Why GAN can not be directly used in NLP?
Generating an image of the real-valued is done with a random vector mapping transformation is a continuous process. Thus the error discriminator may be counter-propagating to the generator.
In natural language processing, encoder generates text decoding process, the process of generating the word model is actually in the process of selected vocabulary words, it is based on the probability of words in the entire current output of the distribution network, select the largest word probabilities. The choice of words argmax process is a discrete process, is a non-conductive. Thus, not by gradient backpropagation D to G, so G is not updated parameters.
Text GAN solution:
1. Direct probability after softmax passed discriminator generator, not argmax sample can backpropagation
2. Strategy gradient policy network
3. Gumble-softmax