Counter Factual Data Augmentation for Mitigating Gender Bias in Morphologically Rich Languages

现代机器学习系统面临的最大挑战之一是消除数据中的偏差。随着人工智能算法越来越多地被用于塑造和决定私人和公共生活中的重要决策，确保这些系统没有人类偏见和错误是至关重要的。尽管如此，这些算法的实用性和准确性取决于“不客观的数据；它们是人类设计的产物”(Crawford，2013)。数据存在固有偏见的一个领域是语言。我目前的硕士论文旨在创建一种新颖的生成模式，将一种性别形式的句子转换为另一种性别形式，用于西班牙语和希伯来语等具有名词、动词和形容词的男性和女性屈折变化的语言。博士将在这项研究的基础上研究三个主要问题：(1)用于平衡性别语言的生成模式的有效性和效率可以在多大程度上得到加强和提高？(2)还可以设计和实施哪些其他技术来消除英语和其他语言的语料库中的偏见？以及(3)，这些模型和技术可以在多大程度上应用于研究和纠正其他类型的偏见，如种族和文化偏见？博士将致力于建立避免偏见放大的系统，促进跨学科应用，并为为NLP任务开发的未来模型提供健壮性。

One of the biggest challenges facing modern machine learning systems is removing bias from data. As AI algorithms areincreasingly being used to shape and determine important decisions in private and public life, it is essential to ensure thatthese systems are devoid of human prejudice and error. Nonetheless, the utility and accuracy of these algorithmsdepend on data which "are not objective; they are creations of human design" (Crawford, 2013). One area in which datais inherently biased is language. My current Master's thesis aims to create a novel generative model that transformssentences of one gender form into another gender form, for languages such as Spanish and Hebrew which possessmasculine and feminine inflections for nouns, verbs, and adjectives. The PhD will seek to build on this research andinvestigate three main questions: (1) To what extent can the efficacy and efficiency of the generative model used tobalance gendered language be strengthened and enhanced? (2) What other techniques can be designed andimplemented to de-bias corpora in English and other languages? And (3), to what extent can these models andtechniques be applied to study and correct other types of biases, such as racial and cultural biases? The PhD will aim tobuild systems that will avoid bias amplification, facilitate interdisciplinary applications, and offer robustness to futuremodels developed for NLP tasks.