This repository contains the dataset of our EMNLP 2022 research paper Revisiting DocRED – Addressing the False Negative Problem in Relation Extraction.
DocRED is a widely used benchmark for document-level relation extraction. However, the DocRED dataset contains a significant percentage of false negative examples (incomplete annotation). We revised 4,053 documents in the DocRED dataset and resolved its problems. We released this dataset as: Re-DocRED dataset.
The Re-DocRED Dataset resolved the following problems of DocRED:
- Resolved the incompleteness problem by supplementing large amounts of relation triples.
- Addressed the logical inconsistencies in DocRED.
- Corrected the coreferential errors within DocRED.
The Re-DocRED dataset is located as ./data directory, the statistics of the dataset are shown below:
| 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 |
If you find our work useful, please cite our work as:
@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}
}