Code and experiments accompanying our paper Fine-Tuning Language Models on Multiple Datasets for Citation Intention Classification at EMNLP 2024
Our code are developed using PyTorch. Please see KimNLP.yml for dependencies.
To fine-tune a PLM for citation intention classification, please run
python main.py --dataset acl --lambdas 1 --data_dir Data/ --workspace Workspace/wksp --lm scibert
To jointly fine-tune a PLM on multiple datasets, run
python main.py --dataset acl-scicite-kim --lambdas 1-0.1-0.1 --data_dir Data/ --workspace Workspace/wksp --lm scibert
where datasets and the corresponding lambda values are concatenated using "-". The first dataset is the primary one while the others are the auxiliary datasets. Validation and Test F1 are evaluated on the primary dataset.
To compute the value of lambda using our TRL method, run
python trl.py --primary_dataset acl --auxiliary_dataset scicite --lm scibert --workspace Workspace/trl_wksp
We are currently resolving the copyright issue of the KIM dataset. We will release it shortly.
Please cite our paper if you find this repo useful:
@misc{shui2024deepcite,
title={Fine-Tuning Language Models on Multiple Datasets for Citation Intention Classification},
author={Zeren Shui and Petros Karypis and Daniel S. Karls and Mingjian Wen and Saurav Manchanda and Ellad B. Tadmor and George Karypis},
year={2024},
eprint={2410.13332},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2410.13332},
}