Awesome

PAIE (Prompting Argument Interaction for Event Argument Extraction)

This is the implementation of the paper Prompt for Extraction? PAIE: Prompting Argument Interaction for Event Argument Extraction. ACL'2022.

Quick links

• Overview
• Preparation
  • Environment
  • Data
• Run the model
  • Quick start
  • Experiments with multiple runs
  • Without bipartite loss
  • Joint/Single prompts
  • Manual/Concat/Soft prompts
  • Few-shot setting
• Citation

Overview

In this work we present PAIE: a simple, effective and low resource-required approach for sentence-/document-level event argument extraction. We formulate our contribution as follow.

1. We formulate and investigate prompt tuning under extractive settings.
2. We extract multiple roles using a joint prompt once a time. It not only considers the interaction among different roles but also reduce time complexity significantly.

Preparation

Environment

To run our code, please install all the dependency packages by using the following command:

pip install -r requirements.txt

Data

We conduct experiments on three common datasets: ACE05, RAMS and WIKIEVENTS.

• ACE05: This dataset is not freely available. Access from LDC and preprocessing following EEQA (2020'EMNLP).
• RAMS / WIKIEVENTS: We write a script for you for data processing. Run the following commands in the root directory of the repo.

bash ./data/download_dataset.sh

Please make sure your data folder structure as below.

data
  ├── ace_eeqa
  │   ├── train_convert.json
  │   ├── dev_convert.json
  │   └── test_convert.json
  ├── RAMS_1.0
  │   └── data
  │       ├── train.jsonlines
  │       ├── dev.jsonlines
  │       └── test.jsonlines
  ├── WikiEvent
  │   └── data
  │       ├── train.jsonl
  │       ├── dev.jsonl
  │       └── test.jsonl
  ├── prompts
  │   ├── prompts_ace_full.csv
  │   ├── prompts_wikievent_full.csv
  │   └── prompts_rams_full.csv
  └── dset_meta
      ├── description_ace.csv
      ├── description_rams.csv
      └── description_wikievent.csv

Run the model

Quick start

You could simply run PAIE with following commands:

bash ./scripts/train_{ace|rams|wikievent}.sh

Folders will be created automatically to store:

1. Subfolder checkpoint: model parameters with best dev set result
2. File log.txt: recording hyper-parameters, training process and evaluation result
3. File best_dev_results.log/best_test_related_results.log: showing prediction results of checkpoints on every sample in dev/test set.

You could see hyperparameter setting in ./scripts/train_[dataset].sh and config_parser.py. We give most of hyperparameters a brief explanation in config_parser.py.

Above three scripts train models with BART-base. If you want to train models with BART-Large, please change --model_name_or_path from facebook/bart-base to facebook/bart-large or run following commands:

bash ./scripts/train_{ace|rams|wikievent}_large.sh

Experiments with multiple runs

Table.3 of our paper shows the fluctuation of results due to random seed and other hyperparameters (learning rate mainly). You could run experiments multiple times to get a more stable and reliable results.

for seed in 13 21 42 88 100
do
    for lr in 1e-5 2e-5 3e-5 5e-5
    do
        bash ./scripts/train_{ace|rams|wikievent}.sh $seed $lr
    done
done

Each run will take ~4h so we highly recommend you to execute above command in parallel way.

Without-bipartite-loss

You could run PAIE without bipartite matching loss by delete the command argument --bipartite or run following commands:

bash ./scripts/train_{ace|rams|wikievent}_nobipartite.sh

Joint-prompt-or-not

Unlike multiple prompt strategy in PAIE, you could also prompt argument using template containing only one role (single prompt). Try it by changing --model_type from paie to base and set proper hyperparameters: --max_span_num, --max_dec_seq_length and --th_delta. Alternatively you could run following commands directly with hyperparameters we tuned:

bash ./scripts/train_{ace|rams|wikievent}_singleprompt.sh

Manual-prompt-or-others

Besides manual prompt, provide another two joint-prompt choices as described in Section 3.2 of our paper. We concelude them in the following:

1. (Default setting) Manual Prompt: All roles are connected manually with natural language
2. Concatenation Prompt: To concatenate all role names belonging to one event type.
3. Soft Prompt: Following previous work about continuous prompt, we connect different roles with learnable, role-specific pseudo tokens.

Run following commands if you want to try Concatenation Prompt:

bash ./scripts/train_{ace|rams|wikievent}_concatprompt.sh

Run following commands if you want to try Soft Prompt:

bash ./scripts/train_{ace|rams|wikievent}_softprompt.sh

Few-shot-setting

PAIE also performs well under low-annotation scenario. You could try it by set hyperparameters --keep_ratio to a number between 0 to 1, which controls the resampling rate from the original training examples. Simply you could also run scripts below:

KEEP_RATIO=0.2 bash ./scripts/train_{ace|rams|wikievent}_fewshot.sh

Note you could adjust the KEEP_RATIO value by yourself.

Citation

Please cite our paper if you use PAIE in your work:

@inproceedings{ma-etal-2022-prompt,
    title = "{P}rompt for Extraction? {PAIE}: {P}rompting Argument Interaction for Event Argument Extraction",
    author = "Ma, Yubo  and
      Wang, Zehao  and
      Cao, Yixin  and
      Li, Mukai  and
      Chen, Meiqi  and
      Wang, Kun  and
      Shao, Jing",
    booktitle = "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = may,
    year = "2022",
    address = "Dublin, Ireland",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.acl-long.466",
    doi = "10.18653/v1/2022.acl-long.466",
    pages = "6759--6774",
}