基于检索器-鉴别器架构的电力地址匹配模型研究  

Power Address Matching Model Based on Retriever Discriminator Architecture

作　　者：赵坚鹏 盛方 徐川子 陈奕 罗庆 陈聪 ZHAO Jianpeng;SHENG Fang;XU Chuanzi;CHEN Yi;LUO Qing;CHEN Cong(State Grid Hangzhou Electric Power Company,Hangzhou 310000,Zhejiang,China)

机构地区：[1]国网杭州供电公司,浙江杭州310000

出　　处：《电力大数据》2023年第1期35-43,共9页Power Systems and Big Data

摘　　要：为解决电力地址库与外部地址库地址的匹配问题,保证电力地址准确性,实现电力系统与外部系统间数据信息共享互通,本文提出一种基于检索器-鉴别器架构的地址匹配模型。首先介绍地址匹配模型的详细结构,包括用于缩小地址检索范围的地址检索器和最终分辨地址是否匹配正确的地址鉴别器,其中地址检索器基于词频-逆文档频率算法构建,地址鉴别器基于中文预训练语言模型NEZHA构建。还提出了一种负样本训练方法提升地址鉴别器辨别效果。详细介绍了实验分析所使用的两个数据集。实验结果表明基于检索器-鉴别器架构的电力地址匹配模型能够准确从外部地址库中找出与电力地址匹配的地址,其中,地址鉴别器能够非常准确地从多个候选地址中找出准确匹配地址,其F1分数达0.99以上。In order to solve the address matching problem between power address database and external address database,ensure the accuracy of power address,and realize the data information sharing and interworking between the power system and the external system,an address matching model based on retriever-discriminator architecture is proposed.Firstly,the detailed structure of the address matching model is introduced,including address retriever to narrow the search scope and the address discriminator to distinguish whether the address matches eventually.The address retriever is constructed based on the word frequency-inverse document frequency algorithm,and on the Chinese pre-training language model NEZHA.A negative sample training method is proposed to improve the discrimination effect of the model.Two datasets used in the experiment are introduced in detail.The experimental results show that the power address matching model based on the searcher-discriminator architecture can accurately find the address that matches the power address from the external address library,among which the address discriminator can accurately find the exact matching address from multiple candidate addresses,and its F1 score is above 0.99.

关 键 词：地址匹配 电力地址 词频-逆文档频率 中文预训练语言模型 负样本 

分 类 号：TM933[电气工程—电力电子与电力传动]