基于 unwordMixup 的半监督深度学习模型
DOI:
作者:
作者单位:

1.山东工商学院;2.大连海事大学

作者简介:

通讯作者:

中图分类号:

TP181

基金项目:

国家自然科学基金项目(面上项目)No.61976124, No.61976125,No.61773244,No.61772319


Semi--supervised Deep learning model based on unwordMixup
Author:
Affiliation:

1.Shandong Technology and Business University;2.Dalian Maritime University

Fund Project:

The National Natural Science Foundation of China (General Program)No.61976124, No.61976125,No.61773244,No.61772319

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    当标注样本匮乏时,半监督学习利用大量未标注样本来解决标注瓶颈的问题,但由于未标注样本和标注样本来自不同领域,可能造成未标注样本存在质量问题,使得模型的泛化能力变差,导致分类精度下降。为此,本文基于 wordMixup 方法,提出了针对未标注样本进行数据增强的 u-wordMixup 方法,并结合一致性训练框架和 Mean Teacher 模型,提出了一种基于 u-wordMixup 的半监督深度学习模型 (Semi-supervised Deep learning model based on u-wordMixup,SD-uwM)。该模型利用 u-wordMixup 方法对未标注样本进行数据增强,在有监督交叉熵和无监督一致性损失的约束下,能够提高未标注样本质量,减少过度拟合。在 AGNews、THUCNews 和 20Newsgroups 数据集上的对比实验结果表明,所提方法能够提高模型的泛化能力,同时也有效提高了时间性能。

    Abstract:

    When labeled data are deficient, semi-supervised learning uses a large number of unlabeled data to solve the bottleneck problem of labeled data. However, the unlabeled data and labeled data come from different fields, it may cause quality problem of unlabeled data, which makes the generalization ability of the model poor and leads to the degradation of classification accuracy. Therefore, based on the wordMixup method, this paper proposes the u-wordMixup method for data augmentation of unlabeled data, and a semi-supervised deep learning model based on u-wordMixup (SD-uwM) by combining the consistent training framework and Mean Teacher model. The model utilizes the u-wordMixup method to augment data of unlabeled data, which can improve the quality of unlabeled data and reduce overfitting under the constraints of supervised cross-entropy and unsupervised consistency loss. The comparative experimental results on the datasets of AGNews, THUCNews and 20 Newsgroups show that the proposed method can improve the generalization ability of the model and also effectively improve the time performance.

    参考文献
    相似文献
    引证文献
引用本文
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2021-10-18
  • 最后修改日期:2022-03-07
  • 录用日期:2022-03-15
  • 在线发布日期:
  • 出版日期: