CUP

This repository is for CUP: Curriculum Learning based Prompt Tuning for Implicit Event Argument Extraction

Dataset

Download RAMS and WikiEvent datasets into ./data folder.

Checkpoints

Our checkpoints are available here.

Download them into ./experiments

Evaluation

Evaluate performance on RAMS

sh script/test_RAMS.sh

Evaluate performance on WikiEvents

sh script/test_Wikievents.sh

Reproduce the results by yourself

In training stage, we utilize documental AMR graph. Hence, preprocess data first.

Transition-based AMR decoder

Follow the instructions in this repository to train an AMR parser.

Then use ./data/amr_parser.py to parse the sentences of the two datasets.

Coreference Resolution

We utilize ready-made coreference resolution tool available here

Use your own AMR decoder or coreference resolutioner

Process training data into the same form as data/WikiEvents/amrs/train.amr.txt and data/WikiEvents/corefered.json

Combine coref with raw training data to build documental AMR graph

For RAMS:

python ./data/preprocess.py --train_dir=./data/RAMS/train.jsonlines --coref_dir=./data/RAMS/corefered.json --output_dir=./data/RAMS/RAMSwithcore/train.jsonl

For WikiEvents:

python ./data/preprocess.py --train_dir=./data/wikievents/informative/train.jsonl --coref_dir=./data/wikievents/corefered.json --output_dir=./data/wikievents/WikiwCoref/informative/train.jsonl

Training

For full data training, store all doc_keys into f'./{args.data_path}/doc_keys.jsonl'

To conduct few-shot training, first select a particular ratio of samples, then store their doc_keys into f'./{args.data_path}/doc_keys.jsonl'

Train on RAMS:

sh scripts/train_RAMS.sh

Train on WikiEvents:

sh scripts/train_WikiEvents.sh

Citation

Please cite our work if this respository inspires you.