《计算机应用》唯一官方网站 ›› 2025, Vol. 45 ›› Issue (4): 1139-1147.DOI: 10.11772/j.issn.1001-9081.2024040536
收稿日期:
2024-04-30
修回日期:
2024-07-29
接受日期:
2024-08-01
发布日期:
2025-04-08
出版日期:
2025-04-10
通讯作者:
杨林旺
作者简介:
安俊秀(1970—),女,山西临汾人,教授,硕士,CCF会员,主要研究方向:数据挖掘、智能计算基金资助:
Junxiu AN, Linwang YANG(), Yuan LIU
Received:
2024-04-30
Revised:
2024-07-29
Accepted:
2024-08-01
Online:
2025-04-08
Published:
2025-04-10
Contact:
Linwang YANG
About author:
AN Junxiu, born in 1970, M. S., professor. Her research interests include data mining, intelligent computing.Supported by:
摘要:
针对离散词扰动和嵌入扰动方法中未充分考虑潜在空间词向量之间距离边界的问题,提出一种邻近性语义感知的对抗性自动编码器(SPAAE)方法。首先,采用对抗自动编码器作为底层模型;其次,根据词向量的邻近距离求得噪声向量概率分布的标准差;最后,通过对概率分布进行随机采样,动态调整扰动参数,从而最大限度模糊自身语义且不影响其他词向量的语义。实验结果表明,与DAAE (Denoising Adversarial Auto-Encoders)和EPAAE (Embedding Perturbed Adversarial Auto-Encoders)方法相比,所提方法在Yelp数据集上的自然流畅度分别提升了14.88%、15.65%;在Scitail数据集上的文本风格迁移(TST)的准确率分别提升了11.68%、6.45%;在Tenses数据集上的BLEU (BiLingual Evaluation Understudy)值分别提升了28.16%、26.17%。可见,SPAAE方法不仅在理论上提供了一种更精确的词向量扰动方式,而且在7个公开数据集上展示了它在不同风格迁移任务中的显著优势。特别是在网络舆论引导中,所提方法可以用于情感文本的风格迁移。
中图分类号:
安俊秀, 杨林旺, 柳源. 基于邻近性语义感知的无监督文本风格迁移[J]. 计算机应用, 2025, 45(4): 1139-1147.
Junxiu AN, Linwang YANG, Yuan LIU. Unsupervised text style transfer based on semantic perception of proximity[J]. Journal of Computer Applications, 2025, 45(4): 1139-1147.
模型 | 自然流畅度 | TST Acc | BLEU-2 | METEOR | ROUGE-L | CIDEr | 嵌入平均 | 向量极值 | 贪婪匹配 |
---|---|---|---|---|---|---|---|---|---|
β-VAE(β=0.15) | 0.829 | 0.856 | 0.060 | 0.075 | 0.206 | 0.299 | 0.768 | 0.445 | 0.596 |
LAAE(λ1=0.05) | 0.819 | 0.838 | 0.055 | 0.066 | 0.172 | 0.224 | 0.741 | 0.422 | 0.571 |
DAAE(p=0.3) | 0.746 | 0.842 | 0.272 | 0.202 | 0.482 | 1.605 | 0.838 | 0.603 | 0.750 |
EPAAE(ζ=2.0,p=0.3) | 0.741 | 0.841 | 0.263 | 0.180 | 0.424 | 1.627 | 0.844 | 0.597 | 0.725 |
SPAAE(p=0) | 0.857 | 0.845 | 0.155 | 0.133 | 0.337 | 0.737 | 0.803 | 0.496 | 0.659 |
SPAAE(p=0.3) | 0.752 | 0.834 | 0.196 | 0.154 | 0.374 | 0.975 | 0.819 | 0.554 | 0.692 |
SPAAE(p=0.1) | 0.795 | 0.838 | 0.295 | 0.214 | 0.499 | 1.647 | 0.858 | 0.615 | 0.750 |
表1 Yelp 数据集上TST任务的定量实验结果
Tab. 1 Quantitative experimental results of TST tasks on Yelp dataset
模型 | 自然流畅度 | TST Acc | BLEU-2 | METEOR | ROUGE-L | CIDEr | 嵌入平均 | 向量极值 | 贪婪匹配 |
---|---|---|---|---|---|---|---|---|---|
β-VAE(β=0.15) | 0.829 | 0.856 | 0.060 | 0.075 | 0.206 | 0.299 | 0.768 | 0.445 | 0.596 |
LAAE(λ1=0.05) | 0.819 | 0.838 | 0.055 | 0.066 | 0.172 | 0.224 | 0.741 | 0.422 | 0.571 |
DAAE(p=0.3) | 0.746 | 0.842 | 0.272 | 0.202 | 0.482 | 1.605 | 0.838 | 0.603 | 0.750 |
EPAAE(ζ=2.0,p=0.3) | 0.741 | 0.841 | 0.263 | 0.180 | 0.424 | 1.627 | 0.844 | 0.597 | 0.725 |
SPAAE(p=0) | 0.857 | 0.845 | 0.155 | 0.133 | 0.337 | 0.737 | 0.803 | 0.496 | 0.659 |
SPAAE(p=0.3) | 0.752 | 0.834 | 0.196 | 0.154 | 0.374 | 0.975 | 0.819 | 0.554 | 0.692 |
SPAAE(p=0.1) | 0.795 | 0.838 | 0.295 | 0.214 | 0.499 | 1.647 | 0.858 | 0.615 | 0.750 |
数据集 | 模型 | 自然流畅度 | TST Acc | BLEU-2 | METEOR | ROUGE-L | CIDEr | 嵌入平均 | 向量极值 | 贪婪匹配 |
---|---|---|---|---|---|---|---|---|---|---|
SNLI | β-VAE(β=0.15) | 0.979 | 0.538 | 0.441 | 0.265 | 0.559 | 2.629 | 0.945 | 0.612 | 0.813 |
LAAE(λ1=0.05) | 0.972 | 0.533 | 0.436 | 0.263 | 0.552 | 2.621 | 0.947 | 0.615 | 0.813 | |
DAAE(p=0.3) | 0.974 | 0.513 | 0.436 | 0.266 | 0.553 | 2.339 | 0.947 | 0.657 | 0.841 | |
EPAAE(ζ=2.5,p=0.3) | 0.979 | 0.509 | 0.490 | 0.300 | 0.604 | 2.924 | 0.959 | 0.696 | 0.865 | |
SPAAE(p=0) | 0.971 | 0.489 | 0.658 | 0.403 | 0.765 | 5.076 | 0.957 | 0.705 | 0.878 | |
SPAAE(p=0.3) | 0.980 | 0.503 | 0.516 | 0.315 | 0.629 | 3.201 | 0.963 | 0.718 | 0.875 | |
SPAAE(p=0.1) | 0.981 | 0.512 | 0.530 | 0.323 | 0.644 | 3.459 | 0.965 | 0.717 | 0.873 | |
DNLI | β-VAE(β=0.15) | 0.959 | 0.610 | 0.229 | 0.138 | 0.346 | 0.558 | 0.900 | 0.489 | 0.713 |
LAAE(λ1=0.05) | 0.960 | 0.629 | 0.232 | 0.140 | 0.346 | 0.577 | 0.902 | 0.495 | 0.715 | |
DAAE(p=0.3) | 0.962 | 0.618 | 0.545 | 0.318 | 0.640 | 3.990 | 0.954 | 0.709 | 0.858 | |
EPAAE(ζ=2.5,p=0.3) | 0.960 | 0.612 | 0.495 | 0.283 | 0.593 | 3.373 | 0.956 | 0.674 | 0.868 | |
SPAAE(p=0) | 0.967 | 0.608 | 0.682 | 0.409 | 0.766 | 5.750 | 0.972 | 0.801 | 0.909 | |
SPAAE(p=0.3) | 0.959 | 0.601 | 0.544 | 0.316 | 0.643 | 4.010 | 0.955 | 0.713 | 0.861 | |
SPAAE(p=0.1) | 0.966 | 0.610 | 0.653 | 0.391 | 0.738 | 5.461 | 0.968 | 0.781 | 0.900 | |
Scitail | β-VAE(β=0.15) | 0.795 | 0.595 | 0.155 | 0.099 | 0.242 | 0.545 | 0.841 | 0.428 | 0.672 |
LAAE(λ1=0.05) | 0.805 | 0.470 | 0.176 | 0.112 | 0.263 | 0.646 | 0.856 | 0.466 | 0.693 | |
DAAE(p=0.3) | 0.826 | 0.488 | 0.308 | 0.208 | 0.441 | 1.893 | 0.910 | 0.629 | 0.810 | |
EPAAE(ζ=2.5,p=0.3) | 0.830 | 0.512 | 0.290 | 0.177 | 0.449 | 1.968 | 0.906 | 0.603 | 0.776 | |
SPAAE(p=0) | 0.794 | 0.545 | 0.375 | 0.221 | 0.463 | 2.001 | 0.896 | 0.635 | 0.818 | |
SPAAE(p=0.3) | 0.797 | 0.503 | 0.336 | 0.203 | 0.423 | 1.731 | 0.914 | 0.618 | 0.807 | |
SPAAE(p=0.1) | 0.833 | 0.491 | 0.342 | 0.207 | 0.442 | 1.879 | 0.918 | 0.621 | 0.807 |
表2 NLI数据集上TST任务的定量实验结果
Tab. 2 Quantitative experimental results of TST tasks on NLI datasets
数据集 | 模型 | 自然流畅度 | TST Acc | BLEU-2 | METEOR | ROUGE-L | CIDEr | 嵌入平均 | 向量极值 | 贪婪匹配 |
---|---|---|---|---|---|---|---|---|---|---|
SNLI | β-VAE(β=0.15) | 0.979 | 0.538 | 0.441 | 0.265 | 0.559 | 2.629 | 0.945 | 0.612 | 0.813 |
LAAE(λ1=0.05) | 0.972 | 0.533 | 0.436 | 0.263 | 0.552 | 2.621 | 0.947 | 0.615 | 0.813 | |
DAAE(p=0.3) | 0.974 | 0.513 | 0.436 | 0.266 | 0.553 | 2.339 | 0.947 | 0.657 | 0.841 | |
EPAAE(ζ=2.5,p=0.3) | 0.979 | 0.509 | 0.490 | 0.300 | 0.604 | 2.924 | 0.959 | 0.696 | 0.865 | |
SPAAE(p=0) | 0.971 | 0.489 | 0.658 | 0.403 | 0.765 | 5.076 | 0.957 | 0.705 | 0.878 | |
SPAAE(p=0.3) | 0.980 | 0.503 | 0.516 | 0.315 | 0.629 | 3.201 | 0.963 | 0.718 | 0.875 | |
SPAAE(p=0.1) | 0.981 | 0.512 | 0.530 | 0.323 | 0.644 | 3.459 | 0.965 | 0.717 | 0.873 | |
DNLI | β-VAE(β=0.15) | 0.959 | 0.610 | 0.229 | 0.138 | 0.346 | 0.558 | 0.900 | 0.489 | 0.713 |
LAAE(λ1=0.05) | 0.960 | 0.629 | 0.232 | 0.140 | 0.346 | 0.577 | 0.902 | 0.495 | 0.715 | |
DAAE(p=0.3) | 0.962 | 0.618 | 0.545 | 0.318 | 0.640 | 3.990 | 0.954 | 0.709 | 0.858 | |
EPAAE(ζ=2.5,p=0.3) | 0.960 | 0.612 | 0.495 | 0.283 | 0.593 | 3.373 | 0.956 | 0.674 | 0.868 | |
SPAAE(p=0) | 0.967 | 0.608 | 0.682 | 0.409 | 0.766 | 5.750 | 0.972 | 0.801 | 0.909 | |
SPAAE(p=0.3) | 0.959 | 0.601 | 0.544 | 0.316 | 0.643 | 4.010 | 0.955 | 0.713 | 0.861 | |
SPAAE(p=0.1) | 0.966 | 0.610 | 0.653 | 0.391 | 0.738 | 5.461 | 0.968 | 0.781 | 0.900 | |
Scitail | β-VAE(β=0.15) | 0.795 | 0.595 | 0.155 | 0.099 | 0.242 | 0.545 | 0.841 | 0.428 | 0.672 |
LAAE(λ1=0.05) | 0.805 | 0.470 | 0.176 | 0.112 | 0.263 | 0.646 | 0.856 | 0.466 | 0.693 | |
DAAE(p=0.3) | 0.826 | 0.488 | 0.308 | 0.208 | 0.441 | 1.893 | 0.910 | 0.629 | 0.810 | |
EPAAE(ζ=2.5,p=0.3) | 0.830 | 0.512 | 0.290 | 0.177 | 0.449 | 1.968 | 0.906 | 0.603 | 0.776 | |
SPAAE(p=0) | 0.794 | 0.545 | 0.375 | 0.221 | 0.463 | 2.001 | 0.896 | 0.635 | 0.818 | |
SPAAE(p=0.3) | 0.797 | 0.503 | 0.336 | 0.203 | 0.423 | 1.731 | 0.914 | 0.618 | 0.807 | |
SPAAE(p=0.1) | 0.833 | 0.491 | 0.342 | 0.207 | 0.442 | 1.879 | 0.918 | 0.621 | 0.807 |
数据集 | 模型 | 自然流畅度 | TST Acc | BLEU-2 | METEOR | ROUGE-L | CIDEr | 嵌入平均 | 向量极值 | 贪婪匹配 |
---|---|---|---|---|---|---|---|---|---|---|
Voices | β-VAE(β=0.15) | 0.774 | 0.980 | 0.137 | 0.109 | 0.198 | 0.708 | 0.759 | 0.427 | 0.589 |
LAAE(λ1=0.05) | 0.783 | 0.982 | 0.091 | 0.078 | 0.159 | 0.418 | 0.736 | 0.382 | 0.554 | |
DAAE(p=0.3) | 0.789 | 0.981 | 0.255 | 0.176 | 0.307 | 1.593 | 0.810 | 0.532 | 0.673 | |
EPAAE(ζ=2.5,p=0.3) | 0.790 | 0.980 | 0.154 | 0.177 | 0.222 | 0.860 | 0.769 | 0.539 | 0.603 | |
SPAAE(p=0) | 0.795 | 0.987 | 0.251 | 0.186 | 0.337 | 1.418 | 0.823 | 0.536 | 0.689 | |
SPAAE(p=0.3) | 0.783 | 0.975 | 0.213 | 0.158 | 0.270 | 1.268 | 0.792 | 0.490 | 0.641 | |
SPAAE(p=0.1) | 0.781 | 0.981 | 0.275 | 0.200 | 0.347 | 1.680 | 0.830 | 0.554 | 0.700 | |
PPR | β-VAE(β=0.15) | 0.742 | 0.956 | 0.221 | 0.225 | 0.419 | 1.492 | 0.817 | 0.554 | 0.721 |
LAAE(λ1=0.05) | 0.746 | 0.945 | 0.189 | 0.184 | 0.364 | 1.212 | 0.794 | 0.510 | 0.682 | |
DAAE(p=0.3) | 0.730 | 0.948 | 0.276 | 0.292 | 0.503 | 2.007 | 0.848 | 0.628 | 0.784 | |
EPAAE(ζ=2.5,p=0.3) | 0.733 | 0.960 | 0.285 | 0.275 | 0.479 | 1.951 | 0.856 | 0.636 | 0.794 | |
SPAAE(p=0) | 0.756 | 0.942 | 0.303 | 0.300 | 0.539 | 2.180 | 0.867 | 0.652 | 0.807 | |
SPAAE(p=0.3) | 0.732 | 0.965 | 0.279 | 0.292 | 0.505 | 2.083 | 0.849 | 0.631 | 0.785 | |
SPAAE(p=0.1) | 0.746 | 0.948 | 0.320 | 0.325 | 0.558 | 2.416 | 0.874 | 0.673 | 0.822 | |
Tenses | β-VAE(β=0.15) | 0.792 | 0.998 | 0.153 | 0.116 | 0.241 | 0.788 | 0.774 | 0.436 | 0.607 |
LAAE(λ1=0.05) | 0.795 | 1.000 | 0.119 | 0.094 | 0.201 | 0.604 | 0.763 | 0.406 | 0.586 | |
DAAE(p=0.3) | 0.779 | 0.999 | 0.316 | 0.241 | 0.423 | 2.143 | 0.841 | 0.588 | 0.720 | |
EPAAE(ζ=2.5,p=0.3) | 0.777 | 0.999 | 0.321 | 0.245 | 0.431 | 2.220 | 0.835 | 0.575 | 0.722 | |
SPAAE(p=0) | 0.801 | 1.000 | 0.338 | 0.247 | 0.460 | 2.106 | 0.853 | 0.607 | 0.741 | |
SPAAE(p=0.3) | 0.770 | 0.998 | 0.398 | 0.303 | 0.507 | 2.805 | 0.868 | 0.653 | 0.770 | |
SPAAE(p=0.1) | 0.788 | 1.000 | 0.405 | 0.307 | 0.528 | 2.887 | 0.877 | 0.665 | 0.783 |
表3 细粒度风格数据集上TST任务的定量实验结果
Tab. 3 Quantitative experimental results of TST tasks on fine-grained style datasets
数据集 | 模型 | 自然流畅度 | TST Acc | BLEU-2 | METEOR | ROUGE-L | CIDEr | 嵌入平均 | 向量极值 | 贪婪匹配 |
---|---|---|---|---|---|---|---|---|---|---|
Voices | β-VAE(β=0.15) | 0.774 | 0.980 | 0.137 | 0.109 | 0.198 | 0.708 | 0.759 | 0.427 | 0.589 |
LAAE(λ1=0.05) | 0.783 | 0.982 | 0.091 | 0.078 | 0.159 | 0.418 | 0.736 | 0.382 | 0.554 | |
DAAE(p=0.3) | 0.789 | 0.981 | 0.255 | 0.176 | 0.307 | 1.593 | 0.810 | 0.532 | 0.673 | |
EPAAE(ζ=2.5,p=0.3) | 0.790 | 0.980 | 0.154 | 0.177 | 0.222 | 0.860 | 0.769 | 0.539 | 0.603 | |
SPAAE(p=0) | 0.795 | 0.987 | 0.251 | 0.186 | 0.337 | 1.418 | 0.823 | 0.536 | 0.689 | |
SPAAE(p=0.3) | 0.783 | 0.975 | 0.213 | 0.158 | 0.270 | 1.268 | 0.792 | 0.490 | 0.641 | |
SPAAE(p=0.1) | 0.781 | 0.981 | 0.275 | 0.200 | 0.347 | 1.680 | 0.830 | 0.554 | 0.700 | |
PPR | β-VAE(β=0.15) | 0.742 | 0.956 | 0.221 | 0.225 | 0.419 | 1.492 | 0.817 | 0.554 | 0.721 |
LAAE(λ1=0.05) | 0.746 | 0.945 | 0.189 | 0.184 | 0.364 | 1.212 | 0.794 | 0.510 | 0.682 | |
DAAE(p=0.3) | 0.730 | 0.948 | 0.276 | 0.292 | 0.503 | 2.007 | 0.848 | 0.628 | 0.784 | |
EPAAE(ζ=2.5,p=0.3) | 0.733 | 0.960 | 0.285 | 0.275 | 0.479 | 1.951 | 0.856 | 0.636 | 0.794 | |
SPAAE(p=0) | 0.756 | 0.942 | 0.303 | 0.300 | 0.539 | 2.180 | 0.867 | 0.652 | 0.807 | |
SPAAE(p=0.3) | 0.732 | 0.965 | 0.279 | 0.292 | 0.505 | 2.083 | 0.849 | 0.631 | 0.785 | |
SPAAE(p=0.1) | 0.746 | 0.948 | 0.320 | 0.325 | 0.558 | 2.416 | 0.874 | 0.673 | 0.822 | |
Tenses | β-VAE(β=0.15) | 0.792 | 0.998 | 0.153 | 0.116 | 0.241 | 0.788 | 0.774 | 0.436 | 0.607 |
LAAE(λ1=0.05) | 0.795 | 1.000 | 0.119 | 0.094 | 0.201 | 0.604 | 0.763 | 0.406 | 0.586 | |
DAAE(p=0.3) | 0.779 | 0.999 | 0.316 | 0.241 | 0.423 | 2.143 | 0.841 | 0.588 | 0.720 | |
EPAAE(ζ=2.5,p=0.3) | 0.777 | 0.999 | 0.321 | 0.245 | 0.431 | 2.220 | 0.835 | 0.575 | 0.722 | |
SPAAE(p=0) | 0.801 | 1.000 | 0.338 | 0.247 | 0.460 | 2.106 | 0.853 | 0.607 | 0.741 | |
SPAAE(p=0.3) | 0.770 | 0.998 | 0.398 | 0.303 | 0.507 | 2.805 | 0.868 | 0.653 | 0.770 | |
SPAAE(p=0.1) | 0.788 | 1.000 | 0.405 | 0.307 | 0.528 | 2.887 | 0.877 | 0.665 | 0.783 |
模型 | 将来时‒过去时 | 过去时‒将来时 | ||
---|---|---|---|---|
基线 | when trading will be halted by them all market liquidity will be gone | simply put there will not be enough business for every store to grow | it suggested that households accumulated wealth across a broad spectrum of assets | its other chemical operations bed continued by the Henderson plant the company said |
EPAAE (ζ=2.5,p=0.3) | but they were all the most prolific market value | i thought i thought it was more like it without much | it will suggest that households will accumulate wealth across a broad spectrum of assets | its |
SPAAE (p=0.1) | when markets were halted by them all that was volatility | simply put there was not enough business for every store to grow | it will suggest that wealth will be accumulated by households across a broad spectrum of assets | its other chemical operations will be continued by the Henderson plant the company will say |
表4 Tenses数据集上的风格迁移(时态迁移)输出样本
Tab. 4 Style transfer (temporal transfer) output samples on Tenses dataset
模型 | 将来时‒过去时 | 过去时‒将来时 | ||
---|---|---|---|---|
基线 | when trading will be halted by them all market liquidity will be gone | simply put there will not be enough business for every store to grow | it suggested that households accumulated wealth across a broad spectrum of assets | its other chemical operations bed continued by the Henderson plant the company said |
EPAAE (ζ=2.5,p=0.3) | but they were all the most prolific market value | i thought i thought it was more like it without much | it will suggest that households will accumulate wealth across a broad spectrum of assets | its |
SPAAE (p=0.1) | when markets were halted by them all that was volatility | simply put there was not enough business for every store to grow | it will suggest that wealth will be accumulated by households across a broad spectrum of assets | its other chemical operations will be continued by the Henderson plant the company will say |
k | SPAAE(p=0.1) | EPAAE(ζ=2.5,p=0.3) |
---|---|---|
1.0 | it | it has |
1.5 | it | it |
2.0 | it | it has |
2.5 | it | it has |
3.0 | it | it has |
表5 Yelp数据集上不同强度的情感风格输出样本
Tab. 5 Emotional style transfer output samples with different intensities on Yelp dataset
k | SPAAE(p=0.1) | EPAAE(ζ=2.5,p=0.3) |
---|---|---|
1.0 | it | it has |
1.5 | it | it |
2.0 | it | it has |
2.5 | it | it has |
3.0 | it | it has |
数据集 | K | 模型 | 自然流畅度 | TST Acc | BLEU-2 | METEOR | ROUGE-L | CIDEr | 嵌入平均 | 向量极值 | 贪婪匹配 |
---|---|---|---|---|---|---|---|---|---|---|---|
Yelp | 1 | SPAAE(p=0) | 0.842 | 0.835 | 0.147 | 0.126 | 0.325 | 0.702 | 0.799 | 0.493 | 0.651 |
SPAAE(p=0.3) | 0.744 | 0.829 | 0.259 | 0.195 | 0.458 | 1.454 | 0.808 | 0.607 | 0.736 | ||
SPAAE(p=0.1) | 0.783 | 0.822 | 0.290 | 0.211 | 0.488 | 1.596 | 0.801 | 0.609 | 0.743 | ||
2 | SPAAE(p=0) | 0.857 | 0.845 | 0.155 | 0.133 | 0.337 | 0.737 | 0.810 | 0.496 | 0.659 | |
SPAAE(p=0.3) | 0.752 | 0.834 | 0.196 | 0.154 | 0.374 | 0.975 | 0.819 | 0.554 | 0.692 | ||
SPAAE(p=0.1) | 0.795 | 0.838 | 0.295 | 0.214 | 0.499 | 1.647 | 0.858 | 0.615 | 0.750 | ||
DNLI | 1 | SPAAE(p=0) | 0.966 | 0.599 | 0.537 | 0.339 | 0.674 | 4.381 | 0.959 | 0.730 | 0.871 |
SPAAE(p=0.3) | 0.959 | 0.604 | 0.559 | 0.326 | 0.645 | 4.244 | 0.954 | 0.715 | 0.863 | ||
SPAAE(p=0.1) | 0.964 | 0.597 | 0.544 | 0.317 | 0.640 | 4.056 | 0.953 | 0.701 | 0.858 | ||
2 | SPAAE(p=0) | 0.967 | 0.608 | 0.682 | 0.409 | 0.766 | 5.750 | 0.972 | 0.801 | 0.909 | |
SPAAE(p=0.3) | 0.959 | 0.601 | 0.544 | 0.316 | 0.643 | 4.010 | 0.955 | 0.713 | 0.861 | ||
SPAAE(p=0.1) | 0.966 | 0.610 | 0.653 | 0.391 | 0.738 | 5.461 | 0.968 | 0.781 | 0.900 | ||
Voices | 1 | SPAAE(p=0) | 0.794 | 0.983 | 0.240 | 0.170 | 0.321 | 1.291 | 0.815 | 0.516 | 0.671 |
SPAAE(p=0.3) | 0.780 | 0.979 | 0.221 | 0.164 | 0.276 | 1.335 | 0.799 | 0.500 | 0.650 | ||
SPAAE(p=0.1) | 0.783 | 0.968 | 0.179 | 0.137 | 0.256 | 0.997 | 0.792 | 0.472 | 0.633 | ||
2 | SPAAE(p=0) | 0.795 | 0.987 | 0.251 | 0.186 | 0.337 | 1.418 | 0.823 | 0.536 | 0.689 | |
SPAAE(p=0.3) | 0.783 | 0.975 | 0.213 | 0.158 | 0.270 | 1.268 | 0.792 | 0.490 | 0.641 | ||
SPAAE(p=0.1) | 0.781 | 0.981 | 0.275 | 0.200 | 0.347 | 1.680 | 0.830 | 0.554 | 0.700 |
表6 不同K值下SPAAE在3种风格迁移任务中的定量实验结果
Tab. 6 Quantitative experimental results of SPAAE under different K values in three style transfer tasks
数据集 | K | 模型 | 自然流畅度 | TST Acc | BLEU-2 | METEOR | ROUGE-L | CIDEr | 嵌入平均 | 向量极值 | 贪婪匹配 |
---|---|---|---|---|---|---|---|---|---|---|---|
Yelp | 1 | SPAAE(p=0) | 0.842 | 0.835 | 0.147 | 0.126 | 0.325 | 0.702 | 0.799 | 0.493 | 0.651 |
SPAAE(p=0.3) | 0.744 | 0.829 | 0.259 | 0.195 | 0.458 | 1.454 | 0.808 | 0.607 | 0.736 | ||
SPAAE(p=0.1) | 0.783 | 0.822 | 0.290 | 0.211 | 0.488 | 1.596 | 0.801 | 0.609 | 0.743 | ||
2 | SPAAE(p=0) | 0.857 | 0.845 | 0.155 | 0.133 | 0.337 | 0.737 | 0.810 | 0.496 | 0.659 | |
SPAAE(p=0.3) | 0.752 | 0.834 | 0.196 | 0.154 | 0.374 | 0.975 | 0.819 | 0.554 | 0.692 | ||
SPAAE(p=0.1) | 0.795 | 0.838 | 0.295 | 0.214 | 0.499 | 1.647 | 0.858 | 0.615 | 0.750 | ||
DNLI | 1 | SPAAE(p=0) | 0.966 | 0.599 | 0.537 | 0.339 | 0.674 | 4.381 | 0.959 | 0.730 | 0.871 |
SPAAE(p=0.3) | 0.959 | 0.604 | 0.559 | 0.326 | 0.645 | 4.244 | 0.954 | 0.715 | 0.863 | ||
SPAAE(p=0.1) | 0.964 | 0.597 | 0.544 | 0.317 | 0.640 | 4.056 | 0.953 | 0.701 | 0.858 | ||
2 | SPAAE(p=0) | 0.967 | 0.608 | 0.682 | 0.409 | 0.766 | 5.750 | 0.972 | 0.801 | 0.909 | |
SPAAE(p=0.3) | 0.959 | 0.601 | 0.544 | 0.316 | 0.643 | 4.010 | 0.955 | 0.713 | 0.861 | ||
SPAAE(p=0.1) | 0.966 | 0.610 | 0.653 | 0.391 | 0.738 | 5.461 | 0.968 | 0.781 | 0.900 | ||
Voices | 1 | SPAAE(p=0) | 0.794 | 0.983 | 0.240 | 0.170 | 0.321 | 1.291 | 0.815 | 0.516 | 0.671 |
SPAAE(p=0.3) | 0.780 | 0.979 | 0.221 | 0.164 | 0.276 | 1.335 | 0.799 | 0.500 | 0.650 | ||
SPAAE(p=0.1) | 0.783 | 0.968 | 0.179 | 0.137 | 0.256 | 0.997 | 0.792 | 0.472 | 0.633 | ||
2 | SPAAE(p=0) | 0.795 | 0.987 | 0.251 | 0.186 | 0.337 | 1.418 | 0.823 | 0.536 | 0.689 | |
SPAAE(p=0.3) | 0.783 | 0.975 | 0.213 | 0.158 | 0.270 | 1.268 | 0.792 | 0.490 | 0.641 | ||
SPAAE(p=0.1) | 0.781 | 0.981 | 0.275 | 0.200 | 0.347 | 1.680 | 0.830 | 0.554 | 0.700 |
1 | JIN D, JIN Z, HU Z, et al. Deep learning for text style transfer: a survey[J]. Computational Linguistics, 2022, 48(1): 155-205. |
2 | SHI K, WANG Y, LU H, et al. EKGTF: a knowledge-enhanced model for optimizing social network-based meteorological briefings[J]. Information Processing and Management, 2021, 58(4): No.102564. |
3 | KASHYAP A R, HAZARIKA D, KAN M-Y, et al. So different yet so alike! constrained unsupervised text style transfer[C]// Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Stroudsburg: ACL, 2022: 416-431. |
4 | TOLSTIKHIN I, BOUSQUET O, GELLY S, et al. Wasserstein auto-encoders[EB/OL]. [2024-07-25].. |
5 | FU S, CHEN J, LEI L. Cooperative attention generative adversarial network for unsupervised domain adaptation[J]. Knowledge-Based Systems, 2023, 261: No.110196. |
6 | SHEN T, MUELLER J, BARZILAY R, et al. Educating text auto-encoders: latent representation guidance via denoising[C]// Proceedings of the 37th International Conference on Machine Learning. New York: JMLR.org, 2020: 8719-8729. |
7 | NARASIMHAN S, DEY S, DESARKAR M S. Towards robust and semantically organised latent representations for unsupervised text style transfer[C]// Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Stroudsburg: ACL, 2022: 456-474. |
8 | LI X L, THICKSTUN J, GULRAJANI I, et al. Diffusion-LM improves controllable text generation[C]// Proceedings of the 36th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2022: 4328-4343. |
9 | HU Z, LEE R K W, AGGARWAL C C, et al. Text style transfer: a review and experimental evaluation[J]. ACM SIGKDD Explorations Newsletter, 2022, 24(1): 14-45. |
10 | SHEN T, LEI T, BARZILAY R, et al. Style transfer from non-parallel text by cross-alignment[C]// Proceedings of the 31st International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2017: 6833-6844. |
11 | FU Z, TAN X, PENG N, et al. Style transfer in text: exploration and evaluation[C]// Proceedings of the 32nd AAAI Conference on Artificial Intelligence. Palo Alto: AAAI Press, 2018: 663-670. |
12 | PRABHUMOYE S, TSVETKOV Y, SALAKHUTDINOV R, et al. Style transfer through back-translation[C]// Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Stroudsburg: ACL, 2018: 866-876. |
13 | JOHN V, MOU L, BAHULEYAN H, et al. Disentangled representation learning for non-parallel text style transfer[C]// Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. Stroudsburg: ACL, 2019: 424-434. |
14 | LI J, JIA R, HE H, et al. Delete, retrieve, generate: a simple approach to sentiment and style transfer[C]// Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). Stroudsburg: ACL, 2018: 1865-1874. |
15 | CHEN M, TANG Q, WISEMAN S, et al. A multi-task approach for disentangling syntax and semantics in sentence representations[C]// Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). Stroudsburg: ACL, 2019: 2453-2464. |
16 | BAO Y, ZHOU H, HUANG S, et al. Generating sentences from disentangled syntactic and semantic spaces[C]// Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. Stroudsburg: ACL, 2019: 6008-6019. |
17 | XU P, CHEUNG J C K, CAO Y. On variational learning of controllable representations for text without supervision[C]// Proceedings of the 37th International Conference on Machine Learning. New York: JMLR.org, 2020: 10534-10543. |
18 | RUBENSTEIN P K, SCHÖLKOPF B, TOLSTIKHIN I. On the latent space of Wasserstein auto-encoders[EB/OL]. [2024-07-25].. |
19 | YANG Z, HU Z, DYER C, et al. Unsupervised text style transfer using language models as discriminators[C]// Proceedings of the 32nd International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2018: 7298-7309. |
20 | BAE K, KIM H I, KWON Y, et al. Unsupervised bidirectional style transfer network using local feature transform module[C]// Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. Piscataway: IEEE, 2023: 740-749. |
21 | HU Y, TAO W, XIE Y, et al. Token-level disentanglement for unsupervised text style transfer[J]. Neurocomputing, 2023, 560: No.126823. |
22 | 蔡国永,李安庆. 提示学习启发的无监督情感风格迁移研究[J]. 计算机工程与应用, 2024, 60(5): 146-155. |
CAI G Y, LI A Q. Prompt-learning inspired approach to unsupervised sentiment style transfer[J]. Computer Engineering and Applications, 2024, 60(5): 146-155. | |
23 | LEE D, TIAN Z, XUE L, et al. Enhancing content preservation in text style transfer using reverse attention and conditional layer normalization[C]// Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). Stroudsburg: ACL, 2021: 93-102. |
24 | MAI H, JIANG W, DENG Z H. Prefix-tuning based unsupervised text style transfer[C]// Findings of the Association for Computational Linguistics: EMNLP 2023. Stroudsburg: ACL, 2023: 14847-14856. |
25 | 李静文,叶琪,阮彤,等. 基于多奖励强化学习的半监督文本风格迁移方法[J]. 计算机科学, 2024, 51(8): 263-271. |
LI J W, YE Q, RUAN T, et al. Semi-supervised text style transfer method based on multi-reward reinforcement learning[J]. Computer Science, 2024, 51(8): 263-271. | |
26 | HO J, JAIN A, ABBEEL P. Denoising diffusion probabilistic models[C]// Proceedings of the 34th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2020: 6840-6851. |
27 | ROMBACH R, BLATTMANN A, LORENZ D, et al. High-resolution image synthesis with latent diffusion models[C]// Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2022: 10674-10685. |
28 | AUSTIN J, JOHNSON D D, HO J, et al. Structured denoising diffusion models in discrete state-spaces[C]// Proceedings of the 35th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2021: 17981-17993. |
29 | WANG Z, ZHAO L, XING W. StyleDiffusion: controllable disentangled style transfer via diffusion models[C]// Proceedings of the 2023 IEEE/CVF International Conference on Computer Vision. Piscataway: IEEE, 2023: 7643-7655. |
30 | YE J, ZHENG Z, BAO Y, et al. DiNoiser: diffused conditional sequence learning by manipulating noises[EB/OL]. [2024-07-25].. |
31 | GHOJOGH B, CROWLEY M, KARRAY F, et al. Adversarial auto-encoders[M]// Elements of dimensionality reduction and manifold learning. Cham: Springer, 2022: 577-596. |
32 | BOWMAN S R, ANGELI G, POTTS C, et al. A large annotated corpus for learning natural language inference[C]// Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing. Stroudsburg: ACL, 2015: 632-642. |
33 | WELLECK S, WESTON J, SZLAM A, et al. Dialogue natural language inference[C]// Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. Stroudsburg: ACL, 2019: 3731-3741. |
34 | KHOT T, SABHARWAL A, CLARK P. SciTail: a textual entailment dataset from science question answering[C]// Proceedings of the 32nd AAAI Conference on Artificial Intelligence. Palo Alto: AAAI Press, 2018:5189-5197. |
35 | LYU Y, LIANG P P, PHAM H, et al. StylePTB: a compositional benchmark for fine-grained controllable text style transfer[C]// Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Stroudsburg: ACL, 2021: 2116-2138. |
36 | BANERJEE S, LAVIE A. METEOR: an automatic metric for MT evaluation with improved correlation with human judgments[C]// Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization. Stroudsburg: ACL, 2005: 65-72. |
37 | LIN C Y. ROUGE: a package for automatic evaluation of summaries[C]// Proceedings of the ACL-04 Workshop: Text Summarization Branches Out. Stroudsburg: ACL, 2004: 74-81. |
38 | VEDANTAM R, ZITNICK C L, PARIKH D. CIDEr: consensus-based image description evaluation[C]// Proceedings of the 2015 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2015: 4566-4575. |
39 | SHARMA S, ASRI L EL, SCHULZ H, et al. Relevance of unsupervised metrics in task-oriented dialogue for evaluating natural language generation[EB/OL]. [2024-07-25].. |
40 | FORGUES G, PINEAU J, LARCHÊVEQUE J M, et al. Bootstrapping dialog systems with word embeddings[EB/OL]. [2024-07-25].. |
41 | RUS V, LINTEAN M. An optimal assessment of natural language student input using word-to-word similarity metrics[C]// Proceedings of the 2012 International Conference on Intelligent Tutoring Systems, LNCS 7315. Berlin: Springer, 2012: 675-676. |
42 | MIR R, FELBO B, OBRADOVICH N, et al. Evaluating style transfer for text[C]// Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). Stroudsburg: ACL, 2019: 495-504. |
43 | MIKOLOV T, YIH W T, ZWEIG G. Linguistic regularities in continuous space word representations[C]// Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Stroudsburg: ACL, 2013: 746-751. |
[1] | 丁美荣, 卓金鑫, 陆玉武, 刘庆龙, 郎济聪. 融合环境标签平滑与核范数差异的领域自适应[J]. 《计算机应用》唯一官方网站, 2025, 45(4): 1130-1138. |
[2] | 潘理虎, 彭守信, 张睿, 薛之洋, 毛旭珍. 面向运动前景区域的视频异常检测[J]. 《计算机应用》唯一官方网站, 2025, 45(4): 1300-1309. |
[3] | 贾洁茹, 杨建超, 张硕蕊, 闫涛, 陈斌. 基于自蒸馏视觉Transformer的无监督行人重识别[J]. 《计算机应用》唯一官方网站, 2024, 44(9): 2893-2902. |
[4] | 夏吾吉, 黄鹤鸣, 更藏措毛, 范玉涛. 基于无监督学习和监督学习的抽取式文本摘要综述[J]. 《计算机应用》唯一官方网站, 2024, 44(4): 1035-1048. |
[5] | 江锐, 刘威, 陈成, 卢涛. 非对称端到端的无监督图像去雨网络[J]. 《计算机应用》唯一官方网站, 2024, 44(3): 922-930. |
[6] | 曹铉, 罗天健. 运动想象脑电信号的跨被试动态多域对抗学习方法[J]. 《计算机应用》唯一官方网站, 2024, 44(2): 645-653. |
[7] | 赵培, 乔焰, 胡荣耀, 袁新宇, 李敏悦, 张本初. 基于多域特征提取的多变量时间序列异常检测[J]. 《计算机应用》唯一官方网站, 2024, 44(11): 3419-3426. |
[8] | 胡能兵, 蔡彪, 李旭, 曹旦华. 基于图池化对比学习的图分类方法[J]. 《计算机应用》唯一官方网站, 2024, 44(11): 3327-3334. |
[9] | 黄梦林, 段磊, 张袁昊, 王培妍, 李仁昊. 基于Prompt学习的无监督关系抽取模型[J]. 《计算机应用》唯一官方网站, 2023, 43(7): 2010-2016. |
[10] | 许喆, 王志宏, 单存宇, 孙亚茹, 杨莹. 基于重构误差的无监督人脸伪造视频检测[J]. 《计算机应用》唯一官方网站, 2023, 43(5): 1571-1577. |
[11] | 葛孟婷, 万鸣华. 基于近邻监督局部不变鲁棒主成分分析的特征提取模型[J]. 《计算机应用》唯一官方网站, 2023, 43(4): 1013-1020. |
[12] | 李文博, 刘波, 陶玲玲, 罗棻, 张航. L1正则化的深度谱聚类算法[J]. 《计算机应用》唯一官方网站, 2023, 43(12): 3662-3667. |
[13] | 刘拥民, 杨钰津, 罗皓懿, 黄浩, 谢铁强. 基于双向循环生成对抗网络的无线传感网入侵检测方法[J]. 《计算机应用》唯一官方网站, 2023, 43(1): 160-168. |
[14] | 郭一阳, 于炯, 杜旭升, 杨少智, 曹铭. 基于自编码器与集成学习的离群点检测算法[J]. 《计算机应用》唯一官方网站, 2022, 42(7): 2078-2087. |
[15] | 刘睿珩, 叶霞, 岳增营. 面向自然语言处理任务的预训练模型综述[J]. 《计算机应用》唯一官方网站, 2021, 41(5): 1236-1246. |
阅读次数 | ||||||
全文 |
|
|||||
摘要 |
|
|||||