Exploring X-GEAR for Zero-Shot Cross-Lingual Event Argument Extraction

Reproducing ACL-2022 paper Multilingual Generative Language Models for Zero-Shot Cross-Lingual Event Argument Extraction.

Repo of the published paper: https://github.com/PlusLabNLP/X-Gear

Setup

• Python=3.8.6-ff

module load python/3.8.6-ff

• Create environment

python -m venv <env_name>
source <env_name>/bin/activate

env_name can be any name.

• install all required dependencies using requirements.txt.

pip install -r requirements.txt

Connect to a GPU

Run the following command to connect to a GPU and assign jobs in the Cluster.

salloc -p gpuq -q gpu --nodes=1 --ntasks-per-node=1 --gres=gpu:A100.80gb:1 --mem=90G --cpus-per-task=24

You can modify the arguments of this command based on specific needs and resources.

Training

• Run ./scripts/generate_data_ace05.sh to generate training examples of different languages for X-Gear. The generated training data will be saved in ./finetuned_data/.

• Run ./scripts/train_ace05.sh to train X-Gear. Alternatively, you can run the following command.

  python ./xgear/train.py -c ./config/config_ace05_mT5copy-base_en.json
  

  This trains X-Gear with mT5-base + copy mechanisim for ACE-05 English. The model will be saved in ./output/. You can modify the arguments in the config file. You can use ./config/config_ace05_mT5copy-base_ar.json and ./config/config_ace05_mT5copy-base_zh.json to train with copy mechanisim for ACE-05 Arabic and Chinese respectively.

Evaluating

• Run the following script to evaluate the performance for ACE-05 English, Arabic, and Chinese.

  ./scripts/eval_ace05.sh [model_path] [prediction_dir]
  

  model_path would be best_model.mdl inside logs(having year, date and time format) folder of output. prediction_dir could be any directory name of where you want to store the predictions.

Authors

  Ramaswamy Iyappan, Masters in Computer Science, riyappan@gmu.edu
  Abhijeet Banerjee, Masters in Computer Science, abanerj@gmu.edu
  Bhargava Canakapally, Masters in Computer Science, bcanakap@gmu.edu
  Title: Exploring X-GEAR for Zero-Shot Cross-Lingual Event Argument Extraction
  Organization: George Mason University
  Year: 2023