SimCSE在中文上无监督 + 有监督 pytorch版
SimCSE：https://arxiv.org/pdf/2104.08821.pdf
ESimCSE: https://arxiv.org/pdf/2109.04380.pdf
1.database: SNS-B (uploaded)
directory: data/SNS-B/
2.environment
torch==1.8.2
transformers==4.12.3video card: 3060Ti 8G
Due to the limitation of the graphics card, the batch_size is set very small.
You can try increasing the batch_size to get better results with video memory allowed.
3.how to run?
SimCSE： python train.py
ESimCSE： python ESimCSE_train.py
4.Result (un-supervised)
spearman corrcoef is shown as result below:
| Model | un_supervised |
|---|---|
| Bert_base | 0.538 |
| SimCSE | 0.692 |
| ESimCSE | 0.707 |
说明：原论文的无监督SimCSE基于英文，从维基百科上挑了100万个句子进行训练的。本项目评测实验是在中文数据集STS-B(已上传)，实现结果以苏剑林科学空间结果 对照。
SimCSE结果与其一致。
以上供参考，码代码不易，有用请点个赞喔。