The code and data for "Can Language Models Be Specific? How?"

S-TEST is a benchmark for measuring the specificity of the language of pre-trained language models.

Currently, S-TEST contains a set of connectors to the following pre-trained language models.

• GPT-2 (Radford et al., 2019)
• BERT-Base (Devlin et al., 2019)
• BERT-Large (Devlin et al., 2019)
• RoBERTa-Base (Liu et al., 2019)
• RoBERTa-Large (Liu et al., 2019)

This repo is build upon the LAMA benchmark.

To reproduce the results:

conda create -n stest37 -y python=3.7 && conda activate stest37
python setup.py install
pip install -r requirements.txt

Install spacy model

python3 -m spacy download en

Download the models

chmod +x download_models.sh
./download_models.sh

The script will create and populate a pre-trained_language_models folder. If you are interested in a particular model please edit the script.

python scripts/run_experiments.py

Temporary results will be logged in output/ and last_results.csv

python eval.py

The details of this repo are described in the following paper. If you find this repo useful, please kindly cite it:

@article{huang2022can,
  title={Can Language Models Be Specific? How?},
  author={Huang, Jie and Chang, Kevin Chen-Chuan and Xiong, Jinjun and Hwu, Wen-mei},
  journal={arXiv preprint arXiv:2210.05159},
  year={2022}
}