Training State-of-the-Art Text Embedding & Neural Search Models

[@EduardoFloresV]

Great presentation. Very useful, thanks for putting the effort and sharing it!

[@枕头不软睡不香]

Very Great presentation!

[@boscojay1381]

16:00, I have heard you talk about the multiple negative ranking loss quiet often, is it the same as the SupCon loss or the contrastive loss? I also noticed in your simCSE paper implementation with sentence transformers library, you again use the MNR loss. Is it a special case of contrastive loss?? I appreciate your response to this question. Thanks

[@nomanshahid9771]

How do we use triplets with multiple negative ranking loss? Doesn't this loss only takes positive pairs?

[@ldrn-b4b]

Thanks! Great content

[@francescocariaggi1145]

Very interesting talk!

[@zurechtweiser]

Es gibt einige vortrainierte Modelle, die für cosine-sim optimiert sind, andere für das Skalarprodukt. Wie wirkt es sich auf das Ergebnis aus, das jeweils andere zu nutzen?

[@tianshuwang7543]

Hello, thank you for the awesome talk. I have a question why contrastive / triplet loss might only optimize the local structure? Don't these losses increase the distance between negative cases, which usually include random pairs in a batch?

[@NilsReimersTalks]

They increase the distance only for the pairs you provide. If you provide poor pairs, they might only optimize some local structures.

[@tianshuwang7543]

@@NilsReimersTalks Thx, I got your point.

[@muhammadhammadkhan1289]

Can you share the slides?

[@NilsReimersTalks]

Slides are here: nils-reimers.de/talks/2021-09-State-of-the-art-Bi-Encoders.zip

[@高长宽]

太棒了！