MAP (Maybe Ambiguous Pronoun)

正确解决文本提及的人从根本上需要对这些人进行推断。这种推论增加了共指解决系统中系统性偏见的风险，包括可能损害二元和非二元反式和顺式利益相关者的偏见。为了更好地理解这种偏见，我们从社会学和社会语言学中突出性别的细微差别概念，并开发了两个新的数据集，用于在人群注释和现有的共指解决系统中询问偏见。通过这些对英文文本进行的研究，我们确认，如果不承认和构建识别性别复杂性的系统，我们构建的系统会导致许多潜在的危害。

Correct coreference resolution for persons mentioned in text fundamentally requires inferential reasoning about these individuals. This inferential reasoning introduces risks of systemic bias in coreference resolution systems, including biases that may harm binary and non-binary trans and cis stakeholders. To better understand such biases, we highlight the nuanced concept of gender drawn from sociology and sociolinguistics, and develop two novel datasets for probing bias in both human annotation and existing coreference resolution systems. Through these studies conducted on English text, we confirm that without systems that recognize and account for the complexity of gender identity, the systems we build will result in numerous potential harms.

搜集汇总

背景与挑战

背景概述

MAP数据集旨在探究共指消解系统中的性别偏见问题，以促进构建更具包容性的自然语言处理系统。该数据集由马里兰大学于2020年发布，基于相关学术研究论文。