«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

期数:

2025年02期

页码:

138-145

栏目:

人工智能

出版日期:

2025-02-10

文章信息/Info

Title:

An Efficient Search Method on Temporal Knowledge Graph Supported by Vector Indexing

文章编号:

1673-629X(2025)02-0138-08

作者:

朱红1; 胡新雨2; 高莉莎2; 张强3; 徐晓轶2; 朱盟盟4

1. 国网南京供电公司,江苏 南京 210009;

2. 国网南通供电公司,江苏 南通 226000;

3. 国网智能电网研究院有限公司,北京 102200;

4. 苏州华天国科电力科技有限公司,江苏 苏州 215000

Author(s):

ZHU Hong1; HU Xin-yu2; GAO Li-sha2; ZHANG Qiang3; XU Xiao-yi2; ZHU Meng-meng4

1. State Grid Nanjing Power Supply Company,Nanjing 210009,China;

2. State Grid Nantong Electric Power Co. ,Ltd. ,Nantong 226000,China;

3. State Grid Smart Grid Research Institute Co. ,Ltd. ,Beijing 102200,China;

4. Suzhou Huatian Guoke Electric Power Technology Co. ,Ltd. ,Suzhou 215000,China

关键词:

知识图谱嵌入; 时态知识图谱; 索引; 搜索; 向量数据库; 机器学习

Keywords:

knowledge graph embedding; temporal knowledge graph; indexing; search; vector database; machine learning

分类号:

TP305

DOI:

10.20165/j.cnki.ISSN1673-629X.2024.0305

摘要:

知识图谱嵌入(Knowledge Graph Embedding,KGE)将实体和关系表示为低维、连续的向量,使机器学习模型能够轻松适应知识图谱(Knowledge Graph,KG)的搜索任务。 然而,在大规模知识库(Knowledge Base,KB)的搜索密集型应用中,现有的模型大多侧重于提高在静态 KG 上搜索的准确性,忽略了在动态时态知识图谱(Temporal Knowledge Graph,TKG)上搜索的时间效率。 为此,提出了一种向量索引支持的 TKG 高效搜索方法,以提高在 TKG 上的搜索效率。 具体来说,首先,将实体,关系和时间信息映射到向量空间,并利用长短期记忆神经网络(Long Short-Term Memory,LSTM)学习关系类型的时间感知,从而建立了具有时间信息感知与关系联合编码的 TKG 向量库。 然后,利用向量数据库建立大规模TKG 的向量索引库(IndexIVFFlat)。 注意,该索引通过聚类操作来划分搜索空间,以提高知识的搜索效率。 最后,在拥有高效索引的 TKG 上通过相似度计算执行近似性搜索与实验评估。 结果显示,该方法在时间效率上优于未建立索引的方法,且在搜索准确度上优于一些强相关的方法。 表明,该向量索引库的建立在保证了搜索准确性的前提下提高了在 TKG上的搜索效率。

Abstract:

Knowledge graph embedding (KGE) represents entities and relations as low-dimensional,continuous vectors,thus enabling machine learning models to be easily adapted to knowledge graph (KG) search task. However,in search-intensive applications of large-scale knowledge base (KB),most of the existing models focus on improving the accuracy of searching on static KGs,while neglecting the time efficiency of searching on dynamic temporal knowledge graph ( TKG). To this end, an efficient search method for TKG supported by vector indexing is proposed to improve the search efficiency on TKG. Specifically,firstly,the entity,relation and time infor-mation are mapped to the vector space,and the time-awareness of relation types is learned using the Long Short-Term Memory (LSTM) neural network,which leads to the establishment of the TKG vector database with the joint encoding of time information awareness and relation. Then,the vector database is utilized to build a vector indexing database ( IndexIVFFlat) for large-scale TKGs. Note that the index divides the search space by clustering operations to improve the search efficiency of knowledge. Finally,approximation search and experimental evaluation are performed by similarity computation on the TKG with efficient indexing. Results show that the proposed method outperforms the unindexed method in time efficiency and outperforms some strongly correlated methods in search accuracy. It is demonstrated that the establishment of this vector indexing database improves the search efficiency on the TKG under the guarantee of the search accuracy.

常用功能

更新日期/Last Update: 2025-02-10