摘要
作为一种新兴的数据结构,知识图谱被广泛用于搜索引擎、推荐系统、对话系统等诸多领域中.知识图谱补全(knowledge graph completion)是通过不同的方法,对图谱中的三元组残缺项进行补充.本文以模型构造方法为视角,从Trans结构、神经网络和张量分解三类方法对已有知识图谱补全的相关技术研究进行深入探讨,分析了不同补全技术的优缺点.指出了Trans结构模型和张量分解模型适用于大规模的知识图谱补全,而神经网络模型适用于关系结构复杂的知识图谱补全;现有知识图谱补全技术存在关系复杂性高、语义信息难以获取、训练代价大、模型扩展性差等不足.从知识图谱中复杂关系处理、上下文语义获取、节点间长期依赖关系捕获、模型融合与可扩展性等方面来展望了知识图谱补全技术未来主要研究方向.
As a new data structure, knowledge graph is widely used in search engines, recommendation systems, dialogue systems and many other fields.Knowledge graph completion is to supplement the triple incomplete items in the knowledge graph through different methods.From the perspective of model construction method, this paper deeply discusses the relevant technology research of existing knowledge graph completion from three methods: trans structure, neural network and tensor decomposition, compares and analyzes the experimental results of each model, and summarizes the advantages and disadvantages of each completion technology.Trans structure model and tensor decomposition model are suitable for large-scale knowledge graph completion, The neural network model is suitable for the complement of knowledge graph with complex relational structure.It is pointed out that knowledge graph completion technology has some shortcomings, such as high relationship complexity, difficult to obtain semantic information, high training cost and poor model scalability.The future research direction of knowledge graph completion technology is discussed from the aspects of relationship complexity, context semantic acquisition, long-term dependency between nodes, model fusion and scalability.
作者
吴国栋
刘涵伟
何章伟
李景霞
王雪妮
WU Guo-dong;LIU Han-wei;HE Zhang-wei;LI Jing-xia;WANG Xue-ni(School of Information&Computer,Anhui Agricultural University,Hefei 230036,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2023年第3期471-482,共12页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(31671589)资助
安徽省自然科学基金项目(2108085MF209)资助
安徽省科技重大专项项目(202103b06020013)资助
嵌入式系统与服务计算教育部重点实验室开放基金项目(ESSCKF2020-03)资助。
关键词
知识图谱补全
Trans结构
神经网络
张量分解
knowledge graph completion
Trans structure
neural network
tensor decomposition
