I would like to know the practical threshold of the TF-IDF (just like the practical p-value cutoff of 0.1 or 0.05 in hypothesis tests).
I tried to look at it in some previous post, and some people suggest to use Zipf curve. As I understand, Zipf curve is referring to the corpus (or even just the document). Does it mean there is no universal cutoff for TF-IDF practically speaking? Can we just look at the TF-IDF value and know whether the term is significantly important to the corpus, without knowing the info of the corpus?