Deep Continuous Prompt for Contrastive Learning of Sentence Embeddings

The performance of sentence representation has been remarkably improved by the framework of contrastive learning. However, recent works still require full fine-tuning, which is quite inefficient for large-scaled pre-trained language models. To this end, we present a novel method which freezes the whole language model and only optimizes the prefix deep continuous prompts. It not only tunes around 0.1% parameters of the original language model, but avoids the cumbersome computation of searching handcrafted prompts. Experimental results show that our proposed DCPCSE outperforms the state-of-the-art method SimCSE by a large margin. We raise the performance of unsupervised BERT$_{base}$ and supervised RoBERTa$_{large}$ by 2.24 and 1.00 points, respectively. Our code is publicly avaliable at https://github.com/YJiangcm/DCPCSE