基于相关性的跨模态信息检索研究  

A Study on Correlation-based Cross-Modal Information Retrieval

作　　者：丁恒[1] 陆伟[1,2] 

机构地区：[1]武汉大学信息管理学院,武汉430072 [2]武汉大学信息资源研究中心,武汉430072

出　　处：《现代图书情报技术》2016年第1期17-23,共7页New Technology of Library and Information Service

基　　金：国家自然科学基金面上项目"基于语言模型的通用实体检索建模及框架实现研究"(项目编号:71173164)的研究成果之一

摘　　要：【目的】梳理基于相关性的跨模态信息检索中的基本策略和核心问题,从提升检索效果的角度探讨偏最小二乘法用于特征子空间投影的优劣。【方法】在Wikipedia跨模态信息检索数据集上,分别采用LDA和BOW模型作为文本和图像资源的特征表达方式,以余弦距离作为相似度度量方法,利用最小二乘法替代典型相关性分析法学习特征子空间投影函数。【结果】从P@K、MAP和NDCG三个检索评价指标上,对比分析典型相关性分析、偏最小二乘回归、偏最小二乘相关三种特征子空间投影法对跨模态信息检索结果的影响,结果表明偏最小二乘相关法具有最佳效果。【局限】偏最小二乘法在处理数据时假设数据之间的关系是线性的,数据基向量之间是正交关系,因而无法解决非线性、非正交问题。【结论】使用偏最小二乘相关法学习的特征子空间投影与原始空间信息的一致性更强,跨模态信息检索结果更稳定。[Objective] Summarize the fundamental strategies and core issues in Cross-Modal lntbrmatlon Retrieval （CMIR） based on correlation, and do research about the pros and cons of using partial least squares in feature subspace projection in order to improve retrieval effect. [Methods] Based on Wikipedia CMIR dataset, LDA and BOW models are used as a characteristic expression of text and image resources, cosine distance as the similarity measure, and the least squares method is used to learn subspace projection function replacing canonical correlation analysis method. [Results] Using comparative analysis of the influence of three features subspace projection methods named canonical correlation analysis, partial least squares regression, partial least squares correlation on CMIR results according to three retrieval evaluation indicators that are P@K, MAP and NDCG, and the results show that partial least squares correlation obtains the best results. [Limitations] In dealing with data, partial least squares method assumes a linear relationship between the data and an orthogonal relationship between the data base vectors, therefore the non-linear, non-orthogonal problem can not be solved. [Conclusions] Feature subspace projection learning by using partial least squares correlation is more consistent with original spatial information, and CMIR results are more stable.

关 键 词：跨模态信息检索 偏最小二乘法 子空间投影 

分 类 号：TP391.3[自动化与计算机技术—计算机应用技术]