I also thought about using GAN for text generation instead of image generation, but I didn’t know how to structure it. This slide summarizes it well and I will refer to it.
p.10 Problems in applying GAN to natural language series data
- Problem 1: When dealing with a series, the output of the Generator is a discrete value, making it difficult to train the Discriminator.
- reinforcement learning (policy gradient), in which the Generator learns as if it were a probabilistic policy
- Approximate the output of the Generator and make it differentiable for learning
- Question 2: What should I use as a Discriminator?
- CNN/RNN → More complex models …
- what constitutes a hostile sample?
I don’t know what you mean by question 1…
- What do you mean by “difficult to learn because of discrete values?”
- Is it because it’s a discrete value? Because the length of the output is indefinite?
p.12 SeqGAN: A classic paper on GAN for series data
-
Yu et al. SeqGAN: Sequence Generative Adversarial Nets with Policy Gradient (AAAI 2017)
-
point
-
- When dealing with a series, it is difficult to learn a discriminator because the output of the generator is a discrete value → Reinforcement learning is used to learn the generator
-
- Discriminator can only evaluate complete series → perform Monte Carlo search
- Discriminator can only evaluate complete series → perform Monte Carlo search
-
-
Interpret word sequences as “sequences of actions in reinforcement learning”.
-
Distribution of the probability of the occurrence of the next word after a word sequence
- The one called a language model in the field of natural language.
- In the field of reinforcement learning, they are called probabilistic measures.
-
- Looking at the next slide, it doesn’t look like it MUST be captured in the framework of reinforcement learning
p.16
- textGAN: Learning GANs for language generation without reinforcement learning
- Zhang et al. Feature Matching for Text Generation (ICML 2017; textGAN)
- point
-
- Differentiate the Discriminator by Soft-argmax instead of reinforcement learning
-
- Optimize in latent space rather than discrete words
-
p.17
p.18
p.19
I don’t know.
p.20 Better than SeqGAN, apparently.
p.22 MaskGAN
-
Can fill in the gaps in the series.
-
Mask a portion of the Encoder and have it fill in there.
- Some kind of Dropout, I guess.
This page is auto-translated from /nishio/言語処理におけるGANの展開 using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I’m very happy to spread my thought to non-Japanese readers.