DiffCSE: Difference-based Contrastive Learning for Sentence Embeddings

TL;DR

DiffCSE introduces a novel unsupervised contrastive learning method that emphasizes differences between original and edited sentences to produce more sensitive and effective sentence embeddings, achieving state-of-the-art results.

Contribution

It presents DiffCSE, a new equivariant contrastive learning framework that enhances sentence embeddings by focusing on meaningful differences caused by sentence edits.

Findings

Outperforms unsupervised SimCSE by 2.3 points on semantic textual similarity.

Achieves state-of-the-art results among unsupervised sentence embedding methods.

Demonstrates effectiveness of difference-based contrastive learning.

Abstract

We propose DiffCSE, an unsupervised contrastive learning framework for learning sentence embeddings. DiffCSE learns sentence embeddings that are sensitive to the difference between the original sentence and an edited sentence, where the edited sentence is obtained by stochastically masking out the original sentence and then sampling from a masked language model. We show that DiffSCE is an instance of equivariant contrastive learning (Dangovski et al., 2021), which generalizes contrastive learning and learns representations that are insensitive to certain types of augmentations and sensitive to other "harmful" types of augmentations. Our experiments show that DiffCSE achieves state-of-the-art results among unsupervised sentence representation learning methods, outperforming unsupervised SimCSE by 2.3 absolute points on semantic textual similarity tasks.