Contrastive Learning: Relaxed Contextualized word Mover Distance (CLRCMD)

This repository reproduces the experimental result of CLRCMD (pronounced as "clear command") reported in the paper to be appeared in ACL 2022 main track.

0. Download checkpoints

We want to upload our checkpoint to model registry such as huggingface hub to make them easily accessible, but due to the complicated process, we decided to just manually upload the checkpoint to our gdrive. Please visit this link to download the checkpoints we used in our experiment. We assume the pytorch_model.bin and model_args.json is located in /home/username/checkpoints/bert-rcmd/

1. Prepare Environment

We assume that the user uses anaconda environment.

conda create -n clrcmd python=3.8
conda activate clrcmd
pip install -r requirements.txt
python setup.py develop

2. Prepare dataset

2-1. Semantic Textual Similarity benchmark (STS12, STS13, STS14, STS15, STS16, STSB, SICKR)

We download the benchmark dataset using the script provided by SimCSE repository.

bash examples/download_sts.sh

• tokenizer.sed: Tokenizer script used in download_sts.bash

2-2. Interpretable Semantic Textual Similarity benchmark (iSTS)

We create a script for downloading iSTS benchmarks.

bash examples/download_ists.sh

2-2-1. Correct wrong input format

• STSint.testinput.answers-students.sent1.chunk.txt
  • 252th example: from a closed path to a closed path.
  • 287th example: from has no gaps to [ has no gaps ]
  • 315th example: from is in a closed path, to [ is in a closed path, ]
  • 315th example: from is in a closed path. to [ is in a closed path. ]
• STSint.testinput.answers-students.sent1.txt
  • 287th example: battery terminal to battery terminal
  • 308th example: switch z, that to switch z, that
• STSint.testinput.answers-students.sent2.chunk.txt
  • 287th example: are not separated by the gap to [ are not separated by the gap ]
  • 315th example: are to [ are ]
  • 315th example: in closed paths to [ in closed path ]

2-3. NLI dataset tailored for self-supervised learning (SimCSE-NLI)

We download the training dataset using the script provided by SimCSE repository.

bash examples/download_nli.bash

3. Conduct experiments

3-1. Evaluate semantic textual similarity benchmark without any training

# Help message
python -m examples.run_evaluate_sts -h

# One example
python -m examples.run_evaluate_sts --data-dir data --model bert-rcmd

3-2. Train model using self-supervised learning (e.g. SimCSE, CLRCMD)

python -m examples.run_train --data-dir data --model bert-rcmd

3-2. Evaluate benchmark performance on the trained checkpoint

python -m examples.run_evaluate_sts --data-dir data --model bert-rcmd --checkpoint /home/username/checkpoints/bert-rcmd

3-3. Evaluate interpretable semantic textual similarity benchmark

# Filter out the alignments which has low score
python -m examples.run_preprocess_ists --alignment-path data/ISTS/test_goldStandard/STSint.testinput.images.wa

# Bert-avg
python -m examples.run_evaluate_ists --data-dir data/ISTS/test_goldStandard/ --source images --checkpoint-dir checkpoints/bert-avg/
./data/ISTS/test_goldStandard/evalF1.pl ./data/ISTS/test_goldStandard/STSint.testinput.images.wa.equi ./checkpoints/bert-avg/images.wa

# Bert-Clrcmd
python -m examples.run_evaluate_ists --data-dir data/ISTS/test_goldStandard/ --source images --checkpoint-dir checkpoints/bert-rcmd/
./data/ISTS/test_goldStandard/evalF1.pl ./data/ISTS/test_goldStandard/STSint.testinput.images.wa.equi ./checkpoints/bert-rcmd/images.wa

4. Report results

4-1. Semantic textual similarity benchmark