文章摘要
涂曼姝,潘接林.关于深度神经网络在交叉领域的情感分类任务中的可迁移性探究[J].情报工程,2018,4(6):013-024
关于深度神经网络在交叉领域的情感分类任务中的可迁移性探究
How Features Transferred in Very Deep Neural Networks on Cross Domain Sentiment Classification
  
DOI:10.3772/j.issn.2095-915X.2018.06.002
中文关键词: 领域自使用  TVDCNN  交叉领域的情感分类
英文关键词: Domain adaptation  TVDCNN  cross domain sentiment classification
基金项目:国家自然科学基金(61650202,11590770-4,U1536117)
作者单位
涂曼姝 1. 中国科学院声学研究所 语言声学与内容理解重点实验室 
潘接林 2. 中国科学院大学 
摘要点击次数: 147
全文下载次数: 121
中文摘要:
      领域自适应中交叉领域的情感分析研究近几年获得了广泛的关注。传统的做法是寻找源域和目标域相同的特征表示,或将源域和目标域的特征映射在高维空间中,使目标域的特征表达在高维空间中更加靠近源域以达到领域自适应的目的。然而由于这些方法都是直接将源域的全部网络权重迁移至目标域,没有考虑到神经网络中文本特征可能具有的层级性,因此我们提出一种基于超深度卷积神经网络(VDCNN) 的层级迁移方法(TVDCNN) 以探究如下几个问题:(1)交叉领域情感分类任务中可迁移的网络层有几层(2)哪几层获得了最好的迁移特征(3)不同的领域可迁移层数是否一致。在中英文的两个数据集的实验结果表明,文本的特征提取确实具有层级特征,前三层的迁移性最好,不同领域的可迁移层数基本一致,且在迁移之后对网络进行微调可以进一步提高正确率。
英文摘要:
      Domain adaptation has raised a lot attention in recent years especially in the field of crossdomain sentiment classification. The traditional approach is to find the same feature representation of the source domain and the target domain, or map the source and target domain in a high dimensional space, so that the target domain features are closed to the source domain to achieve domain adaptation.However, because these methods directly transfer the network weights learned from source domain to the target domain, they do not take into account the hierarchical nature of text feature extraction in the network. Based on these reasons, we proposed a hierarchical transfer function based on the VDCNN to explore these questions: (1)how many layers can be transferred in cross domain sentiment classification.(2) which layer get best transferring features. (3)weather different domain get same transferability. The experiments on two different datasets demonstrated that documents features have hierarchical character.The first three layers have the best transferability. Different domains have same layers for transferring and the accuracy be further improved when fine-turn after the transer.
查看全文   查看/发表评论  下载PDF阅读器
关闭

分享按钮