GLOBESUMM
收藏资源简介:
GLOBESUMM数据集由哈尔滨工业大学等机构创建,旨在解决多语言、跨语言和多文档新闻摘要的挑战。该数据集包含4687条新闻报道,涵盖26种语言,主要来源于GDELT数据库。数据集的创建过程包括大规模数据收集、事件中心化重排和过滤,以及基于协议引导提示的高质量银标摘要注释。GLOBESUMM数据集的应用领域主要集中在多语言新闻摘要,旨在帮助研究人员更好地理解和处理全球新闻事件中的多语言和跨文化差异。
The GLOBESUMM dataset was developed by Harbin Institute of Technology and other institutions, with the goal of addressing the challenges inherent in multilingual, cross-lingual, and multi-document news summarization. This dataset comprises 4,687 news reports spanning 26 languages, and is primarily sourced from the GDELT database. The construction process of the dataset involves large-scale data collection, event-centric rearrangement and filtering, as well as the annotation of high-quality silver-standard summaries using protocol-guided prompts. The primary application scope of the GLOBESUMM dataset lies in multilingual news summarization, with the objective of enabling researchers to better understand and address multilingual and cross-cultural differences in global news events.
GlobeSumm 数据集概述
数据集简介
GlobeSumm 是一个面向多语言、跨语言和多文档新闻摘要的挑战性基准数据集,旨在通过多语言新闻文章提供对全球事件的全面理解。
数据集内容
1. 事件中心的多语言新闻报道
- 数量: 370个新闻事件,包含4687篇新闻文章。
- 语言: 涵盖26种语言。
- 特点: 每个新闻事件至少有10篇不同语言的新闻报道。
2. 新闻相关性判定数据
- 数量: 2104个事件,包含26301篇新闻文章。
- 特点: 包含手动验证的相关性标签,用于判定新闻报道与给定描述的相关性。
3. 全球新闻摘要
- 摘要方法: 采用时间顺序递归摘要(CRS)方法,结合关键信息分割(KIS)、跨语言提示(CLP)和协议引导提示(PGP)技术。
- 特点: 生成高质量的摘要,识别并处理冗余、遗漏和冲突问题。
数据结构示例
python { # Event1 "date": 20230501, "description": "a Cessna 206 light aircraft with seven people on board crashed in the jungle in the Caquetá Department of Colombia.", "category": "Disasters and accidents", "news": [ { # news1 "lang_abbr": "fr", "lang_full": "French", "date": "20230517", "title": "XXX", "article": "XXX", "label_relevant": 1 }, # news2 # news3 ... ] }
引用
@misc{ye2024globesummchallengingbenchmarkunifying, title={GlobeSumm: A Challenging Benchmark Towards Unifying Multi-lingual, Cross-lingual and Multi-document News Summarization}, author={Yangfan Ye and Xiachong Feng and Xiaocheng Feng and Weitao Ma and Libo Qin and Dongliang Xu and Qing Yang and Hongtao Liu and Bing Qin}, year={2024}, eprint={2410.04087}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2410.04087}, }




