Our Paper Sentence Representation Learning with Generative Objective rather than Contrastive Objective is accepted by the main conference of EMNLP 2022. In this repo, we provide the code for training in the unsupervised setting.
pytorch=1.10.1, python=3.7.11
for other dependencies:
pip install -r requirements.txt
We provide both the raw data and the augmened data we used in our paper in data/raw.csv and data/augmented.csv. This data file is processed using the code from EDA. We preprocess data/raw.csv using synoynm replacement with a ratio of 0.5.
bash scripts/run_unsup_example.sh.
bash eval.sh
The performance in the unsupervised setting is a little unstable. Different machines may lead to different results. In our experiments we use 24G RTX and 80G A100. Both machines lead to an average performance above 76.15.
| Device | STS12 | STS13 | STS14 | STS15 | STS16 | STSBenchmark | SICKRelatedness | Avg. |
|---|---|---|---|---|---|---|---|---|
| A100 | 70.75 | 83.53 | 72.78 | 83.69 | 77.64 | 79.27 | 65.37 | 76.15 |
| RTX | 70.21 | 83.88 | 73.06 | 83.87 | 77.60 | 79.19 | 65.31 | 76.16 |