适用于网络新闻数据的未配对跨模态哈希方法  

Unpaired Cross-modal Hashing Method for Web News Data

作　　者：武昭盟 张成刚[2] WU Zhao-meng;ZHANG Cheng-gang(School of Cyberspace Security,Changchun University,Changchun 130012,China;College of Computer Science and Technology,Inner Mongolia Minzu University,Tongliao 028000,China)

机构地区：[1]长春大学网络空间安全学院,吉林长春130012 [2]内蒙古民族大学计算机科学与技术学院,内蒙古通辽028000

出　　处：《计算机与现代化》2024年第3期54-60,共7页Computer and Modernization

摘　　要：针对当前大部分跨模态哈希方法只能在提供完全配对的实例时才能训练,而不适用于现实世界中存在的大量未配对数据这一情况,提出一种基于网络新闻数据的未配对跨模态哈希方法。首先,构建特征融合网络处理未配对的训练数据,补充和完善模态信息,并采用对抗性损失加强学习的公共表示。其次,使用亲和矩阵对样本的特征分布和生成的二进制码进行优化,使样本之间的语义关系更加明确。最后,添加类别预测损失以提高二进制码的判别能力。在真实的网络新闻数据集上分别进行了配对场景和非配对场景的实验,实验结果表明了本文提出的方法具有扩展到实际应用中的能力。Most of the current cross-modal Hashing methods can only be trained when fully paired instances are provided,and are not suitable for a large number of unpaired data in the real world.In order to solve this problem,an unpaired cross-modal Hashing method for Web news data is proposted.Firstly,a feature fusion network is constructed to process the unpaired training data,the modal information is supplemented and improved,and the adversarial loss is used to strengthen the common representa⁃tion of learning.Secondly,the affinity matrix optimizes the feature distribution of samples and the generated binary codes,so that the semantic relationship between samples is more explicit.Finally,we add a class prediction loss to enhance the discrimina⁃tion ability of binary codes.Experiments on real network news datasets with paired scenes and unpaired scenes respectively,the results show that the proposed method can be extended to practical applications.

关 键 词：跨模态哈希 特征融合 未配对数据 对抗性学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]