21-EuroParl
收藏资源简介:
21-EuroParl是一个多语言并行语料库,它基于欧洲议会(EP)的议会程序,包含了每位发言人的政治隶属关系信息。该数据集由1.5M个句子组成,总计4千万个单词和2.49亿个字符。它覆盖了从2009年到2011年的三年时间,来自27个国家的1000多名发言者,12个欧盟政党,25个欧盟委员会和数百个国内政党。数据集以句子级别进行对齐,并包含了丰富的元数据,如发言人的政治隶属关系(国家和欧盟政党)、日期和语言。这个数据集的创建过程涉及了对LinkedEP数据集的查询、清洗、句子对齐等步骤。21-EuroParl旨在用于研究多语言大型语言模型(LLMs)的政治公平性,特别是通过机器翻译来评估LLMs是否存在政治偏见。
21-EuroParl is a multilingual parallel corpus based on the parliamentary proceedings of the European Parliament (EP), which contains the political affiliation information of each speaker. The dataset consists of 1.5 million sentences, totaling 40 million words and 249 million characters. It spans a three-year period from 2009 to 2011, with over 1,000 speakers from 27 countries, 12 EU political parties, 25 EU committees, and hundreds of national political parties. The dataset is aligned at the sentence level and includes rich metadata, such as the speakers' political affiliations (national and EU political parties), dates, and languages. The creation of this dataset involves steps including querying, cleaning, and sentence alignment of the LinkedEP dataset. 21-EuroParl is designed for research on political fairness of multilingual large language models (LLMs), particularly to evaluate whether LLMs exhibit political biases via machine translation.




