty in capturing the deep intention behind users' discourse.(3)In order to solve the above problems,this thesis proposes to design an improved generative dialogue system based on knowledge graph.After designing and comparing a variety of improved algorithms,we finally choose to add two modules in the dialogue: dialogue scene and background information of characters participating.On the one hand,the introduction of the method of reducing the search scope makes the dialogue more targeted;on the other hand,the introduction of the method of providing background information makes the model better understand the users,so as to generate response more in line with the needs of users.The experimental results on the public Du Rec Dial data-set show that the UPST?G model built in this study is better than the original GL?G model based on knowledge graph.Hits@1and Hits@3,the indexes related to retrieval,both are increased by 0.6%.F1 and BLEU2,the indexes related to generation,are increased by 1.0% and 1.4% respectively.At the same time,the PPL index is decreased by 0.3%.Knowledge P / R / F1,the indexes related to knowledge utilization,rise by 1.5%,2.9% and 4.2% respectively.In addition,through the mode of single round dialogue and multi round dialogue,the response content generated by the dialogue model is analyzed,which proves that the response content generated by the algorithm is of higher quality.To sum up,the UPST?G algorithm model based on knowledge graph is more effective than the traditional generative dialogue algorithm.The research proves the feasibility and effectiveness of the algorithm in the generative dialogue system.