Research On Neural Network-Based Knowledge Graph Reasoning

Knowledge graphs(triplet collection composed of entities and relations),as a structured and semantic representation of knowledge,have attracted massive attention from researchers.However,due to the omission of knowledge acquisition,existing knowledge graphs are confronted with issues of completeness and newly-added knowledge.Knowledge graph reasoning,as an important means of complementing and updating knowledge graphs,has played an important role in many education fields such as learning resource recommendation,knowledge search,intelligent Q&A,and machine translation.And it has become one of the cores to promote the development of Internet and artificial intelligence education.Knowledge graph reasoning aims to infer new knowledge or identify wrong knowledge based on existing knowledge.The main idea is to convert the traditional reasoning process into semantic vector calculation based on the distributed representation of entities and relations.Existing knowledge graph reasoning models can be roughly categorized into translational distance-based models,semantic matching-based models,and deep learning-based models.Among them,the deep learning-based models utilize the different neural network to extract the semantic features of entities and relations,and then used for the reasoning of knowledge triplets.Because of its advantages in semantic feature learning and extraction,it has become the mainstream research direction of knowledge graph reasoning.In this context,this paper explores the adoption of the neural network to solve the challenges in knowledge graph reasoning.The main contents and contributions of this paper are summarized as follows:Aiming at the challenge of complex relation reasoning in knowledge graphs,a complex relation reasoning model based on the convolutional neural network is proposed.There are complex relations such as one-to-many,many-to-one,and many-to-many in the knowledge graphs.The performance of the existing knowledge graph reasoning model needs to be improved when dealing with complex relations.To address this issue,this paper proposes a dynamic multi-scale convolutional neural network(M-DCN),which dynamically generates relation-specific convolution kernels,so that the neural network can dynamically obtain the features of entities,and then used for knowledge graph reasoning.Besides,M-DCN will generate multi-scale convolution kernels to extract different aspects of the semantic features between entities and relations.Experiments show that the proposed model effectively improves the reasoning effect of complex relations in the knowledge graphs.Aiming at the challenge of few-shot knowledge graph reasoning,a few-shot knowledge graph reasoning model HRAN based on the heterogeneous graph neural network is proposed.The traditional knowledge graph reasoning model requires a large number of samples for training to accurately learn the semantic features of entities and relations.To improve the performance of the few-shot knowledge graph reasoning,this paper proposes a heterogeneous graph attention network(HRAN),which can aggregate the features of neighbor nodes according to the entity attributes,and then used to better represent the features of the central node.Besides,HRAN utilizes the attention mechanism to obtain the importance of different relation-paths and selectively aggregates important neighbor node features.Experiments on several datasets prove that the proposed model significantly improves the performance of knowledge graph reasoning compared with previous models.Aiming at the challenge of polysemy phenomenon in the knowledge graphs,an interactive knowledge graph learning reasoning model IE_RCN based on recalibration neural network is proposed.It can observe that most of the knowledge graphs cannot be accurately constructed,which leads to the polysemous phenomenon of entities and relations.To solve this issue,this paper proposes a recalibration convolutional network(IE_RCN)based on interactive knowledge graphs learning.IE_RCN aims to establish cross-semantic interactions from relation to entity and from entity to relation,so that entities and relations will be in different vector subspaces when involving different triplet predictions.Besides,this paper proposes a recalibration mechanism to reconstruct the feature map,thereby selectively extracting important features and suppressing useless features.The experimental results show that the proposed reasoning model has achieved a consistent improvement in various evaluation metrics.