• This repo contains the code and pre-trained model checkpoints for our EMNLP 2022 paper.
• Our code is based on the SimCSE.

We propose a method that makes high-quality positives for contrastive sentence representation learning. A pivotal ingredient of our approach is the use of prefix that attached to a pre-trained language model, which allows for differentiable data augmentation during contrastive learning. Our method can be summarized in two steps: supervised prefix-tuning followed by joint contrastive fine-tuning with unlabeled or labeled examples. The following figure is an overview of the proposed two-stage training strategy.

First, install PyTorch on the official website. All our experiments are conducted with PyTorch v1.8.1 with CUDA v10.1. So you may use the following code to download the same PyTorch version:

pip install torch==1.8.1+cu101 torchvision==0.9.1+cu101 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html

Then run the following script to install the remaining dependencies:

pip install -r requirements.txt

We use the same training and evaluation datasets as SimCSE. Therefore, we adopt their scripts for downloading the datasets.

To download the unlabeled Wikipedia dataset, please run

cd data/
bash download_wiki.sh

To download the labeled NLI dataset, please run

cd data/
bash download_nli.sh

To download the evaluation datasets, please run

cd SentEval/data/downstream/
bash download_dataset.sh

Following previous works, we use SentEval to evaluate our model.

We prepared two example scripts for reproducing our results under the semi-supervised and supervised settings respectively.

To train a semi-supervised model, please run

bash run_semi_sup_bert.sh

To train a supervised model, please run

bash run_sup_bert.sh

To evaluate the model, please run:

python evaluation.py \
    --model_name_or_path <model_checkpoint_path> \
    --mode <dev|test>

The results are expected to be shown in the following format:

*** E.g. Supervised model evaluatation results ***
+-------+-------+-------+-------+-------+--------+---------+-------+
| STS12 | STS13 | STS14 | STS15 | STS16 |  STSB  |  SICKR  |  Avg. |
+-------+-------+-------+-------+-------+--------+---------+-------+
| 77.40 | 85.24 | 80.50 | 86.85 | 82.59 | 84.12  |  80.29  | 82.43 |
+-------+-------+-------+-------+-------+--------+---------+-------+

We prepare two model checkpoints:

• diffaug-semisup-bert-base-uncased
• diffaug-sup-bert-base-uncased

Here is an example about how to evaluate them on STS tasks:

python evaluation.py \
    --model_name_or_path Tianduo/diffaug-semisup-bert-base-uncased \
    --mode test

Please cite our paper if it is helpful to your work:

@inproceedings{wang2022diffaug,
   title={Differentiable Data Augmentation for Contrastive Sentence Representation Learning},
   author={Wang, Tianduo and Lu, Wei},
   booktitle={Empirical Methods in Natural Language Processing (EMNLP)},
   year={2022}
}