I and my co-authors recently wrote a blogpost on our ICLR 2021 paper “A Distributional Approach to Controlled Text Generation”…

I and my co-authors recently wrote a blogpost on our ICLR 2021 paper : A Distributional Approach to Controlled Text Generation. From this post, you should learn:

  • What we mean by distributionally controlled text generation, and how previous approaches have completely ignored this aspect of control.
  • How this task reduces to Exponential Family Energy-Based Models, and how we can leverage the EBM through a “distillation” process into an autoregressive policy for generation.
  • An overview of our results and experiments in the paper.
  • Finally, possibly yet unsolved problems with our technique.

The blogpost is on naver’s website here. Enjoy!