X-Gear: Multilingual Generative Language Models for Zero-Shot Cross-Lingual Event Argument Extraction

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

Setup

• Python=3.7.10

$ conda env create -f environment.yml

Data and Preprocessing

• Go into the folder ./preprocessing/
• If you follow the instruction in the README.md, then you can get your data in the folder ./processed_data/

Training

• Run ./scripts/generate_data_ace05.sh and ./scripts/generate_data_ere.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 or ./scripts/train_ere.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 or replace the config file with other files in ./config/.

Evaluating

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

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

  If you want to test X-Gear with mT5-large, remember to modify the config file in ./scripts/eval_ace05.sh.

• Run the following script to evaluate the performance for ERE English and Spanish.

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

  If you want to test X-Gear with mT5-large, remember to modify the config file in ./scripts/eval_ere.sh.

We provide our pre-trained models and show their performances as follows.

ACE-05

	en Arg-I	en Arg-C	ar Arg-I	ar Arg-C	zh Arg-I	zh Arg-C
X-Gear-ace05-mT5-base+copy-en	73.39	69.28	47.64	42.09	57.81	54.46
X-Gear-ace05-mT5-base+copy-ar	33.87	27.17	72.97	66.92	31.14	28.84
X-Gear-ace05-mT5-base+copy-zh	59.85	55.15	38.04	34.88	72.93	68.99
X-Gear-ace05-mT5-large+copy-en	75.16	71.85	54.18	50.00	63.14	58.40
X-Gear-ace05-mT5-large+copy-ar	38.81	34.57	73.49	67.75	39.26	36.13
X-Gear-ace05-mT5-large+copy-zh	61.44	55.40	38.71	36.14	70.45	66.99

ERE

	en Arg-I	en Arg-C	es Arg-I	es Arg-C
X-Gear-ere-mT5-base+copy-en	78.26	71.55	64.31	58.70
X-Gear-ere-mT5-base+copy-es	69.21	59.79	70.67	66.37
X-Gear-ere-mT5-large+copy-en	78.10	73.04	64.82	60.35
X-Gear-ere-mT5-large+copy-es	69.03	63.73	71.47	68.49

Citation

If you find that the code is useful in your research, please consider citing our paper.

@inproceedings{acl2022xgear,
    author    = {Kuan-Hao Huang and I-Hung Hsu and Premkumar Natarajan and Kai-Wei Chang and Nanyun Peng},
    title     = {Multilingual Generative Language Models for Zero-Shot Cross-Lingual Event Argument Extraction},
    booktitle = {Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (ACL)},
    year      = {2022},
}