WinoBias

WinoBias 包含 3,160 个句子，由熟悉该项目的研究人员创建，平均分配用于开发和测试。创建句子以遵循两个原型模板，但鼓励注释者提出实体可以以合理方式交互的场景。选择的模板具有挑战性，旨在涵盖分别需要语义和语法的情况。

WinoBias contains 3,160 sentences created by researchers familiar with the project, which are evenly split between development and test sets. The sentences were designed to follow two prototype templates, while annotators were encouraged to develop scenarios where entities could interact in plausible ways. The selected templates are challenging and aim to cover cases that require semantic and syntactic reasoning respectively.

搜集汇总

背景与挑战

背景概述

WinoBias是一个用于评估共指消解中性别偏见的数据集，包含3,160个句子，平均分配用于开发和测试。数据集由研究人员创建，遵循特定模板，涵盖需要语义和语法分析的场景。