摘要
针对文本生成模型训练所需的特定领域数据集匮乏的问题,提出一种基于TDTGAN的文本数据增强模型,用以扩充特定领域的文本数据。该模型生成器部分采用RMC代替常用的LSTM架构,来使模型获得更好的表达能力;通过Perturb-Softmax技术来优化模型对离散数据的处理能力;构建动态余弦相似度损失函数来提高模型增强文本的多样性;并通过引入分类器来使模型能够增强特定领域的文本。在Oracle、MR和AR数据集上进行了多组对比实验,结果显示,该模型相比现有文本数据增强模型,能更有效地提高所增强本文的质量和多样性。
A text data augmentation model based on TDTGAN is proposed to address the problem of insufficient domain specific datasets required for training text generation models,in order to expand domain specific text data.The model generator employs RMC instead of the traditional LSTM architecture,to enhance the model’s expressive power.The model’s ability is optimized to handle discrete data through the Perturb-Softmax technique.We construct a dynamic cosine similarity loss function to improve the diversity of text enhanced by the model.And by introducing classifiers,the model can enhance text in specific domains.Comparative experiments were conducted on Oracle,MR,and AR datasets,and the results showed that compared to existing text data augmentation models,our model can more effectively improve the quality and diversity of the enhanced text.
作者
闫雒恒
薛丰衣
YAN Luoheng;XUE Fengyi(College of Information Engineering,North China University of Water Resources and Electric Power,Zhengzhou 450000,China)
出处
《河南科学》
2025年第6期787-796,共10页
Henan Science
基金
河南省科技攻关项目(242102210213)
教育教学资源建设项目(YHJC22ZD059)。
关键词
生成对抗网络
文本数据增强
离散数据
分类器
余弦相似度损失函数
generative adversarial networks
text data augmentation
discrete data
classifier
cosine similarity loss function
