Implementation of Document-level Relation Extraction with Knowledge Distillation and Adaptive Focal Loss - Findings of ACL 2022 (https://aclanthology.org/2022.findings-acl.132)

We would also like to highlight our subsequent work on the revision of DocRED. We have shown that the DocRED dataset is incompletely annotated. There are more than 60% of the triples that are not annotated in the evaluation split of DocRED. Therefore, it may not serve as a fair evaluation to the task of document-level relation extraction.

We would like to recommend to use the Re-DocRED dataset for this task. This dataset is a revised version of the original DocRED dataset and resolved the false negative problem in DocRED. Models trained and evaluated on Re-DocRED gains around 13 F1 compared to DocRED. The Re-DocRED dataset can be downloaded at : https://github.com/tonytan48/Re-DocRED. The leaderboard of Re-DocRED is hosted on Paperswithcode: https://paperswithcode.com/sota/relation-extraction-on-redocred. The files contained in Re-DocRED are:

root
 |-- data
 |    |-- train_revised.json        
 |    |-- dev_revised.json
 |    |-- test_revised.json

• Python (tested on 3.7.4)
• CUDA (tested on 10.2)
• PyTorch (tested on 1.10.2)
• Transformers (tested on 4.8.2)
• numpy (tested on 1.19.4)
• apex (tested on 0.1)
• opt-einsum (tested on 3.3.0)
• axial-attention (tested on 0.6.1)
• ujson
• tqdm

The DocRED dataset can be downloaded following the instructions at https://github.com/thunlp/DocRED

root
 |-- dataset
 |    |-- docred
 |    |    |-- train_annotated.json        
 |    |    |-- train_distant.json
 |    |    |-- dev.json
 |    |    |-- test.json
 |    |    |-- wikidata-properties.csv
 

 |-- meta
 |    |-- rel2id.json
 

Train the BERT model on DocRED with the following command:

Step 1: Training Teacher Model

>> bash scripts/batch_roberta.sh  # for RoBERTa

Step 2: Inference logits for the distantly supervised data

>> bash scripts/inference_logits_roberta.sh  

Step 3: Pre-train the student model

>> bash scripts/knowledge_distill_roberta.sh  

Step 4: Continue fine-tuning on the human annotated dataset.

>> bash scripts/continue_roberta.sh  

The program will generate a test file --output_name in the official evaluation format. You can compress and submit it to Codalab for the official test score.

Our pre-trained models at each stage can be found at: https://drive.google.com/drive/folders/1Qia0lDXykU4WPoR16eUtEVeUFiTgEAjQ?usp=sharing You can download the models and make use of the weights for inference/training.

Evaluating the trained models.

>> bash scripts/eval_roberta.sh  

Part of the code is adapted from ATLOP: https://github.com/wzhouad/ATLOP.

If you find our work useful, please cite our work as:

@inproceedings{tan-etal-2022-document,
    title = "Document-Level Relation Extraction with Adaptive Focal Loss and Knowledge Distillation",
    author = "Tan, Qingyu  and
      He, Ruidan  and
      Bing, Lidong  and
      Ng, Hwee Tou",
    booktitle = "Findings of ACL",
    year = "2022",
    url = "https://aclanthology.org/2022.findings-acl.132",


}

@inproceedings{tan2022revisiting,
  title={Revisiting DocRED – Addressing the False Negative Problem in Relation Extraction},
  author={Tan, Qingyu and Xu, Lu and Bing, Lidong and Ng, Hwee Tou and Aljunied, Sharifah Mahani},
  booktitle={Proceedings of EMNLP},
  url={https://arxiv.org/abs/2205.12696},
  year={2022}
}