UniRel

Released code for our EMNLP22 paper: UniRel: Unified Representation and Interaction for Joint Relational Triple Extraction.

Join the Discord if there are any questions.

Updates

• 2023-06-01
  • Add multi-token entity implementation.
  • Provide UniRel class in predict.py for easy inference and a checkpoint trained on nyt (multi-token) for trying.

Model

Results

Usage

Prerequisites

UniRel is implemented with Python == 3.8 and pytorch == 1.7.1, Other main requirments are:

• tdqm
• transformers == 4.12.5
• wandb

The detail requirments could be found at requirements.txt

Data

We obtain the data from TPLinker, please kindly refer to TPLinker officail repository. Change two filename of the download data:

• train_data.json -> train_split.json
• test_triples.json -> test_data.json

You can also download the data from here

Pretrained Model

We use the bert-base-cased model from Huggingface, you can download it by following their instrcution or let Transformers to automatically download. After that, place the files at the root directory of the project (./bert-base-cased).

I provided a checkpoint for trying predict. You can download here.

Train & Evalutaion

All parameter are listed in the script run_nyt.sh and run_webnlg.sh. By run with command bash run_nyt.sh can do train and evaluation.

Citation

@inproceedings{tang-etal-2022-unirel,
    title = "{U}ni{R}el: Unified Representation and Interaction for Joint Relational Triple Extraction",
    author = "Tang, Wei  and
      Xu, Benfeng  and
      Zhao, Yuyue  and
      Mao, Zhendong  and
      Liu, Yifeng  and
      Liao, Yong  and
      Xie, Haiyong",
    booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, United Arab Emirates",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.emnlp-main.477",
    pages = "7087--7099",
}

Have a nice day.