S-SimCSE: Sampled Sub-networks for Contrastive Learning of Sentence Embedding
This is an incremental improvement for natural language processing tasks involving semantic text similarity.
The paper tackled improving sentence embedding performance by enhancing contrastive learning with sampled dropout rates, achieving over 1% improvement on BERT base compared to the state-of-the-art SimCSE.
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$_{base}$