12
$\begingroup$

In semi-supervised learning, there are hard labels and soft labels. Could someone tell me the meaning and definition of the two things?

$\endgroup$
1
  • 3
    $\begingroup$ Where have you come across these terms? Could you provide a source for that? $\endgroup$ Commented Dec 21, 2018 at 10:28

1 Answer 1

11
$\begingroup$

According to Galstyan and Cohen (2007), a hard label is a label assigned to a member of a class where membership is binary: either the element in question is a member of the class (has the label), or it is not.

A soft label is one which has a score (probability or likelihood) attached to it. So the element is a member of the class in question with probability/likelihood score of eg 0.7; this implies that an element can be a member of multiple classes (presumably with different membership scores), which is usually not possible with hard labels.

One use of soft labels in semi-supervised learning could be that the training set consists of hard labels; a classifier is trained on that using supervised learning. The classifier is then run on unlabelled data, and adds soft labels to the elements. This enlarged data set is then used for further training, where the algorithm can treat hard and soft labels differently.

$\endgroup$
1
  • $\begingroup$ Hi, My question is what models in NLP can treat differently the soft and the hard labels in the same dataset. Can they recognize the difference and treat the "hard" as the source of truth? A link also would be very helpfull about these models. Thank you! $\endgroup$ Commented Oct 18, 2022 at 7:50

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .