Evaluate GAN effectiveness for generating answer variations.

[hairuoguo]

It might be possible to introduce text/language variation to our model by training a GAN to generate new variations. This could be used in conjunction with AMT (having turkers act as discriminators).

Maybe other generative models (like a VAE) could work as well?

[brando90]

this could be a good idea, however, its important to prioritize the things that have a higher chance to succeed before doing something as ambitious "automatic generation of text". Right now there are reputable DL research groups working on things like this (and probably industrial labs with a lot of resources like google brain too) and right now they have not really succeeded in solving "natural language generation" and instead got some strong criticism like this:

https://medium.com/@yoav.goldberg/an-adversarial-review-of-adversarial-generation-of-natural-language-409ac3378bd7

so before we attempt to do this its important to get done the main features of the real framework like the code for users/students to generate questions automatically. For example:

https://github.com/brando90/eit_proj1/issues/26 https://github.com/brando90/eit_proj1/issues/27

I have created the priority label to identify urgent stuff.

[brando90]

there are other options too that seem more reliable, lets check the literature more carefully before attempting this (crowdsourcing could be good too if we plan it well, feel free to share ur ideas and plans here as u think about it)