WALS-Bench
收藏资源简介:
WALS-bench是一个基于世界语言结构图谱(WALS)的大规模多语言基准测试数据集,旨在评估大型语言模型中的元语言知识。该数据集覆盖了192个语言类型学特征,涉及2,660种语言。数据集提供了两种格式:格式1包含192个问题(每个特征一个问题),格式2包含76,475个问题(每个特征-语言组合一个问题)。每个问题都包含一个语言特征描述、可能的答案选项以及特定语言的真实答案。数据集分为训练集(134个特征)、验证集(29个特征)和测试集(29个特征)。数据以JSONL格式存储,包含特征ID、特征名称、领域、问题文本、可能答案和真实答案等信息。语言特征覆盖了词序、名词类别、简单从句、音系学、动词类别、词汇、形态学、名词句法、复杂句子、手语等多个领域。数据集采用CC BY 4.0许可协议,原始数据来自Max Planck进化人类学研究所的世界语言结构图谱。
WALS-bench is a large-scale multilingual benchmark dataset based on the World Atlas of Language Structures (WALS), designed to evaluate meta-linguistic knowledge in Large Language Models (LLMs). This dataset covers 192 language typological features and involves 2,660 languages. The dataset offers two formats: Format 1 contains 192 questions (one question per feature), while Format 2 contains 76,475 questions (one question per feature-language pair). Each question includes a description of the linguistic feature, a set of possible answer options, and the ground-truth answer for a specific language. The dataset is split into training set (134 features), validation set (29 features), and test set (29 features). The data is stored in JSONL format, containing information such as feature ID, feature name, domain, question text, possible answers, and ground-truth answer. The linguistic features cover multiple domains including word order, noun classes, simple clauses, phonology, verb classes, lexicon, morphology, nominal syntax, complex sentences, sign languages, and more. The dataset is released under the CC BY 4.0 license, and its original data is sourced from the World Atlas of Language Structures at the Max Planck Institute for Evolutionary Anthropology.




