Eider

Code for ACL 2022 Finding paper "EIDER: Empowering Document-level Relation Extraction with Efficient Evidence Extraction and Inference-stage Fusion"

Dataset

The DocRED dataset can be downloaded following the instructions at link. Noted that the dev.json file has been modified in Aug, 2021. The modified version contains 998 documents. We use the original version of dev.json, which contains 1000 documents.

The expected structure of files is:

Eider
 |-- dataset
 |    |-- docred
 |    |    |-- train_annotated.json        
 |    |    |-- train_distant.json
 |    |    |-- dev.json
 |    |    |-- test.json
 |-- meta
 |    |-- rel2id.json

Coreference Results Required by Eider_rule

We use hoi as the coreference model for Eider_rule. The processed data can be found here.

The expected structure of files is:

Eider
 |-- coref_results
 |    |-- train_annotated_coref_results.json
 |    |-- dev_coref_results.json
 |    |-- test_coref_results.json

Training and Inference

Train Eider-BERT on DocRED with the following commands:

>> bash scripts/train_bert.sh eider test hoi
>> bash scripts/test_bert.sh eider test hoi

Alternatively, you can train Eider-RoBERTa using:

>> bash scripts/train_roberta.sh eider test hoi
>> bash scripts/test_roberta.sh eider test hoi

The commands for Eider_rule is similar:

>> bash scripts/train_bert.sh eider test hoi # BERT
>> bash scripts/test_bert.sh eider test hoi

>> bash scripts/train_roberta.sh eider test hoi # RoBERTa
>> bash scripts/test_roberta.sh eider test hoi

Citation

If you make use of this code in your work, please kindly cite the following paper:

@inproceedings{xie2021eider,
      title={EIDER: Empowering Document-level Relation Extraction with Efficient Evidence Extraction and Inference-stage Fusion}, 
      author={Yiqing Xie and Jiaming Shen and Sha Li and Yuning Mao and Jiawei Han},
      year={2022},
      booktitle = {Findings of the 60th Annual Meeting of the Association for Computational Linguistics},
      publisher = {Association for Computational Linguistics},
}