[Adapt] [Seminar]RelGAN: Relational Generative Adversarial Network For Text Generation

[任思宇]

Hello Adapters,


Generative adversarial networks have achieved promising results on generating continuous realistic data, such as image and audio.However, due to the discrete nature of textual data, the text generation still remain a challenging task for modern GANs.


In this paper, the authors propose a novel GAN architecture that mainly consists of three components to tackle existing mode collapse and non-differentiability issue: 1) a relational memory-based generator; 2) gumbel-softmax relaxation and 3) multiple embedded representation in discriminator. Relational memory generator and multiple representation in discriminator are introduced mainly to alleviate mode collapsed problem that is commonly observed in current textual GANs. Gumbel-softmax relaxation is used to overcome the notorious non-differentiability issue.


Experimental results demonstrate that RelGAN outperforms previous dominant textual GANs in terms of both sample quality and diversity.

Time: Wed 4:30pm
Venue: SEIEE 3-414

See you there!
Best Regards,
Roy
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://cs.sjtu.edu.cn/pipermail/adapt/attachments/20191008/d774145e/attachment.html>