annotations_creators:
- unknown
language_creators:
- unknown
language:
- fr
license:
- other
multilinguality:
- unknown
pretty_name: OrangeSum
size_categories:
- unknown
source_datasets:
- original
task_categories:
- summarization
task_ids:
- unknown
Dataset Card for GEM/OrangeSum
Dataset Description
- Homepage: [Needs More Information]
- Repository: https://github.com/Tixierae/OrangeSum
- Paper: https://aclanthology.org/2021.emnlp-main.740
- Leaderboard: N/A
- Point of Contact: [Needs More Information]
Link to Main Data Card
You can find the main data card on the GEM Website.
Dataset Summary
OrangeSum is a French summarization dataset inspired by XSum. It features two subtasks: abstract generation and title generation. The data was sourced from "Orange Actu" articles between 2011 and 2020.
You can load the dataset via:
import datasets
data = datasets.load_dataset('GEM/OrangeSum')
The data loader can be found here.
paper
Dataset Overview
Where to find the Data and its Documentation
Download
Paper
BibTex
@inproceedings{kamal-eddine-etal-2021-barthez,
title = "{BART}hez: a Skilled Pretrained {F}rench Sequence-to-Sequence Model",
author = "Kamal Eddine, Moussa and
Tixier, Antoine and
Vazirgiannis, Michalis",
booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
month = nov,
year = "2021",
address = "Online and Punta Cana, Dominican Republic",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2021.emnlp-main.740",
doi = "10.18653/v1/2021.emnlp-main.740",
pages = "9369--9390",
abstract = "Inductive transfer learning has taken the entire NLP field by storm, with models such as BERT and BART setting new state of the art on countless NLU tasks. However, most of the available models and research have been conducted for English. In this work, we introduce BARThez, the first large-scale pretrained seq2seq model for French. Being based on BART, BARThez is particularly well-suited for generative tasks. We evaluate BARThez on five discriminative tasks from the FLUE benchmark and two generative tasks from a novel summarization dataset, OrangeSum, that we created for this research. We show BARThez to be very competitive with state-of-the-art BERT-based French language models such as CamemBERT and FlauBERT. We also continue the pretraining of a multilingual BART on BARThez{'} corpus, and show our resulting model, mBARThez, to significantly boost BARThez{'} generative performance.",
}
Has a Leaderboard?
no
Languages and Intended Use
Multilingual?
no
Covered Languages
French
License
other: Other license
Primary Task
Summarization
Credit
Dataset Structure
Dataset in GEM
Rationale for Inclusion in GEM
Similar Datasets
no
GEM-Specific Curation
Modificatied for GEM?
no
Additional Splits?
no
Getting Started with the Task
Pointers to Resources
Papers about abstractive summarization using seq2seq models:
- Abstractive Text Summarization using Sequence-to-sequence RNNs and Beyond
- Get To The Point: Summarization with Pointer-Generator Networks
- BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension
- BARThez: a Skilled Pretrained French Sequence-to-Sequence Model
Papers about (pretrained) Transformers:
- Attention is All you Need
- BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Technical Terms
No unique technical words in this data card.
Previous Results
Previous Results
Measured Model Abilities
The ability of the model to generate human like titles and abstracts for given news articles.
Metrics
ROUGE, BERT-Score
Proposed Evaluation
Automatic Evaluation: Rouge-1, Rouge-2, RougeL and BERTScore were used.
Human evalutaion: a human evaluation study was conducted with 11 French native speakers. The evaluators were PhD students from the computer science department of the university of the authors, working in NLP and other fields of AI. They volunteered after receiving an email announcement. the best-Worst Scaling (Louviere et al.,2015) was used. Two summaries from two different systems, along with their input document, were presented to a human annotator who had to decide which one was better. The evaluators were asked to base their judgments on accuracy (does the summary contain accurate facts?), informativeness (is important in-formation captured?) and fluency (is the summary written in well-formed French?).
Previous results available?
no
Broader Social Context
Previous Work on the Social Impact of the Dataset
Usage of Models based on the Data
no
Impact on Under-Served Communities
Addresses needs of underserved Communities?
no
Discussion of Biases
Any Documented Social Biases?
no
Are the Language Producers Representative of the Language?
The dataset contains news articles written by professional authors.
Considerations for Using the Data
PII Risks and Liability
Licenses
Copyright Restrictions on the Dataset
open license - commercial use allowed
Copyright Restrictions on the Language Data
open license - commercial use allowed