A span-based joint named entity recognition (NER) and relation extraction model with RoBERTa
This code repository has been restructured based on JointIE for the purpose of dependency transition models. It mainly references the following models and codes:
Generalizing Natural Language Analysis through Span-relation Representations (ACL2020). [paper] (https://arxiv.org/abs/1911.03822)
LSTM/BERT-CRF Model for Named Entity Recognition (or Sequence Labeling) code
torch==2.2.0
Transformer==4.37.2
python==3.8.0
Train and evaluate model with default configure.（RoBERTa-Large, Learing rate 1e-5）
python transformers_trainer.py --dataset scierc| Dataset | NER (F1) | Relation (F1) |
|---|---|---|
| SciERC | 70.98（best） | 46.19 |
| SciERC | 69.87 | 47.27（best） |
| NYT24(NYT) | 96.52（best） | 84.90 |
| NYT24(NYT) | 96.21 | 85.06（best） |
| NYT29 | 内容6 | 内容6 |
| WebNLG | 97.94 | 92.64(best) |
| ACE2004_fold1 | 90.79 (best) | 49.79 |
| ACE2004_fold1 | 89.70 | 60.81 (best) |
| ACE2004_fold2 | 90.11(best) | 51.11 |
| ACE2004_fold2 | 88.04 | 53.98(best) |
| ACE2004_fold3 | 90.00(best) | 42.39 |
| ACE2004_fold3 | 88.86 | 46.77(best) |
| ACE2004_fold4 | 92.03(best) | 46.72 |
| ACE2004_fold4 | 91.04 | 48.67(best) |
| ACE2004_fold5 | 88.32(best) | 45.16 |
| ACE2004_fold5 | 87.26 | 46.35(best) |
| ACE2004_NER（avg best） | 90.25 | 47.03 |
| ACE2004_RE（avg best） | 88.98 | 51.32 |
| ACE2005_NRE(best) | 90.12 | 62.43 |
| ACE2005_RE(best) | 89.51 | 64.84 |