metadata

license: apache-2.0
language:
  - el
pipeline_tag: summarization
task_categories:
  - summarization
  - text-generation
  - text2text-generation
task_ids:
  - summarization
  - text-generation
  - title-generation
tags:
  - GreekNLP
  - Text Summarization
  - Text Generation
  - Title Generation
  - Greek
  - Wikipedia
pretty_name: Greek Wikipedia
size_categories:
  - 10K<n<100K

GreekWikipedia

A Greek abstractive summarization dataset collected from the Greek part of Wikipedia, which contains 93,433 articles, their titles and summaries.
This dataset has been used to train our best-performing model GreekT5-umt5-base-greekwikipedia as part of our research article:
Giarelis, N., Mastrokostas, C., & Karacapilidis, N. (2024) Greek Wikipedia: A Study on Abstractive Summarization For information about dataset creation, limitations etc. see the original article.

Supported Tasks and Leaderboards

This dataset supports:

Text summarization: Given the text of an article, a text generation model learns to generate an abstractive summary.
Title Generation: Given the text of an article, a text generation model learns to generate a post title.

Languages

All articles are written in Greek.

Dataset Structure

Data Instances

The dataset is structured as a .csv file, while three dataset splits are provided (train, validation and test).

Data Fields

The following data fields are provided for each split:

title: (str) A short title.
article: (str) The full text of the article.
summary: (str): The abstractive summary of the article.
url: (str) The URL which links to the original unprocessed article.

Data Splits

Split	No of Documents
Train	83,433
Validation	5000
Test	5000

Example code

from datasets import load_dataset

# Load the training, validation and test dataset splits.
train_split = load_dataset('IMISLab/GreekWikipedia', split = 'train')
validation_split = load_dataset('IMISLab/GreekWikipedia', split = 'validation')
test_split = load_dataset('IMISLab/GreekWikipedia', split = 'test')

print(test_split[0])

Contact

If you have any questions/feedback about the dataset please e-mail one of the following authors:

[email protected]
[email protected]
[email protected]

Citation

TBA