KoBBQ datasets
收藏资源简介:
The Bias Benchmark for Question Answering (BBQ) is designed to evaluate social biases of language models (LMs), but it is not simple to adapt this benchmark to cultural contexts other than the US because social biases depend heavily on the cultural context. In this paper, we present ₩KoBBQ₩, a Korean bias benchmark dataset, and we propose a general framework that addresses considerations for cultural adaptation of a dataset. Our framework includes partitioning the BBQ dataset into three classes--Simply-Transferred (can be used directly after cultural translation), Target-Modified (requires localization in target groups), and Sample-Removed (does not fit Korean culture)-- and adding four new categories of bias specific to Korean culture. We conduct a large-scale survey to collect and validate the social biases and the targets of the biases that reflect the stereotypes in Korean culture. The resulting KoBBQ dataset comprises 268 templates and 76,048 samples across 12 categories of social bias. We use KoBBQ to measure the accuracy and bias scores of several state-of-the-art multilingual LMs. The results clearly show differences in the bias of LMs as measured by KoBBQ and a machine-translated version of BBQ, demonstrating the need for and utility of a well-constructed, culturally-aware social bias benchmark.
问答偏见基准数据集(Bias Benchmark for Question Answering,BBQ)旨在评测语言模型(Language Models,LMs)的社会偏见,但由于社会偏见高度依赖文化语境,将该基准适配至美国以外的文化场景并非易事。本文提出了KoBBQ——一款韩语偏见基准数据集,同时构建了一套通用框架,可解决数据集文化适配过程中的各类核心考量问题。该框架将BBQ数据集划分为三类:简单迁移类(仅需完成文化翻译即可直接使用)、目标修改类(需针对目标群体进行本土化调整)与样本移除类(不符合韩语文化语境,需予以剔除),并新增了四类针对韩语文化的特有偏见类别。我们通过大规模调研收集并验证了反映韩语文化刻板印象的社会偏见及其靶向对象。最终构建的KoBBQ数据集共包含268条模板与76048条样本,覆盖12类社会偏见场景。我们利用KoBBQ对多款顶尖多语言大语言模型的准确率与偏见得分进行了评测。实验结果清晰表明,基于KoBBQ与机器翻译版BBQ评测得到的大语言模型偏见表现存在显著差异,这充分证明了构建具备文化适配性的高质量社会偏见基准数据集的必要性与实用价值。
数据集概述
数据集名称
- KoBBQ: Korean Bias Benchmark for Question Answering
数据集位置
数据集统计
| Category | # of Templates | # of Samples |
|---|---|---|
| Age | 21 | 3,608 |
| Disability Status | 20 | 2,160 |
| Gender Identity | 25 | 768 |
| Physical Appearance | 20 | 4,040 |
| Race/Ethnicity/Nationality | 43 | 51,856 |
| Religion | 20 | 688 |
| Socio-Economic Status | 27 | 6,928 |
| Sexual Orientation | 12 | 552 |
| Domestic Area of Origin | 22 | 800 |
| Family Structure | 23 | 1,096 |
| Political Orientation | 11 | 312 |
| Education Background | 24 | 3,240 |
| Total | 268 | 76,048 |
评估方法
- 评估代码和提示可在KoBBQ/evaluation找到。
- 评估步骤包括将模型输出放入特定列,并运行评估脚本。
伦理考虑
- 数据集不得用于恶意目的,如自动生成和发布针对特定群体的偏见语言。应鼓励用于减少语言模型偏见的研究。




