Environment

Run the following command to create the required conda environment:

conda env create -f environment.yml -n your_new_environment_name

Data

• Train files:

  Run the following command in the data/ directory:

  bash download_hybrid.sh

• Evaluation files:

  Run the following command in the SentEval/data/downstream/ directory:

  bash download_dataset.sh

How we hold out 10% of the training data and how some data augmentations are performed are shown in tools.py.

Training

• Train with final performance in the "Wiki.STS_HT" training setting :

  bash scripts/train_bert_wiki_sts.sh

• Train with final performance in the "NLI.STS_HT" training setting :

  bash scripts/train_bert_nli_sts.sh

The scripts to perform experiments before Final Performance section are listed in scripts/data_domain.

Evaluation

bash scripts/evaluation.sh path_to_the_result

Plotting

How we plot figures in the paper are shown in plot.py.