• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

计算机工程与科学 ›› 2020, Vol. 42 ›› Issue (09): 1704-1710.

• 人工智能与数据挖掘 • 上一篇    

引入自编码机制对抗网络的文本生成模型

韩虎1,2,孙天岳1,赵启涛1   

  1. (1.兰州交通大学电子与信息工程学院,甘肃 兰州 730070;

    2.甘肃省人工智能与图形图像工程研究中心,甘肃 兰州 730070)

  • 收稿日期:2019-11-11 修回日期:2020-02-22 接受日期:2020-09-25 出版日期:2020-09-25 发布日期:2020-09-25
  • 基金资助:
    国家社会科学基金(17BXW071);国家自然科学基金(61562057);甘肃省科技计划(18JR3RA104)

Generative adversarial networks with autoencoder for text generation

HAN Hu1,2,SUN Tian-yue1,ZHAO Qi-tao1   

  1. (1.School of Electronic & Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070;

    2.Gansu Provincial Engineering Research Center for 

    Artificial Intelligence and Graphic & Image Processing,Lanzhou 730070,China)

  • Received:2019-11-11 Revised:2020-02-22 Accepted:2020-09-25 Online:2020-09-25 Published:2020-09-25

摘要: 自编码模型是一种无监督的学习算法,主要用于数据的降维和特征提取。在对抗神经网络模型基础上引入自编码模型,旨在提高输入数据的特征表示。主要使用前馈神经网络和Seq2seq模型学习原文本特征,将随机数据变为具有特征的数据作为输入,加快训练的速度,提高模型的准确率。同时使用强化学习模型解决反向传播中离散化数据梯度难以下降的问题。模型的鉴别器使用CNN网络和交叉熵模型,使生成的文本具有新颖性。使用BELU评价指标和主观评价的实验结果均表明了该模型的有效性。


关键词: 生成对抗神经网络, 自编码模块, 强化学习, 交叉熵

Abstract: Autoencoder is an unsupervised learning algorithm, mainly used for data dimensionality reduction and feature extraction. Based on adversarial neural network model, autoencoder is introduced to improve the feature representation of input data. Feedforward neural network and Seq2seq model are mainly used to learn the source text features, and the random data are transformed into characteristic data as input, which greatly accelerates the speed and accuracy of training. At the same time, reinforcement learning model is used to solve the problem that the gradient of discretized data is difficult to descend. The discriminator of the model uses CNN network and cross entropy model to make the generated text innovative and novel. In the experimental part, the results of automatic evaluation and subjective evaluation show that the model is effective.

Key words: generative adversarial neural network, autoencoder, reinforcement learning, cross entropy