Existing measures of human rights abuses are often only available at the country-year level. Several more fine-grained measures exhibit spatio-temporal inaccuracies or reporting biases due to the primary sources upon which they rely. To address these challenges, and to increase the diversity of available human rights measures more generally, this study provides the first quantitative effort to measure human rights abuses from textual records of citizen-government interactions. Using a dataset encompassing over 1.5 million access-to-information (ATI) requests made to the Mexican federal government from June 2003 onward, supervised classification is used to identify the subset of these requests that pertain to human rights abuses of various types. The results from this supervised machine learning exercise are validated against (i) gold standard ATI requests pertaining to past human rights abuses in Mexico and (ii) several accepted external measures of sub-national and sub-annual human rights abuses. In doing so, we demonstrate that the measurement of human rights abuses from citizen-submitted ATI request texts can provide measures of human rights abuse that exhibit both high validity and notable spatio-temporal specificity, relative to existent human rights datasets and variables.

现有的侵犯人权措施通常仅适用于国家年度级别。由于它们依赖的主要来源，几个更细粒度的度量表现出时空不准确或报告偏差。为了应对这些挑战，并更广泛地增加可用人权措施的多样性，本研究首次提供了从公民与政府互动的文本记录中衡量侵犯人权行为的定量工作。使用包含从 2003 年 6 月起向墨西哥联邦政府提出的超过 150 万个信息获取 (ATI) 请求的数据集，使用监督分类来识别这些请求的子集，这些请求与各种类型的侵犯人权行为有关。此受监督的机器学习练习的结果已根据 (i) 与墨西哥过去侵犯人权行为有关的黄金标准 ATI 请求和 (ii) 若干公认的次国家和次年度侵犯人权行为的外部措施进行验证。在此过程中，我们证明了从公民提交的 ATI 请求文本中衡量侵犯人权行为可以提供具有高有效性的人权侵犯衡量标准和显着的时空特异性，相对于现有的人权数据集和变量。