面向链接预测的知识图谱表示学习方法综述

作为人工智能的重要基石, 知识图谱能够从互联网海量数据中抽取并表达先验知识, 极大程度解决了智能系统认知决策可解释性差的瓶颈问题, 对智能系统的构建与应用起关键作用. 随着知识图谱技术应用的不断深化, 旨在解决图谱欠完整性问题的知识图谱补全工作迫在眉睫. 链接预测是针对知识图谱中缺失的实体与关系进行预测的任务, 是知识图谱构建与补全中不可或缺的一环. 要充分挖掘知识图谱中的隐藏关系, 利用海量的实体与关系进行计算, 就需要将符号化表示的信息转换为数值形式, 即进行知识图谱表示学习. 基于此, 面向链接预测的知识图谱表示学习成为知识图谱领域的研究热点. 从链接预测与表示学习的基本概念出发, 系统性地介绍面向链接预测的知识图谱表示学习方法最新研究进展. 具体从知识表示形式、算法建模方式两种维度对研究进展进行详细论述. 以知识表示形式的发展历程为线索, 分别介绍二元关系、多元关系和超关系知识表示形式下链接预测任务的数学建模. 基于表示学习建模方式, 将现有方法细化为4类模型: 平移距离模型、张量分解模型、传统神经网络模型和图神经网络模型, 并详细描述每类模型的实现方式与解决不同关系元数链接预测任务的代表模型. 在介绍链接预测的常用的数据集与评判标准基础上, 分别对比分析二元关系、多元关系和超关系3类知识表示形式下, 4类知识表示学习模型的链接预测效果, 并从模型优化、知识表示形式和问题作用域3个方面展望未来发展趋势.

Survey on Representation Learning Methods of Knowledge Graph for Link Prediction

As an important cornerstone of artificial intelligence, knowledge graphs can extract and represent a priori knowledge from massive data on the Internet, which greatly solves the bottleneck problem of the poor interpretability of cognitive decisions of intelligent systems and plays a key role in the construction and application of intelligent systems. As the application of knowledge graph technology continues to deepen, the knowledge graph completion that aims to solve the problem of the incompleteness of graphs is imminent. Link prediction is the task of predicting the missing entities and relations in the knowledge graph, which is indispensable in the construction and completion of the knowledge graph. The full exploitation of the hidden relations in the knowledge graph and the use of massive entities and relations for computation require the conversion of the symbolic representations of information into the numerical form, i.e., knowledge graph representation learning. Hence, link prediction-oriented knowledge graph representation learning has become a popular research topic in the field of knowledge graphs. This study systematically introduces the latest research progress of link prediction-oriented knowledge graph representation learning methods from the basic concepts of link prediction and representation learning. Specifically, the research progress is discussed in detail in terms of knowledge representation forms and algorithmic modeling methods. The development of the knowledge representation forms is used as a clue to introduce the mathematical modeling of link prediction tasks in the knowledge representation forms of binary relations, multi-relations, and hyper-relations. On the basis of the representation learning modeling, the existing methods are refined into four types of models: translation distance models, tensor decomposition models, traditional deep learning models, and graph neural network models. The implementation methods of each type are described in detail together with representative models for solving link prediction tasks with different relational metrics. The common datasets and criteria for link prediction are then introduced, and on this basis, the link prediction effects of the four types of knowledge representation learning models under the knowledge representation forms of binary relations, multi-relations, and hyper-relations are presented in a comparative analysis. Finally, the future development trends are given in terms of model optimization, knowledge representation forms, and problem scope.