OpenLID-v3
收藏资源简介:
OpenLID-v3是由奥斯陆大学语言技术组开发的语言识别数据集,旨在解决网络数据中相近语言区分和噪声过滤的难题。该数据集覆盖194种语言及一个‘非语言’类别,通过整合GlotLID的开放许可数据和最新维基百科语料进行扩展,特别优化了拉丁语、塞尔维亚语拉丁字母变体等易混淆语言的训练样本。数据来源包括多语言网页文档、平行语料库及人工标注的噪声样本,主要应用于大规模多语言预训练数据清洗,可有效提升低资源语言数据的纯净度。
OpenLID-v3 is a language identification dataset developed by the Language Technology Group at the University of Oslo, designed to address the challenges of distinguishing closely related languages and filtering noise in web-scale data. This dataset covers 194 languages plus a "non-language" category, and is expanded by integrating openly licensed data from GlotLID and the latest Wikipedia corpora, with specially optimized training samples for easily confusable languages such as Latin and Latin-script variants of Serbian. Its data sources include multilingual web documents, parallel corpora, and manually annotated noise samples. It is primarily applied to large-scale multilingual pre-training data cleaning, and can effectively improve the purity of low-resource language data.



