proxectonos/GlBBQ
收藏资源简介:
GlBBQ是BBQ(偏见基准问答)的加利西亚语改编版,旨在测量多选问答系统中的社会偏见。具体来说,它源自EsBBQ(西班牙语偏见基准问答),并遵循BBQ框架,在模糊上下文(正确答案未知,模型应避免依赖刻板印象)和去歧义上下文(正确答案明确支持,偏见可能导致错误预测)下评估模型。数据集扩展至加利西亚语,提供EsBBQ的语言适应版本。它包括27,320个实例、323个模板和10个社会偏见类别(如年龄、残疾状况、性别、LGBTQIA、国籍、外貌、种族/民族、宗教、社会经济地位和西班牙地区)。每个实例包含一个上下文、一个问题及三个答案选项,任务是根据上下文选择正确答案。数据集结构支持评估偏见在不确定性下的表现和偏见敏感条件下的准确性。用途包括评估语言模型的社会偏见、研究跨语言偏见转移、基准测试加利西亚语NLP系统以及多语言问答公平性研究。局限性包括专注于刻板印象偏见、继承EsBBQ的西班牙特定社会文化假设、未明确建模交叉性,以及翻译可能引入轻微语言伪影。
GlBBQ is a Galician adaptation of the BBQ (Bias Benchmark for QA), a benchmark designed to measure social bias in multiple-choice question answering (QA) systems. More specifically, GlBBQ is derived from EsBBQ, the Spanish adaptation of BBQ. The dataset follows the BBQ framework, where models are evaluated under ambiguous contexts (where the correct answer is unknown and models should avoid relying on stereotypes) and disambiguated contexts (where the correct answer is explicitly supported and bias may lead to incorrect predictions). GlBBQ extends this framework to Galician by providing a linguistically adapted version of EsBBQ. It includes 27,320 instances, 323 templates, and 10 social bias categories (such as Age, Disability Status, Gender, LGBTQIA, Nationality, Physical Appearance, Race/Ethnicity, Religion, Socioeconomic Status, and Spanish Region). Each instance consists of a context, a question, and three answer options, with the task being to select the correct answer given the context. The dataset structure enables evaluation of both bias under uncertainty and accuracy under bias-sensitive conditions. Intended uses include evaluating social bias in language models, studying cross-lingual bias transfer, benchmarking Galician NLP systems, and research on fairness in multilingual QA. Limitations include a focus on stereotype-based bias, inheritance of EsBBQs Spain-specific sociocultural assumptions, lack of explicit modeling of intersectionality, and potential minor linguistic artifacts from translation.




