Closed EeyoreLee closed 1 year ago
Hi,
For now our framework does not support real-number similarities from the supervision data. You can either adjust the code to set different weights on different data examples, or truncate the similarity and only see pairs with high-enough similarities as positive pairs.
Thanks for your reply. In fact, I wanna sort a set of sentences by their similarity to another text. So the real-number similarities seems like important. Therefore, "truncate the similarity and only see pairs with high-enough similarities as positive pairs." may not be an suitable way. Like that, you also agree that shouldn't use the hard negative
and should adjust the code to support the idea?
Hi,
Yeah, hard negatives won't suit your need. Maybe contrastive learning is not a good objective in your case, and you can probably use a regression objective.
Hi,
Yeah, hard negatives won't suit your need. Maybe contrastive learning is not a good objective in your case, and you can probably use a regression objective.
Thanks. I will try some and feedback here if I adjust SimCSE and it's useful.
I saw there is a hard negative weight in SimCSE. Should I use it to give the low similarity a higher punishment. or I make some changes from loss. Like Facol Loss, but not it , to give the high similarity a higher weight. Looking forward to your reply and understanding of the matter. Thanks in advance! :)