知识图谱表示学习技术综述

A Survey of Knowledge Representation Learning -Translation Models

摘要:

摘 要：知识图谱表示学习，是将知识图谱中实体与关系以低维稠密向量表示的技术，在知识图谱驱动的人工智能研究中发挥着基础性支撑作用,已是当下研究热点，引起学者广泛关注，并取得很多研究成果。从表示学习的基本概念出发,系统性地阐述知识图谱表示学习方法最新研究进展，具体从算法模型的问题背景、算法模型原理、算法模型特点等方面进行详细论述。聚焦平移模型类算法，将模型算法细分成：单数据空间、多数据空间、概率空间、外部信息融合等类型，详细分析代表性模型，并梳理各算法间演化关系，从定量和定性两个维度归纳总结平移类算法模型。从表示空间类型、编码模型、外部信息融合、实时知识表示学习等方面展望未来发展趋势。

Abstract:

Abstract：Knowledge representation learning is a fundamental technology that represents entities and relations in knowledge graph in low-dimensional semantic space, playing a significant role in the research of artificial intelligence driven by knowledges. It has become a research hotspot, attracting massive attention from scholars and reaping many research achievements. Starting from the basic concepts of the knowledge representation learning, this paper systematically elaborates on the latest research progress of knowledge representation learning methods, specifically discussing the motivations, principle, and characteristics of the algorithm models in detail. Focusing on translation-based models, the algorithms are subdivided into types: single data space, multiple data space, probability space, and external information fusion. The evolutions are discussed between typical algorithms, and extensively quantitative comparisons are conducted, according this subdivisions. Finally, this study discusses the remaining challenges and outlook the future directions for knowledge representation learning, from aspects such as representation space types, encoding models, external information fusion, real-time knowledge representation learning.