tonytan48/Re-DocRED
收藏资源简介:
Re-DocRED数据集是为了解决DocRED数据集中的假阴性问题(即不完整标注)而创建的。该数据集通过补充大量关系三元组解决了DocRED的不完整性问题,解决了逻辑不一致性问题,并纠正了共指错误。Re-DocRED数据集包含3,053个训练文档、500个开发文档和500个测试文档,平均每个文档包含28.1个三元组、19.4个实体和7.9个句子。
The Re-DocRED dataset was constructed to address the false negative issue (i.e., incomplete annotation) inherent in the DocRED dataset. This dataset resolves the incompleteness of DocRED by supplementing a substantial number of relational triples, fixes logical inconsistencies, and rectifies coreference errors. It comprises 3,053 training documents, 500 development documents, and 500 test documents. On average, each document contains 28.1 relational triples, 19.4 entities, and 7.9 sentences.
Re-DocRED Dataset 概述
数据集描述
Re-DocRED Dataset 是对原始 DocRED 数据集的修订版本,旨在解决原始数据集中的 false negative 问题。该数据集由 EMNLP 2022 研究论文 Revisiting DocRED – Addressing the False Negative Problem in Relation Extraction 提出。
修订内容
- 解决不完全性问题:补充了大量关系三元组。
- 处理逻辑不一致:解决了 DocRED 中的逻辑矛盾。
- 修正共指错误:纠正了 DocRED 中的共指错误。
数据集统计
数据集位于 ./data 目录下,具体统计数据如下:
| Train | Dev | Test | |
|---|---|---|---|
| # Documents | 3,053 | 500 | 500 |
| Avg. # Triples | 28.1 | 34.6 | 34.9 |
| Avg. # Entities | 19.4 | 19.4 | 19.6 |
| Avg. # Sents | 7.9 | 8.2 | 7.9 |
引用信息
若使用此数据集,请引用以下文献: bibtex @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} }




