Research On Knowledge Graph Completion Based On Graph Convolutional Neural Network

As the latest achievement of knowledge engineering,knowledge graphs(KG)are widely used in many applications of artificial intelligence.The effects of the applications depends on the data amount and data quality.The research of knowledge graph completion is helpful to improve the quality of the data.The research of knowledge graph completion achieve completing knowledge base with the internal information.Nowadays,the internal and overseas researches of knowledge graph completion mostly focus on knowledge units.In order to predict missing relations between nodes in KG,these studies firstly learn the semantic and structural features to obtain embedded vectors,and then judge potential relations between entities in theory of similarity to achieve completing knowledge base.The theories of the existing approaches are mostly in the perspective of 3-tuples,because as the basic unit in KG 3-tuples reflect the connections between entities and relations.However,these approaches rarely involve the information from higher dimension of KG,such as multiple relations in relational paths,structural information and so on.Graph convolutional neural networks(GCN)have been used in various graph representation learning,which adopt message passing mechanism to realize global feature learning by local iterations.There are some troubles applying GCN to knowledge graph completion:The features of nodes are attenuated in the progress of message passing;Knowledge graphs belong to heterogeneous graphs,and relations contain direction and semantic information.In order to solve the problems mentioned above,this paper improves graph convolutional networks to apply to the task of knowledge graph completion.The main works in this paper are as follows:(1)Aiming at the problem of insufficient utilization of relational paths in knowledge graph,a knowledge graph embedding model based on the hierarchical attention is proposed by improving the graph convolutional networks.This work introduces the hierarchical attention mechanism on KBGAT to learn the features of entities and relations at different depths.Firstly,improved by Random Walks,a relational path sampling strategy is designed to obtain the relational path samples for training.Then sampled paths are transformed into a constructed triple in theory of translation to realize deep knowledge representation.At last,the hierarchical attention encoder captures the internal features of multiple hops from the knowledge base.ConvKB model is chosen as the decoder to evaluate the similarity of the triples.The experimental results on WN18 RR,FB15k-237,UMLS,Kinship and NELL-995 datasets show that proposed approach effectively takes use of hierarchical information and outperforms related works especially in the knowledge graphs containing much structural information.Aiming at poor performance of hierarchical attention mechanism in dealing with many-to-many multiple relations,a strategy of relational path attenuation is proposed.The strategy of relational path attenuation adopt attenuation weight to differ many-to-many multiple relations.The attenuation weights are based on the frequency of relations in the sampled relational paths.The experimental results on the WN18 RR,FB15k-237,UMLS,Kinship and NELL-995 datasets show that the hierarchical attention completion approach with the strategy of relational path attenuation has a better effect in link prediction experiment,which verifies the effectiveness of the optimization strategy.(2)Aiming at the problem of insufficient utilization of time dimension by dynamic knowledge graph completion approach,a temporal aware completion approach is proposed.This completion approach is under the perspective of the 4-tuples.Temporal aware encoder improves the graph attention networks to extract features among entities,relations and times,and temporal convolution decoder evaluates the similarity of the 4-tuples embedded vector after feature extraction.Experimental results on ICEWS14,ICEWS05-15,Wikidata12 k and YAGO11 k datasets demonstrate the effectiveness of temporal aware completion approach.Meanwhile ablation experiment is conducted to verify the effectiveness of encoder and decoder respectively.