S-SimCSE: Sampled Sub-networks for Contrastive Learning of Sentence Embedding

Abstract

Contrastive learning has been studied for improving the performance of learning sentence embeddings. The current state-of-the-art method is the SimCSE, which takes dropout as the data augmentation method and feeds a pre-trained transformer encoder the same input sentence twice. The corresponding outputs, two sentence embeddings derived from the same sentence with different dropout masks, can be used to build a positive pair. A network being applied with a dropout mask can be regarded as a sub-network of itsef, whose expected scale is determined by the dropout rate. In this paper, we push sub-networks with different expected scales learn similar embedding for the same sentence. SimCSE failed to do so because they fixed the dropout rate to a tuned hyperparameter. We achieve this by sampling dropout rate from a distribution eatch forward process. As this method may make optimization harder, we also propose a simple sentence-wise mask strategy to sample more sub-networks. We evaluated the proposed S-SimCSE on several popular semantic text similarity datasets. Experimental results show that S-SimCSE outperforms the state-of-the-art SimCSE more than $1\%$ on BERT$_{base}$