Text Generation "tricks"

In Lesson 4 it is said that the model is not made for text generation as the intention is classification.
However it is mentioned that if text generation was the purpose there are a few “tricks” that would make the generated text more consistent.

Which are those tricks? Can anyone point me out to some good practices for text generation with the knowledge obtained in Lesson 4 - NLP?


(Edit: I have seen other posts and beam search being mentioned but after playing a bit just using that instead of “predict” provides worse results. I guess I have to also do something else?)


Hi again,
Does anyone has any feedback for this? I am trying to generate new text, not to classify it.

Thanks a lot!

1 Like

Hii, I am also waiting to know the tricks and its implementation.
In another thread jeremy said something called "beam search’’ for text generation models.
But how to implement it?


have you seen https://youtu.be/3oEb_fFmPnY