Can we control the output of neural generation systems?
Can we build a better chatbot using this control?
And what makes a good conversation anyway?

Check out the blog post for our NAACL paper:

https://t.co/UAAUxSeA0L

#AI #NLProc #DeepLearning #ConversationalAI @facebookai

This post got a little delayed since the paper's publication, but the intervening time has given me an opportunity to reflect.

So this post has some extra thoughts regarding more recent NLG developments - e.g., pretrained models like GPT(2), and the suitability of beam search.