File size: 1,273 Bytes
f1c7827
8e85d9e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f1c7827
f96455a
193d9b3
f96455a
 
 
 
 
 
 
3fa10cf
f96455a
3fa10cf
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
---
language:
- en
license:
- mit
multilinguality:
- monolingual
pretty_name: DocEE
size_categories:
- 10K<n<100K
source_datasets:
- original
tags:
- wiki
- news
- event-detection
task_categories:
- text-classification
task_ids:
- multi-class-classification
---

This is a [DocEE dataset](https://github.com/tongmeihan1995/docee) adapted for easy use with HuggingFace ecosystem.

Dataset is described in [DocEE: A Large-Scale and Fine-grained Benchmark for Document-level Event Extraction](https://aclanthology.org/2022.naacl-main.291/) and can be downloaded from [here](https://github.com/tongmeihan1995/docee).

Train split contains 21949 documents while test split contains 5536 documents. In total, dataset contains 27485 documents classified into 59 event types.

## Differences from the original split(s)

Originally, the dataset is split into three splits: train, validation and test. For the purposes of this repository, original splits were joined back together and divided into train and test splits while making sure that splits were stratified across document sources (news and wiki) and event types.

Originally, the `title` column additionally contained `date` and `metadata` columns. They are now separated into three columns: `date`, `metadata` and `title`.