薛琪, 孟祥福, 张峰, 张霄雁, 朱金侠, 朱尧, 王丹丹. HLMGAN:分层学习的多奖励文本生成对抗网络[J]. 云南大学学报(自然科学版), 2022, 44(1): 64-72. doi: 10.7540/j.ynu.P00173
引用本文: 薛琪, 孟祥福, 张峰, 张霄雁, 朱金侠, 朱尧, 王丹丹. HLMGAN:分层学习的多奖励文本生成对抗网络[J]. 云南大学学报(自然科学版), 2022, 44(1): 64-72. doi: 10.7540/j.ynu.P00173
XUE Qi, MENG Xiang-fu, ZHANG Feng, ZHANG Xiao-yan, ZHU Jin-xia, ZHU Yao, WANG Dan-dan. HLMGAN: Generation adversarial network based on hierarchical learning with multi-reward text[J]. Journal of Yunnan University: Natural Sciences Edition, 2022, 44(1): 64-72. DOI: 10.7540/j.ynu.P00173
Citation: XUE Qi, MENG Xiang-fu, ZHANG Feng, ZHANG Xiao-yan, ZHU Jin-xia, ZHU Yao, WANG Dan-dan. HLMGAN: Generation adversarial network based on hierarchical learning with multi-reward text[J]. Journal of Yunnan University: Natural Sciences Edition, 2022, 44(1): 64-72. DOI: 10.7540/j.ynu.P00173

HLMGAN:分层学习的多奖励文本生成对抗网络

HLMGAN: Generation adversarial network based on hierarchical learning with multi-reward text

  • 摘要: 文本生成是自然语言处理的一项重要任务. 针对生成的文本大多缺乏多样性,且当生成文本过长时,文本生成的质量会有明显下降的问题,提出了一种采用Sentences and Words(SW)奖励机制的传递向量文本生成对抗网络. 首先,为生成器提出了层次结构设计,包括传递特征向量训练模块和生成向量训练模块,同时传递判别模型中真实文本特征向量给生成器的传递特征向量训练模块,由此来提高长文本生成的准确率,生成向量训练模块接收其生成词序列;然后,在训练过程中,使用关系存储核心代替传统的长短期记忆循环神经网络模型作为生成器,提高了模型的表达能力和捕获信息的能力;最后,采用SW奖励机制提高文本生成的多样性. 实验结果表明,分层学习的多奖励文本生成对抗网络(Generation Adversarial Network Based on Hierarchical Learning with Multi-reward Text,HLMGAN)模型在合成数据负对数似然度和双语互译质量评估辅助工具指标中均有所提升.

     

    Abstract: Text generation is a significant task in natural language processing. In order to solve the problem that most of the generated texts lack diversity and the quality of the generated texts is decreased (due to the long length of the generated texts), a transfer vector text generation adversarial network using the Sentences and Words (SW) reward mechanism was born. It is first for generator hierarchical structure design, including pass training module of characteristic vector and generates vector training module. The discriminant model in the actual text vector feature vector is passed to the generator transfer training module to improve the accuracy of long texts generated, generate vector training module receives its generated word sequence. Then, in the training process, Relational Memory Core (RMC) was used to replace the traditional long-short term memory model as the generator, which improved the model's expression ability and captured information. Finally, the SW reward mechanism is adopted to improve the diversity of text generation. The experimental results show that the evaluation indexes of the Generation Adversarial Network Based on Hierarchical Learning with Multi-reward Text(HLMGAN) model are improved compared with the existing models.

     

/

返回文章
返回