基于卷积神经网络的文档列表排序学习算法  

LEARNING ALGORITHM OF DOCUMENT LIST SORTING BASED ON CONVOLUTIONAL NEURAL NETWORK

作　　者：曹军梅[1] 马乐荣[1] Cao Junmei;Ma Lerong(College of Mathematics and Computer Science,Yan an University,Yan an 716000,Shaanxi,China)

机构地区：[1]延安大学数学与计算机科学学院,陕西延安716000

出　　处：《计算机应用与软件》2022年第7期256-260,303,共6页Computer Applications and Software

基　　金：国家自然科学基金项目(61751217,61866038);延安大学科研引导项目(YDY2018-11);大学生创新创业训练计划项目。

摘　　要：在许多信息检索任务中,为了进一步提高检索性能,通常需要对检索到的文档进行重新排序。现有的排序学习算法主要集中在损失函数的构造上,而没有考虑特征之间的关系。将多通道深度卷积神经网络作用于文档列表排序学习算法,即ListCNN,实现了信息检索的精确重排序。对于从文档中提取的多模态特征,其中一些特征具有局部相关性和冗余性,因此利用卷积神经网络来重新提取特征以提高列表方法的性能。ListCNN架构考虑了原始文档特征的局部相关性,能够有效地重新提取代表性特征。在公共数据集上对ListCNN进行了验证,实验结果表明其性能优于已有文档列表排序学习算法。In many information retrieval tasks, it is necessary to reorder the retrieved documents to further improve retrieval performance. The existing sorting learning approaches mainly focus on the construction of loss functions without considering the relationship among features. This paper applied the multi-channel deep convolutional neural networks(CNNs) to learning algorithm of document list sorting, namely ListCNN, which could achieve accurate re-sorting for information retrieval. For the multi-modal features extracted from documents, we found that some features were in local correlation with redundancy. Accordingly, we adopted convolutional deep neural networks to re-extract features to boost the performance of list approaches. The ListCNN architecture could effectively re-extract representational features by considering the local correlation from original document features. The proposed ListCNN architecture was validated on public datasets. The result shows that its performance is superior to the existing document list sorting learning algorithm.

关 键 词：文档列表 排序学习 多模态特征 卷积神经网络 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]