aisingapore/SEA-NLI
收藏资源简介:
SEA-NLI 是一个基于文化的自然语言推理基准数据集,专门设计用于评估大型语言模型对东南亚多样文化细微差别的理解能力。它超越了通常依赖简单词汇重叠或翻译英语内容的标准NLI数据集,专注于围绕特定东南亚文化概念的前提-假设对。每个示例都提供英语和相应的区域本地语言版本,以测试跨语言推理能力。数据集覆盖八个国家及其对应语言:柬埔寨(高棉语)、缅甸(缅甸语)、马来西亚(马来语)、泰国(泰语)、新加坡(泰米尔语)、菲律宾(菲律宾语/他加禄语)、印度尼西亚(印度尼西亚语)和越南(越南语)。数据集包含两个主要配置:normal(标准文化基础示例,共1,443个例子)和hard(具有挑战性的示例,需要深入文化知识,共717个例子),总计2,160个示例。数据字段包括entry_id、premise_english、hypothesis_english、premise_native、hypothesis_native、true_label等,涉及文化概念标题、描述、类别等信息。
SEA-NLI is a culturally grounded Natural Language Inference (NLI) benchmark specifically designed to evaluate how well Large Language Models (LLMs) understand the diverse cultural nuances of Southeast Asia. It moves beyond standard NLI datasets that often rely on simple lexical overlap or translated English content, focusing on premise–hypothesis pairs centered on specific Southeast Asian cultural concepts. Each example is provided in both English and a corresponding regional native language to test cross-lingual reasoning. The dataset covers eight countries and their respective languages: Cambodia (Khmer), Myanmar (Burmese), Malaysia (Malay), Thailand (Thai), Singapore (Tamil), Philippines (Filipino/Tagalog), Indonesia (Indonesian), and Vietnam (Vietnamese). It includes two primary configurations: normal (standard culturally grounded NLI examples with filtered artifacts, 1,443 examples) and hard (challenging examples with minimal lexical overlap, requiring deep cultural knowledge, 717 examples), totalling 2,160 examples. Data fields include entry_id, premise_english, hypothesis_english, premise_native, hypothesis_native, true_label, concept_title, concept_description, concept_category, culture, and others.




