Research On Knowledge Graph Completion Based On Multi-source Heterogeneous Information Fusion

In recent years,with the rapid development of information technology and the popularity of the Internet,the explosive growth of data has confronted people with a huge and scattered amount of information.In this context,knowledge graph is proposed and attracts much attention as a powerful tool for organizing and representing knowledge,and is widely used in such fields as smart search and smart Q&A.In the era of big data,a large amount of data is continuously generated,which leads to the existing knowledge graph information completeness is low and a large amount of potential knowledge is missing.In order to solve this problem,it is especially important to use the existing knowledge mining to get new knowledge,and the knowledge graph complementation technology comes into being.The hot research topic in recent years is to use knowledge representation learning to complement the knowledge graph,which can capture the complex semantic relationships between entities and relations and has better performance compared with traditional methods.However,the existing models have the following problems:(1)most knowledge representation models only focus on the structural information of the knowledge graph,lacking the utilization of rich exogenous information,resulting in the lack of semantics of the obtained knowledge;(2)the existing knowledge representation models that fuse exogenous information only utilize a single type of exogenous information,lacking the effective fusion and interaction of multi-source heterogeneous information such as text and images,and cannot more effectively solve the problem of data sparsity of knowledge graph cannot be solved more effectively.In order to solve the above problems,the following research works are carried out in this paper:1.To address the problem of insufficient utilization of exogenous information in the knowledge graph completion model,this paper proposes DKGC(Description-based Knowledge Graph Completion),a knowledge graph completion model based on the fusion of entity description information.This model improves the entity description encoder by using a pre-trained language model to obtain the feature representation of entity description information,and then fuses the entity description information with the triadic structure information of the knowledge graph based on the translation model.In this paper,experiments are conducted on the publicly available datasets FB15 K,WN18and YAGO3-10.The experimental results show that DKGC has better performance on several evaluation tasks compared to the benchmark model on different datasets,indicating that entity exogenous information can improve the representation of knowledge representation learning.2.To address the problem that the knowledge graph completion model only utilizes a single exogenous information,this paper designs a knowledge graph completion model DIKGC(Description and Image-based Knowledge Graph Completion)based on the fusion of multi-source heterogeneous information,which combines two kinds of exogenous information,namely entity image information and entity description text information,and fully interacts with them to replace the traditional knowledge representation learning.It replaces the traditional unimodal model that relies only on structural information or fuses only one type of exogenous information.This model mines entity description information using pre-trained language model,learns image information of entities in knowledge graph using convolutional neural network,and then fuses it with triadic structure information.In this paper,experiments are conducted on the publicly available datasets WN9-DI and WN11-DI,and the performance of DIKGC is further improved compared to DKGC on different evaluation tasks,which verifies the effectiveness of multi-source heterogeneous information fusion in enhancing the performance of the model.