How to extract the best candidate after token classification?

Let’s assume the model predicts the following for an input sequence.

The O            
creation. O
date O
is O
2020 I-DATE
and  O
update. O
date.  O
is  O
01-09-2020 B-DATE

How do you pick the best candidate for creation date from logist values?

Can you clarify what you want to achieve? What do you mean by candidate here?

I am trying to extract the best date candidate here, but there are multiple date candidates here, 27 Aug 2020 and 01-09-2020. What is the best was way to aggregate confidence scores across for the candidate 27, Aug and 2020. Should I take average confidence scores for the split tokens (27, Aug, 2020) and compare against 01-09-2020 to decide which is the best candidate for creation date?

A classification model will not help you for that. It was not trained to answer this kind of question.

Maybe you need a syntactic parser to tell which date is associated with “The creation date”?