Research And Implementation Of Natural Language Dialogue Based On Knowledge Graph

With the breakthrough development of deep learning and natural language processing technology,dialogue systems have been widely used in scenarios such as intelligent customer service.The construction of generative dialogue systems often faces the problem of inaccurate generated responses due to lack of commonsense knowledge.The knowledge graph,as highly structured knowledge,can provide common sense background knowledge to the dialogue system,and promote semantic understanding and response generation during the dialogue process.However,in the existing dialogue systems based on knowledge graphs,the specific context is ignored in the process of knowledge encoding,which leads to the introduction of redundant knowledge,and it is difficult to make rational use of knowledge in the generation process to generate informative and diverse responses.Therefore,the core research problem of this paper is knowledge fusion and knowledge selection in dialogue systems.The main innovations and contributions are as follows:Firstly,for the task of knowledge-aware open-domain dialogue generation,existing studies have found that there are many redundant knowledge triples in the retrieved subgraphs,and most models lack modeling of the current conversation topic,resulting in generated responses that are irrelevant to the current conversation topic.In order to solve the above problems,the knowledge-enhanced encoder based on the graph attention mechanism and the knowledge selection module based on the prediction of topic fact.The model utilizes modeling conversation topics to improve the accuracy of using knowledge to generate responses Experiments show that the model outperforms current baseline models in knowledge selection.Secondly,the answer generation task based on knowledge graph is further studied.Existing methods ignore the relationship information in the question in the process of knowledge encoding,and the knowledge probability distribution will be put into the decoding process in a static form,making it difficult to generate questions containing correct entity or multi-entity answers.To solve the above problems,a knowledge encoding module based on relation prediction and a multi-layer decoding module based on dynamic knowledge selection are proposed.The model realizes the ability to dynamically select knowledge as the decoding state changes,improving the accuracy and richness of the answers generated by the model.The comparison experiment with the baseline model on the public dataset proves the effectiveness of the model design.Finally,based on the above model and related data resources,a dialogue question answering system integrating open domain chat dialogue and knowledge graph question answering is designed and implemented.Combined with the actual project needs and characteristics,it provides users with a simple and easy-to-operate dialog question and answer interactive web page,which verifies the practical application value of the natural language dialog technology based on knowledge graphs.