Prompt&Paraphrase

Repository for COLING 2022 paper "Prompt Combines Paraphrase: Teaching Pre-trained Models to Understand Rare Biomedical Words"

Preparation for data

We conducted our experiments on two biomedical datasets MedNLI and MedSTS. Both datasets are ONLY accessible with the corresponding confidentiality agreement signed.

MedNLI

https://physionet.org/content/mednli/1.0.0/

MedSTS

https://arxiv.org/abs/1808.09397

The MedSTS can be accessed by E-mail contact with the data maintainers.

Rare biomedical words and paraphrases

As introduced in the paper, the rare biomedical words are retrieved through an API provided by an online dictionary - "Wiktionary".

We consider the words with tags including 'medical', 'medicine', 'disease', 'symptom', and 'pharma-' as biomedical words. We loop through the common pre-training corpora for biomedical pre-trained models to obtain the word frequency for all possible words. Rare words are the words whose frequency in the pre-training corpora is smaller than a manually-selected threshold. Here, we provide the rare biomedical word list with a threshold of 200k and the corresponding paraphrases in the file paraphrase.txt.

Questions

If you have any question about our work, feel free to open an issue or email me (hcwang@ir.hit.edu.cn).

Citation

@inproceedings{wang-etal-2022-prompt,
    title = "Prompt Combines Paraphrase: Teaching Pre-trained Models to Understand Rare Biomedical Words",
    author = "Wang, Haochun and Liu, Chi and Xi, Nuwa and Zhao, Sendong and Ju, Meizhi and Zhang, Shiwei and Zhang, Ziheng and Zheng, Yefeng and Qin, Bing and Liu, Ting",
    booktitle = "Proceedings of the 29th International Conference on Computational Linguistics",
    month = oct,
    year = "2022",
    address = "Gyeongju, Republic of Korea",
    publisher = "International Committee on Computational Linguistics",
    url = "https://aclanthology.org/2022.coling-1.122",
    pages = "1422--1431",
    
}