Research On Key Techniques Of Knowledge Reasoning For Knowledge Graph Completion

Knowledge graph is an efficient knowledge representation model,and its concept was formally proposed by Google.The original intention of the knowledge graph is to improve the capabilities of search engines.With the development of artificial intelligence technology,the knowledge graph has supported many practical applications,including knowledge question answering,recommendation systems and other natural language understanding tasks.However,large-scale knowledge graphs such as YAGO,DBPedia,and Freebase that contain billions of facts cannot avoid the problem of missing relations or nodes.This problem leads to the study of knowledge graph reasoning tasks oriented to the completion of knowledge graphs.On the one hand,knowledge graph reasoning can use known facts to reason about unknown or potential facts to improve and expand the knowledge graph;on the other hand,it supports downstream tasks,such as intelligent question answering,recommendation systems,etc.,to obtain more accurate and faster responses.Therefore,knowledge graph reasoning has important theoretical and practical significance for the construction and application of knowledge graph.This paper proposes a reasoning model oriented to traditional knowledge graphs,and three reasoning models oriented to temporal knowledge graphs:(1)At present,knowledge graph reasoning have not allocated attention based on the target relation in the reasoning problem.Different types of relations of a given entity in the knowledge graph should be given different attention.In response to this problem,this paper proposed a target-oriented relation-oriented attention reasoning model,which pays more attention to the relation that matches the relation in the reasoning problem.This model pays more attention to the relation that matches the relation in the reasoning problem.In addition,we proposed a hierarchical attention mechanism with high-order propagation characteristics,which calculates the attention at the node level and the attention between subgraphs separately,which alleviates the over-smoothing problem to a certain extent.(2)Aiming at the current temporal knowledge graph reasoning research,only considering the problem of fusing the current timestamp information with the embedded representation of entities/relations,this paper proposed a temporal knowledge graph reasoning model based on GRU nesting,and the latest k timestamps are integrated into the embedded representation of entities/relations as a temporal series.This method first uses an inner GRU to process the time sequence to obtain integrated temporal information;then,the hidden state obtained by the inner GRU encoding and the time-independent embedding representation of the entity/relation are used as the input of the outer GRU to obtain the embedding representation after the entity/relation fusion temporal information;through this nested GRU,the quadruple in the temporal knowledge graph can be converted into a triple without losing the temporal information.(3)Aiming at the problem that the current temporal knowledge graph reasoning model cannot model the temporal combination relation,this paper proposed a temporal multi-relation reasoning model based on Householder transformation.First,perform Householder transformation on the embedded representation of time and relation in the temporal knowledge graph to obtain an orthogonal matrix about time and relation,and then regard the orthogonal matrix as the rotation of the head entity to the tail entity.Orthogonal matrix mainly changes the direction of the matrix through rotation,so it maintains the modeling ability of symmetric/antisymmetric relations,inverse relations and combination relations.This method overcomes the limitations of existing reasoning models and has the ability to learn and infer various relational patterns over time.(4)In view of the existing research on temporal knowledge graph reasoning,most of them are shallow and fast models,which may lead to insufficient feature extraction of knowledge graphs.This paper proposed a temporal knowledge graph reasoning model based on convolutional neural networks.The model uses 2D convolution and fully connected layers to model the interaction between input entities,relations,and temporal embedding representations.The tensor obtained after the initial embedding of entities,relations,and timestamps is cocated as the input of the model,and the features are extracted through the convolution,pooling and fully connected layers,and the output is the score of the fact.In summary,this thesis presents the key technology of knowledge reasoning for knowledge graph completion.It is significant to the theoretical research and practical applications on the construction and application of knowledge graphs.