基于图约简的知识联想关系网络建模

Title:

Modeling knowledge network on associative relations based on graph reduction

作者:

王坤, 谢振平, 陈梅婕

1. 江南大学 数字媒体学院, 江苏 无锡 214122;
2. 江苏省媒体设计与软件技术重点实验室, 江苏 无锡 214122

Author(s):

WANG Kun, XIE Zhenping, CHEN Meijie

1. School of Digital Media, Jiangnan University, Wuxi 214122, China;
2. Jiangsu Key Laboratory of Media Design and Software Technology, Wuxi 214122, China

关键词:

知识图谱; 联想记忆; 知识建模; 图约简; 知识网络; 知识联想; 记忆保留; 关系网络

Keywords:

knowledge graph; associative memory; knowledge modeling; graph reduction; knowledge network; knowledge association; memory reservation; relational network

分类号:

TP391

DOI:

10.11992/tis.201808009

摘要:

考虑到人类知识在大脑中以联想记忆形式存在，尝试从联想关系的视角深入探索知识体系的内在关系网络模型，旨在为知识图谱的建模提供一种可参考的新思路。对于给定的知识语料库，首先考虑以直接联想关系生成方式构建初始的知识关系网络，随后引入多种图约简方法优化知识联想关系网络的建模。研究中特别地提出了随机选择、局部联想最大记忆保留、全局联想最大记忆保留等3种知识联想关系约简重整策略，并通过实验手段对这3种策略进行建模分析。实验结果表明：3种方法呈现出了价值意义清晰的共同性能特征，而其中的全局联想最大记忆保留策略能最优地平衡知识联想关系网络的规模和联想记忆效率，可为相关应用提供有效的方法基础，也可为进一步探索类脑联想记忆的知识关系网络生成建模提供十分有益的启发。

Abstract:

Inspired by the fact that knowledge is stored in the form of associative memory in the human brain, we discuss the internal associative network model of knowledge system using associative relations, in order to provide a new train of thoughts referential to modeling knowledge graph. For a given knowledge corpus, an initial knowledge relation network was first constructed by producing direct associative relations, and then several graph reduction methods were introduced to optimize the modeling efficiency. Random selection and local and global strong memory reservation strategies were designed to reform the associative relations and their associative intensities, and experimental datasets were used to analyze these three modeling strategies. The experimental results show that the three different strategies exhibit interesting common characteristics. Moreover, global strong memory reservation strategy can optimize balance between the size of knowledge associative relation network and the associative memory efficiency. The results can provide a basis for related applications, as well as provide a meaningful understanding for exploring human-like knowledge associative memory modeling problems.