Implement CoSENT loss

[tomaarsen]

Hello!

Context

This issue describes a feature that I am planning to be included in a release before v3, or alternatively, in v3 of Sentence Transformers.

Details

Some recent work by Li & Li, 2023 brought my attention to CoSENT (blogpost, GitHub repo). It proposes a new loss function with the purpose of more effectively and directly causing the cosine similarity of a positive pair to be larger than that of a negative pair.

Implementation snippets:

• in Keras: https://github.com/bojone/CoSENT/blob/124c368efc8a4b179469be99cb6e62e1f2949d39/cosent.py#L79-L90
• in Torch: https://github.com/SeanLee97/AnglE/blob/ce6d402c794fe3ddd1c1c08a756124cd04f6bb65/angle_emb/angle.py#L67-L77

@johneckberg has expressed an interest in contributing this loss.

cc @bwanglzu @ir2718 @johneckberg @aamir-s18 as I know you're interested in my TODO list. John will lead this development, and I will review his PR.

• Tom Aarsen