Research On Technologies Of Knowledge Graph Representation Learning Based On Semantic Analysis

Knowledge graph is a state-of-the-art approach to knowledge storage and management,distinguished from traditional databases due to its versatile heterogeneous graph structure,which is more human-friendly.As a result,knowledge graphs are applicable in diverse scenarios such as natural language processing,intelligent question answering,and recommendation systems.Knowledge graphs play a vital role in advancing the development of artificial intelligence and big data technology.Vectorizing entities and relations within a knowledge graph can improve semantic computation efficiency,thereby enhancing the potential of knowledge graphs to support various applications.However,existing representation learning models for knowledge graphs fail to fully utilize the semantic information of the core elements-entities and relations-resulting in limited capacity to learn deep latent semantics in the graph,ultimately hindering optimal performance.This thesis addresses these issues by proposing models that utilize techniques such as semantic propagation and data augmentation to enhance the semantic perspective of entities and relations.The primary contributions of this thesis include:1.To address the issue of underutilization of semantic information in entity text resulting from the simplistic combination of representations,a novel model that incorporates semantic propagation on a text-augmented knowledge graph is proposed.In some knowledge graphs,text descriptions containing rich knowledge related to entities can be used to supplement entity semantics alongside triple information.However,previous studies have primarily relied on simple combinations of triple and entity text representations,leading to insufficient integration of semantic information and inadequate performance improvements.To address this issue,this thesis proposes a novel method that utilizes semantic propagation on a text-augmented knowledge graph.First,named entities are extracted from the text and linked to corresponding entities to obtain an augmented knowledge graph.Graph convolutional networks are then employed to propagate semantic information between nodes in the augmented graph,effectively fusing both text and triple semantics.Finally,link prediction is conducted on four benchmark datasets-FB15 K,WN18,FB15K-237,and WN18 RR.The model surpasses the baseline Pretrain model,achieving the highest performance improvements of 5.1% and 2.0% on the MRR and HITS@10 metrics,respectively,demonstrating its effectiveness and advancement.2.To address the issue of low accuracy and difficulty in training few-shot relations caused by the long-tail distribution of knowledge graphs,a model based on variational autoencoder data augmentation is proposed.The relations in knowledge graphs often follow a long-tail distribution,which means that most relations have only a small number of related triples,known as "few-shot relations." These relations are difficult to fully capture semantics during the representation learning process,limiting the model’s final performance.To address this issue,this thesis proposes a few-shot representation learning model based on variational autoencoder data augmentation.Firstly,entity pair representations are obtained by aggregating entity context information from triples.Then,the variational autoencoder extracts latent features of the entity pair representations and generates new samples that are highly similar to the original samples to aid model training.Finally,two few-shot benchmark datasets,NELL and Wiki,were used for link prediction.The proposed model achieved improvements of 1.7% and 6.7% on the MRR and 5.0%and 8.5% on the HITS@1,respectively,compared with the baseline FAAN,verifying the effectiveness and advancement in few-shot scenarios.3.To address the issue of underutilizing entity context and relation rule semantics,which results in poor ontology and rule modeling,a novel model that incorporates meta information and logical rule enhancements is proposed.Entities in knowledge graphs typically exist in a contextual setting,and there is often rich contextual information around them.The context of the same entity type is usually similar,while there is a significant difference between different entity types,which can be used to differentiate them.Additionally,knowledge graphs often contain similar substructures with the same relations but different entities,which can be considered as rules used to infer unknown triples.However,existing models have not fully exploited these characteristics.Therefore,this thesis proposes a knowledge graph representation learning model based on meta information and logical rules enhancements.First,the entity meta information is modeled based on the surrounding structure of entities to learn ontology type knowledge.Then,high-confidence triples are inferred based on logical rules mined from the graph to expand the training set.Finally,the model combines the meta information,logical rules,and triple data for training to obtain representations of entities and relations.Experimental results on standard datasets FB15 K,WN18,FB15K-237,WN18 RR,and their sparse counterparts show that the proposed model outperforms existing baseline models in multiple evaluation metrics.For instance,in FB15K-237,the proposed model achieved improvements of 3.6% and 6.2% on the MRR and HITS@1,respectively,and performance improvements were particularly evident on sparse datasets,such as 15.4% and 28.1% improvements on the MRR and HITS@1 of the FB15 KSparse dataset.