Research And Applications Of Knowledge Graph Embedding Based On Convolutional Neural Network

Knowledge Graph Embedding(KGE)aims to learn the distributed representations of entities and relations in the Knowledge Graph(KG),thereby making the knowledge stored in the graph easier to be applied to various downstream tasks,such as knowledge graph completion,knowledge reasoning,and retrieval.In the past few years,many research achievements have been proposed,such as embedding technology based on translation distance and semantic matching.However,most of these conventional models use linear methods to model the association relationship in knowledge(or called fact).The model structure is too shallow,and its expressiveness is therefore limited.Recently,Convolutional Neural Network(CNN)has been applied to KGE.The CNN-based models have both parameter efficiency and expressiveness and achieved excellent performance in link prediction tasks.However,the existing embedding methods ignore the correlation among channels in CNN,and fail to further improve the performance of the model.The following research are included in this paper: 1)We analyze the existing deep learning-based embedding models,and propose a CNN-based model,ConvD,to learn the local and global relationship and integrate cross channel information in triples by catching the Dimension-wise interactions.Then the performance of link prediction task executed by ConvD on the public datasets WN18RR and FB15k-237 is compared with the previous research work.2)To enable the embedding technology to be applied to the Open-World Knowledge Graph Completion(Open-World KGC),we studied the combination of embedding technology and related textual information.Based on the Open-World Extension,Enhanced Open-World Extension(EOWE)is proposed.The performance of extended model ConvD-EOWE is verified by executing the Open-World link prediction task on dataset FB15k-237-OWE and DBpedia50k.3)To satisfy the upgrade of intelligence level of the city,a smart city KG for firefighting and elderly care,called SCKG,is constructed.ConvD is applied to learn the representations of entities and relations in SCKG as the basis for subsequent applications.Link prediction task is executed on the corresponding dataset SCKG159,which validates the performance of ConvD on real-world dataset.The experimental results show that,to the best of our knowledge,ConvD obtains the best MR score on the FB15k-237 and the best MRR and Hits@N score on the WN18RR.Compared with using OWE to extend ConvD,EOWE can improve all the metrics on the FB15k-237-OWE and DBpedia50k,especially in the MR score by 11.7%and 12.3% respectively.Compared with other classical embedding models,ConvD has better MRR,Hits@3 and Hits@10 score on SCKG159.