This GitHub repo belongs to the paper Prompt Tuning or Fine-Tuning - Investigating Relational Knowledge in Pre-Trained Language Models published on AKBC2021.
@inproceedings{
fichtel2021prompt,
title={Prompt Tuning or Fine-Tuning - Investigating Relational Knowledge in Pre-Trained Language Models},
author={Leandra Fichtel and Jan-Christoph Kalo and Wolf-Tilo Balke},
booktitle={3rd Conference on Automated Knowledge Base Construction},
year={2021},
url={https://openreview.net/forum?id=o7sMlpr9yBW},
doi={}
}
-
Create an environment:
conda create --name bertriple python=3.6.9
-
Install the requirements:
pip install -r requirements.txt
-
Run our experiments:
./run_experiments.sh
-
The evaluation results are under
results/
-
The plot and the tables of our paper:
Experiment 1: results/bertBCF_AUTOPROMPT41_common_all_obj_3_LAMA.csv results/bertBCF_AUTOPROMPT41_common_all_obj_3_LAMA_uhn.csv Experiment 2: bertBCF_AUTOPROMPT41_prec@1_3_runs.pdf Experiment 3: prec_per_props_bertBCF_AUTOPROMPT41_common_all_obj_3_LAMA_1.tex prec_per_props_bertBCF_AUTOPROMPT41_common_all_obj_3_LAMA_2.tex Experiment 4 (appendix): results/baselines_common_vocab/bertLC_LAMA.csv results/baselines_common_vocab/bertLC_LAMA_uhn.csv.csv results/bertLCF_AUTOPROMPT41_common_all_obj_3_LAMA.csv results/bertLCF_AUTOPROMPT41_common_all_obj_3_LAMA_uhn.csv results/baselines_common_vocab/distilbertBC_LAMA.csv results/baselines_common_vocab/distilbertBC_LAMA_uhn.csv results/distilbertBCF_AUTOPROMPT41_common_all_obj_3_LAMA.csv results/distilbertBCF_AUTOPROMPT41_common_all_obj_3_LAMA_uhn.csv results/baselines_different_vocab/robertaB_LAMA.csv results/baselines_different_vocab/robertaB_LAMA_uhn.csv results/robertaBF_AUTOPROMPT41_different_all_obj_3_LAMA.csv results/robertaBF_AUTOPROMPT41_different_all_obj_3_LAMA_uhn.csv results/baselines_different_vocab/facebookbartB_LAMA.csv results/baselines_different_vocab/facebookbartB_LAMA_uhn.csv results/facebookbartBF_AUTOPROMPT41_different_all_obj_3_LAMA.csv results/facebookbartBF_AUTOPROMPT41_different_all_obj_3_LAMA_uhn.csv