基于语义信息的大规模知识图谱补全算法  

Large-scale Knowledge Graph Complementation Algorithm Based on Semantic Information

作　　者：李鑫[1,2,3] 何芳州[1,2,3] LI Xin;HE Fang-zhou(Shenyang Institute of Computing Technology,Chinese Academy of Sciences,Shenyang Liaoning 110168,China;Criminal Investigation Police University of China,Shenyang Liaoning 110854,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院沈阳计算机技术研究所,辽宁沈阳110168 [2]中国刑事警察学院,辽宁沈阳110854 [3]中国科学院大学,北京100049

出　　处：《计算机仿真》2023年第12期428-433,共6页Computer Simulation

基　　金：2020年度辽宁省社科基金重点项目(L20ATQ002);辽宁省教育厅2021年科研经费项目(LJKR0026);公安部公安理论及软科学研究计划面上项目研究(2021LL49);中国刑事警察学院院级项目(D2022046)。

摘　　要：针对现有知识图谱补全算法中存在三元组复杂关系表示能力弱,缺失实体与实体关系三元组预测精度低的问题,提出一种基于改进Trans H算法与DSICNN算法相结合的知识图谱补全算法,提高了缺失三元组的预测精度。上述算法首先通过构建语义信息超平面Si提高Trans H算法性能;然后利用改进算法提取三元组偏导语义信息向量DSI,提高三元组复杂关系表示能力;接着将DSI链接后作为卷积神经网络的输入,通过卷积、池化与投影处理,构建DSICNN模型;最后利用损失函数迭代并用打分评价函数对构建的模型进行评价。链接预测实验与三元组分类实验表明,提出的DSICNN算法针对实体关系预测在MR、MRR以及Hits@10指标上均有着最高性能,且在FB15K-237和NELL-995大数量文本数据集中有着较好的表现,表明提出的算法在提高三元组预测精度降低了三元组复杂关系的表示能力,且可以用于大规模知识图谱补全。To address the problems of weak representation of complex relationships of triads and low prediction accuracy of missing entity-entity relationship triads in existing knowledge graph complementation algorithms.In this paper,we propose a knowledge graph complementation algorithm based on the combination of an improved Trans H algorithm and DSICNN algorithm to improve the prediction accuracy of missing triples.The algorithm first improves the performance of the Trans H algorithm by constructing the semantic information hyperplane Si;then the improved algorithm is used to extract the triad biased semantic information vector DSI to improve the ability to represent com⁃plex relations of triads;then the DSI is linked and used as the input of the convolutional neural network,and the DSICNN model is constructed by convolution,pooling and projection processing;finally,the constructed model is e⁃valuated by using loss function iteration and scoring function.The link prediction experiments and triad classification experiments show that the DSICNN algorithm proposed in this paper has the highest performance in MR,MRR and Hits@10 metrics for entity relationship prediction,and has better performance in FB15K-237 and NELL-995 large number of text datasets,which indicates that the algorithm proposed in this paper can be used to improve the triad prediction accuracy to reduce the triad complex relationship representation capability and can be used for large-scale knowledge graph complementation.

关 键 词：知识图谱 语义信息 补全算法 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置]