基于知识图谱的零样本文档检索伪查询生成  

Knowledge Graph Aware Pseudo Query Generation for Zero-Shot Document Retrieval

作　　者：刘军平 孙医贵 朱强 胡新荣[1,2,3] 彭涛 姚迅[1,2,3] 王帮超 LIU Junping;SUN Yigui;ZHU Qiang;HU Xinrong;PENG Tao;YAO Xun;WANG Bangchao(School of Computer Science and Artificial Intelligence,Wuhan Textile University;Engineering Research Center of Hubei Province for Clothing Information;Hubei Provincial Engineering Research Center for Intelligent Textile and Fashion,Wuhan 430200,China)

机构地区：[1]武汉纺织大学计算机与人工智能学院 [2]湖北省服装信息化工程技术研究中心 [3]纺织服装智能化湖北省工程研究中心,湖北武汉430200

出　　处：《软件导刊》2024年第11期47-52,共6页Software Guide

基　　金：教育部人文社会科学研究一般项目(23YJAZH082);湖北省教育科学规划重点课题(2022GA046)。

摘　　要：为提高文档检索模型性能,减轻手工标记训练数据的工作量,提出一种基于知识图谱的零样本文档检索伪查询生成方法KGQG。该方法利用知识图谱增强伪查询,将外部信息与伪查询相结合,以生成更丰富、更具信息量的伪查询。实验结果表明,在BEIR基准测试的12个公开数据集中,KGQG方法比经典的稀疏检索模型、稠密检索模型以及最新的基于外部知识扩展的零样本稠密检索模型在归一化折现累计效益(NDCG)指标方面分别提升了4.6、11.88、7.96个百分点。KGQG方法不仅能提高检索性能,而且减少了手动标记训练数据需求,为文档检索模型的未来研究与应用提供了有益参考。To improve the performance of document retrieval models and reduce the workload of manually labeling training data,a zero sam⁃ple document retrieval pseudo query generation method KGQG based on knowledge graph is proposed.This method utilizes knowledge graphs to enhance pseudo queries,combining external information with pseudo queries to generate richer and more informative pseudo queries.The experimental results showed that in the 12 public datasets of the BEIR benchmark test,the KGQG method improved the normalized discounted cumulative benefit(NDCG)index by 4.6,11.88,and 7.96 percentage points respectively compared to the classical sparse retrieval model,dense retrieval model,and the latest zero sample dense retrieval model based on external knowledge extension.The KGQG method not only im⁃proves retrieval performance,but also reduces the need for manual labeling of training data,providing useful references for future research and application of document retrieval models.

关 键 词：稠密检索 信息检索 零样本学习 查询扩展 知识图谱 自然语言处理 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]