DistillCSE: Distilled Contrastive Learning for Sentence Embeddings

Jiahao Xu, Wei Shao, Lihui Chen, Lemao Liu

Submission Type: Regular Long Paper

Submission Track: Semantics: Lexical, Sentence level, Document Level, Textual Inference, etc.

Keywords: sentence embedding, self-training, contrastive learning

TL;DR: This paper improves self-training for sentence embeddings by mitigating the issue induced by high variance teacher logits, and achieves the state-of-the-art performance.

Abstract: This paper proposes the DistillCSE framework, which performs contrastive learning under the self-training paradigm with knowledge distillation. The potential advantage of DistillCSE is its self-enhancing feature: using a base model to provide additional supervision signals, a stronger model may be learned through knowledge distillation. However, the vanilla DistillCSE through the standard implementation of knowledge distillation only achieves marginal improvements. The quantitative analyses demonstrate its reason that the standard knowledge distillation exhibits a relatively large variance of the teacher model's logits due to the essence of contrastive learning. To mitigate the issue induced by high variance, this paper accordingly proposed two simple yet effective solutions for knowledge distillation: a Group-P shuffling strategy as an implicit regularization and the averaging logits from multiple teacher components. Experiments on standard benchmarks demonstrate that the proposed DistillCSE outperforms many strong baseline methods and yields a new state-of-the-art performance.

Submission Number: 1012