TaG

The official repository for ACL 2023 paper "A Novel Table-to-Graph Generation Approach for Document-Level Joint Entity and Relation Extraction".

Requirements

• Python (tested on 3.8.10)
• CUDA (tested on 11.2)
• PyTorch (tested on 1.12.1)
• Transformers (tested on 4.12.4)
• apex (tested on 0.1)
• opt-einsum (tested on 3.3.0)
• numpy
• scikit-learn
• scipy
• wandb
• ujson
• tqdm

Dataset

The DocRED dataset can be downloaded following the instructions at here. The Re-DocRED dataset is a revised version of DocRED and can be downloaded from here. The expected structure of files is:

TaG
 |-- dataset
 |    |-- docred/re-docred
 |    |    |-- train_annotated.json        
 |    |    |-- dev.json
 |    |    |-- test.json
 |    |    |-- rel_info.json
 |    |    |-- rel2id.json (in this repository)

We provide a copy of DocRED in data/docred. It is worth noting that TaG also need to generate intermediate data ({}-gc.json) from mention extraction prediction results, and we also provide the example files in data/docred.

Training & Evaluation

Since we detach the mention extraction (ME) stage from the end-to-end pipeline, you can train a TaG model in the following steps:

1. Run mention extraction with src/train_me.py. An example script is run_me.sh.
2. Preprocess span prediction with src/prepro.py. Specify the dev and test predictions with --dev_file and --test_file arguments respectively.
3. Run coreference resolution & relation extraction with src/train_tag.py. An example script is run_tag.sh.

The evaluation results are provided in the log. To evaluate RE result on test data, you should first save the model using --save_path argument before training. The model correponds to the best dev results will be saved. After that, You can evaluate the saved model by setting the --load_path argument, and the program will generate a test file result.json.

Acknowledgement

In this repository, we refer to and use some code from ATLOP. Thanks for their open-source efforts!🍻

Citation

@inproceedings{zhang-etal-2023-novel,
    title = "A Novel Table-to-Graph Generation Approach for Document-Level Joint Entity and Relation Extraction",
    author = "Zhang, Ruoyu and Li, Yanzeng and Zou, Lei",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.acl-long.607",
    doi = "10.18653/v1/2023.acl-long.607",
    pages = "10853--10865",
}