Research On Semantic Knowledge Reasoning Technology In Intelligent Dialogue System

With the progress of science and technology,as well as the development of artificial intelligence industry,intelligent dialogue system has been widely studied.Because the traditional intelligent dialogue system only pays attention to the semantic relevance among the dialogue content and ignores the mining of logical features,the generated response often has logical errors,which seriously affects the user experience.In order to deal with the above shortcomings,this paper explores the semantic knowledge reasoning technology in intelligent dialogue system,and analyzes the challenges in the following scenarios:intelligent dialogue question answering algorithm based on knowledge graph,response generation algorithm based on dialogue history.Specifically,the challenges faced by the current intelligent dialogue reasoning technology in the above two scenarios are as follows:(1)In the reasoning process,due to the static representation of relation information and the lack of reasoning direction,the relation information and direction information contained in the knowledge graph are not fully mined;(2)Due to the limitation of the graph convolutional network,it is difficult to model the long-distance node relation in the knowledge graph,resulting in the insufficient mining of the information in the knowledge graph,which limits the reasoning ability;(3)With the dialogue history,when the intelligent dialogue system generates the response,it pays more attention to calculate the similarity between the response and the dialogue history,which ignores the inference of the logical consistency between the response and the dialogue history.To overcome the above challenges,in this paper,we study the knowledge graph dialogue reasoning technology based on enhanced relation and direction representation,the knowledge graph dialogue reasoning technology based on modeling the long-distance node relation,and the multi-turn dialogue reasoning technology based on multi view fine-grained difference perception.The main contributions of this paper are summarized as follows:(1)Knowledge graph dialogue reasoning technology based on enhanced relation and direction representation.In the knowledge graph based dialogue reasoning technology,given the question-related subgraph,the graph convolutional reasoning algorithm mainly focuses on how to update the node features with the aggregation of neighbor information,ignoring the further mining of the representation ability of relational features,resulting in the inability of relational features to capture the rich structured information of different subgraphs.In order to solve the above challenge,this paper designs an algorithm to nodize the relation edges in the subgraph,so as to update the relation representation features by the aggregation of neighbor information,and finally obtain the relation features,which capture the subgraph structure information.At the same time,the arbitrariness of reasoning direction in the graph convolutional reasoning makes it contrary to the directional reasoning reality of human reasoning,which limits the reasoning ability in the knowledge graph.To solve the above problems,this paper proposes a relation direction conversion algorithm with seed node as the core,which is used to introduce direction information into reasoning,so that the information propagates along the seed node to nodes far away from the seed node.By imitating human behavior,this paper integrates the direction information of reasoning into the graph convolutional algorithm,so as to effectively improve the reasoning ability of the dialogue system and enhance the accuracy of answer prediction.(2)Knowledge graph dialogue reasoning technology based on modeling the long-distance node relation.In the knowledge graph,the reasoning technology based on graph convolutional algorithm updates and infers the information through the aggregation of neighbor nodes,which proves the importance of information propagation between nodes to reasoning.However,because the graph convolutional algorithm is a special form of Laplacian smoothing,in order to avoid the transition smoothing of node features,the graph convolutional algorithm can not overlay too many convolutional layers,resulting in the information between long-distance nodes can not be propagated,which limits the ability to mine the information contained in the knowledge graph in reasoning.In order to solve the above problems,this paper proposes a new model GlobalGraph,which aims to model the long-distance node relation from two perspectives while stacking less convolutional layers:1)Global type label similarity between nodes:GlobalGraph designs the global type label assignment algorithm,assigns the global type label to each central node based on the structured information of neighbor relation,and models the long-distance node relation through the label similarity algorithm;2)Relevance between nodes and the input question:GlobalGraph designs a semantic similarity algorithm to model the correlation between each node and the input question.Finally,a question aware dynamic graph is modeled under nodes with high correlation,which aims to model the long-distance node relations.Finally,the dialogue system propagates information and reasoning between long-distance nodes which have the relations based on the above two perspectives.(3)The multi-turn dialogue reasoning technology based on multi view fine-grained difference perception.When generating responses based on dialogue history,the generation model pays attention to the semantic correlation between candidate responses and dialogue history,ignores the modeling and reasoning of logical consistency between them,and limits the performance of the generation model.In order to solve the above problems,inspired by human focusing on perceiving fine-grained details in reasoning,this paper proposes a multi perspective fine-grained difference perception reasoning algorithm to enhance the reasoning ability of dialogue generation model.Specifically,this paper models the logical consistency inference features from multiple perspectives:infer the consistency difference between candidate responses,infer the consistency difference between candidate responses and dialogue history,and infer the consistency difference between candidate responses and the current speaker’s own dialogue history.At the same time,in order to capture the fine-grained logical inconsistency clues in the text,this paper designs a bidirectional attention mechanism based on word level in each perspective,which aims to perceive the fine-grained logical differences.Finally,this paper sorts the candidate replies through the consistency features.