Questions about the code semantic about colbert

[okhat]

It's just standard cross-entropy with softmax, over two classes, with the labels corresponding to the positive passages.

So, yes, this is binary classification between the positive and negative examples. Notice this is different from pointwise binary classification (namely, classifying each passage as positive or negative in isolation).

[okhat]

Hi Kyoungrok! Thanks for your question.

For optimization, we treat the task as binary classification between score(query, +ve passage) and score(query, -ve passage). The scores of the positive passages are placed in position zero (and the negatives in position one), hence the labels are all zeros.

[kyoungrok0517]

Thanks for the answer. As I observe out = (bsize, 2) and labels = (bsize) with zeros. Following your explanation If I do F.cross_entropy(out, labels) then the labels evaluated only on the first column (positive samples) so we have binary classification effect? I think I'm lack of knowledge how this cross_entropy function works.