Research On Knowledge Graph Completion Model Based On Deep Learning

The knowledge graph is a multi-relational graph used to represent different types of entities,and it has gradually become one of the key tools for knowledge management.However,there is a common problem of missing entities or relationships in the process of building knowledge graphs,so knowledge graph completion techniques have been proposed to complement the missing information in knowledge graphs,and many knowledge completion models have been constructed by using various representation techniques.With the development of deep learning,more and more knowledge graph completion models begin to use deep learning technology to improve the model’s completion capabilities.For example,both completion models using capsule networks and those incorporating additional information such as entity descriptions have achieved good results.Although the capsule network completion model solves the problems of spatial invariance and low coding efficiency in the convolutional neural network model,it is difficult to capture the potential dependencies between entities.And the use of Softmax in dynamic routing will increase the error of model prediction.In addition,both the models that use entity information and those that use ordered relational path information for knowledge graph completion use only a single additional information,and fail to effectively use the additional information in multiple knowledge maps for knowledge representation,thus affecting the ability of knowledge graph completion.Aiming at the problems existing in capsule networks,we propose a relational memory-based decentralized dynamic routing capsule network model(RDMCaps E).The model can use the relational memory network to effectively capture the potential dependencies of triples to generate embedding vectors,and then use the multi-scale capsule network model for feature processing,and use our improved disperse dynamic routing in the dynamic routing part of the capsule network model to improve the capsule network performance.Experiments have proved that introducing the relational memory network into the capsule network model can effectively improve the ability of knowledge graph completion,and the dynamic routing can improve the model more significantly in the more relevant datasets.To address the problems of a single additional information model,we propose a knowledge graph completion model(OPDRL)that integrates ordered relational paths and entity description information.The model can efficiently obtain the ordered relational path representation using two-level pooling,then represent the entity description information using Transformer and relational attention mechanisms,and finally we fuse these two parts together with the Trans R model that can effectively handle complex relationships in the knowledge graph.Experiments show that the OPDRL model trained with the fusion of ordered relational path and entity description information has superior performance compared with the model only using any one of the additional information.Moreover,in Transformer,we use power normalized PN instead of layer normalized LN for normalization,and ablation experiments show that it can effectively improve the ability of knowledge graph completion.