P-ICL

Data Preparation

Please save your dataset in data folder. Note that CoNLL2003 and WNUT2017 are open-source datasets, ACE2004 and ACE2005 are not free. We keep our CoNLL2003 and WNUT2017 train and test JSON files in data folder.

• CoNLL2003: https://huggingface.co/datasets/conll2003
• WNUT2017: https://huggingface.co/datasets/wnut_17
• ACE2004: https://catalog.ldc.upenn.edu/LDC2005T09
• ACE2005: https://catalog.ldc.upenn.edu/LDC2006T06

Generation

Please review generate.py for P-ICL and utils.py for prompts, and change some important parameters.

python generate.py --dataset your_dataset --mode your_mode --picl_cnt your_picl_cnt --icl_cnt your_icl_cnt

Evaluation

Please review eval.py for computing entity-level F1 score.

python eval.py --label_path your_label_file_path --pred_path your_model_output_file_path