I'm doing NER.
If I use Softmax as the classifier, it's outputs are probabilities of all labels for a character.
I use standard method of training. But when making predictions, I don't want to use the standard way, 'find maximum probability', to determin the label.
I want to label a whole entity to 'O' if one of its character's label's maximum probability is below a threthold. In other words, I would rather miss an entity, than labeling a not-so-certain entity.
But I also want to try CRF as classifier. As far as I know, CRF seems to score a whole sentence's labeling. Is a single label's probability still meaningful in CRF? What can I do if I still want a entity-level threthold similar to Softmax?