1.记录一次与大神们的关于GAN应用于NLP的讨论(简书)
http://www.jianshu.com/p/32e164883eab
总结
- 目前尝试的内容:纯文本生成,诗歌生成,唐诗生成,机器翻译,IR,中文分词,文本分类
- 主要思路:考虑使用D进行真假判别,然后用RL的policy gradient的方式来打分和更新
- 使用word2vec之类的连续向量,微调之后不能代表一个有意义的词语,可以考虑取最近的点,但是存在bias,对于部分生成的序列,十分难判断其在之后生成整个序列时的分数
- IR GAN 没有使用attention机制
- 生成文本的方法:gumble softmax,policy gradient
- GAN可能适合的方向:半监督学习,文本分类
- 后续各种文本生成+GAN都是用seqGAN的套路,另外一种做法是在文本自编码器中的编码层做对抗训练
2.SeqGAN(CSDN)
http://blog.csdn.net/yinruiyang94/article/details/77675586
论文:https://arxiv.org/abs/1609.05473
3.GAN在自然语言处理方面有哪些有趣的文章和应用?(知乎)
https://www.zhihu.com/question/54463527
4.Role of RL in Text Generation by GAN(强化学习在生成对抗网络文本生成中扮演的角色)(知乎)
https://zhuanlan.zhihu.com/p/29168803
5.好玩的文本生成
http://geek.csdn.net/news/detail/131622
6.Generative Adversarial Text to Image Synthesis
http://www.cnblogs.com/wangxiaocvpr/p/6015535.html
论文:https://arxiv.org/pdf/1605.05396.pdf
7.生成对抗网络(专知)
http://www.zhuanzhi.ai/topic/2001150162715950/awesome
必读paper
1.Sequence Generative Adversarial Nets with Policy Gradient
2.GANS for Sequences of Discrete Elements with the Gumbel-softmax Distribution Generating
3.Text via Adversarial Training
4.Generating Sentences from a Continuous Space
5.Generating Text with Deep Reinforcement Learning
6.Deep Reinforcement Learning for Dialogue Generation
7. Generating Text via Adversarial Training
8.Adversarial Learning for Neural Dialogue Generation
9.SeqGAN: Sequence Generative Adversarial Nets with Policy Gradient
10.GANs for sequence of discrete elements with the Gumbel-softmax distribution
11.Generative Adversarial Text to Image Synthesis