Datasets:
metadata
license: cc-by-4.0
task_categories:
- summarization
language:
- ur
tags:
- Urdu
- Summarization
pretty_name: Urdu Summarization (BBC and DW Urdu News)
size_categories:
- 10K<n<100K
Urdu_DW-BBC-512
Dataset Description
- Homepage:
- Repository:
- Paper: https://doi.org/10.48550/arXiv.2310.02790
- Point of Contact: [email protected]
Dataset Summary
- Urdu Summarization Dataset containining 76,637 records of Article + Summary pairs scrapped from BBC Urdu and DW Urdu News Websites.
- Preprocessed Version: upto 512 tokens (~words); removed URLs, Pic Captions etc
Supported Tasks and Leaderboards
Summarization: Extractive and Abstractive
- urT5 adapted from mT5 having monolingual vocabulary only; 40k tokens of Urdu.
- Fine-tuned version @ https://huggingface.co/mbshr/urt5-base-finetuned, ref to https://doi.org/10.48550/arXiv.2310.02790 for details.
- ROUGE-1 F Score: 40.03 combined, 46.35 BBC Urdu datapoints only and 36.91 DW Urdu datapoints only)
- BERTScore: 75.1 combined, 77.0 BBC Urdu datapoints only and 74.16 DW Urdu datapoints only
Languages
Urdu.
Data Fields
- url: URL of the article from where it was scrapped (BBC Urdu URLs in english topic text with number & DW Urdu with Urdu topic text)
dtype: {string}
- Summary: Short Summary of article written by author of article like highlights.
dtype: {string}
- Text: Complete Text of article which are intelligently trucated to 512 tokens.
dtype: {string}