MultiBLiMP 1.0
收藏资源简介:
MultiBLiMP 1.0是由格罗宁根大学和德克萨斯大学奥斯汀分校的研究人员创建的一个大规模多语言语料库,包含101种语言,6种语言现象,超过125,000个最小对立体。该数据集通过利用Universal Dependencies和UniMorph的大规模语言资源,采用全自动管道创建。它不仅是一个评估基准,还是一个可以自动创建高度多语言基准的管道,可扩展到更多的句法现象。该数据集主要用于评估大型语言模型在多语言环境下的语言能力,并揭示了当前最先进技术在处理低资源语言方面的不足。
MultiBLiMP 1.0 is a large-scale multilingual corpus developed by researchers from the University of Groningen and The University of Texas at Austin, which covers 101 languages, 6 linguistic phenomena, and over 125,000 minimal pairs. This dataset was built through a fully automated pipeline that leverages large-scale linguistic resources from Universal Dependencies and UniMorph. It not only acts as an evaluation benchmark but also provides a pipeline for automatically generating highly multilingual benchmarks that can be expanded to cover more syntactic phenomena. This dataset is mainly used to assess the linguistic abilities of large language models in multilingual scenarios, and exposes the deficiencies of current state-of-the-art technologies in handling low-resource languages.




