KoBALT (Korean Benchmark for Advanced Linguistic Tasks)
收藏资源简介:
KoBALT是一个包含700个多项选择题的全面的语言学基准测试,涵盖了五个语言学领域:句法、语义、语用、语音/音位和形态学。该数据集旨在通过引入专家策划的、具有语言学动机的问题来推进大型语言模型在韩语中的评估,这些问题与标准韩语语料库的最小n-gram重叠最小,从而显著降低了数据污染的风险,并允许对真正的语言理解进行更稳健的评估。通过评估20个当代大型语言模型,揭示了它们在各个语言学领域中的性能差异。此外,通过95位标注者的偏好评估,展示了KoBALT评分与人类判断之间的强相关性,从而验证了该基准测试作为衡量韩语语言理解的有效性和区分性。
KoBALT is a comprehensive linguistic benchmark consisting of 700 multiple-choice questions spanning five core linguistic domains: syntax, semantics, pragmatics, phonetics/phonology, and morphology. This benchmark aims to advance the evaluation of large language models (LLMs) for Korean by introducing expert-curated, linguistically motivated questions that exhibit minimal n-gram overlap with standard Korean corpora, thereby significantly reducing the risk of data contamination and enabling more robust assessment of genuine language understanding. An evaluation of 20 contemporary large language models using KoBALT reveals performance disparities across the five linguistic domains. Furthermore, a preference evaluation involving 95 annotators demonstrates a strong correlation between KoBALT scores and human judgments, validating the benchmark’s effectiveness and discriminative power as a metric for Korean language understanding.

- 1KoBALT: Korean Benchmark For Advanced Linguistic Tasks首尔国立大学, LG AI Research · 2025年



