How to separate the documents that need to be classified from the rest before classifying the text?

Question

big data

There are certain topics that the classifier is trained to separate, and there are topics about which he knows nothing. If the classifier is given topics unfamiliar to him, he will find some answer, but it will not be correct.
I suppose that it is possible to build a common vector for each of the known topics and, before classifying, compare the document vector with the topic vectors to find a deviation, where some value can be considered a threshold.
But since the number of words is high and large texts can contain many different words, I doubt this option.
Are there any better suggestions?

Answer 1

well, if you want to remove non-essential text -> collect the most commonly used common words from all documents, then first remove them from all texts, thereby leaving only the most specific text.