Generative Adversarial Network for NMT

I’m an undergraduate student in computer science. I’m going to implement a translator. I used OpenNMT-py with parallel corpus data previous time. Now I need to use only monolingual corpora data to implement a translator.
I hope to use GAN architecture for this.
Is it possible to use only monolingual data for GAN architecture?. Because I’m still reading papers and collecting information about this architecture. In addition to that, can someone please let me know is there any feature in OpenNMT-py for implement the GAN architecture.

You can use monolingual corpara to train a language model or use it for back-translation, but openNMT isn’t an unsupervised translation system. Have a look at Masked Sequence to Sequence Pre-training for Language Generation.

1 Like

Thank you for your valuable information. It seems that it is a good clue for my work. I will go through it.
If I use Generative Adversarial Network, then do I need to use parallel corpus data for pre-train the generator model or discriminator model?

GANs usually use labeled data and generated data to learn the model. You can use only monolingual data by mapping together the language spaces and enhance the translations with a GAN. But it is an open question how the GAN-method performs compared to the widely used back-translation.

1 Like