CAREER: Language Technologies Against the Language of Social Discrimination

该奖项全部或部分由2021年美国救援计划法案（公法117-2）资助。在线社交平台的指数增长为获取专家和群众智慧，寻找教育，就业和友谊提供了前所未有的平等机会。一个可能严重阻碍这些体验的关键根源是暴露于隐性社会偏见。这种风险很高，因为偏见是有害的，而且普遍存在，而且众所周知，语言是传播和延续陈规定型观念和偏见的主要手段。该项目开发语言技术，以检测和干预社会歧视的语言性别歧视，种族主义，同性恋恐惧症的microaggressions，屈尊俯就，客观化，非人性化的隐喻，等等，这可能是无意识和无意的，但会造成长期的个人和职业伤害。该计划开辟了新的研究机会，涉及自然语言处理，机器学习，数据科学和计算社会科学。它开发了新的网络规模算法，以自动检测隐含和伪装的毒性，以及在线仇恨言论和辱骂性语言。在技术上，它开发了新方法来揭示和降级深度学习模型中的虚假模式，以及解释深度学习模型的新技术，从而为可靠和可解释的机器学习开辟了新途径。该计划的成功完成将为现有方式的范式转变铺平道路，以监测网络空间的文明，保护弱势群体免受歧视和侵略，并减少平台版主的精神负担。因此，这个项目可以使大量的个人受益并赋予他们权力，这些人是受性别、种族、年龄、性取向、民族歧视的弱势群体的代表，他们使用基于用户生成内容的社交媒体或人工智能技术。最后，该计划开发的教育课程将为未来的技术人员提供构建道德人工智能的理论和实践工具，并将大大促进STEM教育的多样性，公平性和包容性，帮助培养新一代更多样化的研究人员进入人工智能领域。这个CAREER项目的总体目标是开发轻监督，可解释的机器学习方法-以社会心理学和因果推理为基础-以检测书面语篇和叙事文本中的隐性社会偏见。更具体地说，该项目的第一阶段开发算法和模型，用于识别和解释社交媒体上的简短评论中的性别微攻击，首先是无监督的，然后是主动学习，由训练有素的注释者进行有限的监督。它提供了变革性的解决方案，使现有的过度参数化的黑盒神经网络更强大，更可解释。由于微攻击往往是隐含的，它还开发了方法来产生解释微攻击检测器的决定。在第二阶段，该项目解决了在数字媒体的叙事领域中检测LGBTQ社区成员的偏见框架的挑战性任务，并通过跨语言操作化完善的社会心理学理论来开发数据分析工具。这个五年计划的预期成果包括新的数据集、算法和模型，这些数据集、算法和模型提供以人为本的文本分析，并查明和解释跨语言、数据域和社会背景的潜在偏见框架。该奖项反映了NSF的法定使命，并被认为值得通过使用基金会的智力价值和更广泛的影响审查标准进行评估来支持。

This award is funded in whole or in part under the American Rescue Plan Act of 2021 (Public Law 117-2).The exponential growth of online social platforms provides an unprecedented source of equal opportunities for accessing expert- and crowd-wisdom, for finding education, employment, and friendships. One key root cause that can deeply impede these experiences is the exposure to implicit social bias. The risk is high, since biases are pernicious and pervasive, and it has been well established that language is a primary means through which stereotypes and prejudice are communicated and perpetuated. This project develops language technologies to detect and intervene in the language of social discrimination—sexist, racist, homophobic microaggressions, condescension, objectification, dehumanizing metaphors, and the like—which can be unconscious and unintentional, but cause prolonged personal and professional harms. The program opens up new research opportunities with implications to natural language processing, machine learning, data science, and computational social science. It develops new Web-scale algorithms to automatically detect implicit and disguised toxicity, as well as hate speech and abusive language online. Technologically, it develops new methods to surface and demote spurious patterns in deep-learning models, and new techniques to interpret deep-learning models, thereby opening new avenues to reliable and interpretable machine learning. Successful completion of the program will pave the ground for a paradigm shift in existing ways for monitoring civility in cyberspace, shielding vulnerable populations from discrimination and aggression, and reducing the mental load of platform moderators. Therefore, this project can benefit and empower a dramatic number of individuals—representatives of disadvantaged groups discriminated by gender, race, age, sexual orientation, ethnicity—who use social media or AI technologies built upon user generated content. Finally, the educational curriculum developed by this program will equip future technologists with theoretical and practical tools for building ethical AI, and will substantially promote diversity, equity and inclusion in STEM education, helping to foster a new, more diverse generation of researchers entering AI. The overarching goal of this CAREER project is to develop lightly supervised, interpretable machine learning approaches—grounded in social psychology and causal reasoning—to detect implicit social bias in written discourse and narrative text. More specifically, the first phase of the project develops algorithms and models for identifying and explaining gendered microaggressions in short comments on social media, first unsupervisedly, then with active learning, given limited supervision by trained annotators. It provides transformative solutions to making existing overparameterized black-box neural networks more robust and more interpretable. Since microaggressions are often implicit, it also develops approaches to generate explanations to the microaggression detector’s decisions. In the second phase, the project addresses the challenging task of detecting biased framing about members of the LGBTQ community in narrative domains of digital media and develops data analytic tools by operationalizing, across languages, well-established social psychology theories. The expected outcomes of this five-year program include new datasets, algorithms, and models that provide people-centered text analytics, and pinpoint and explain potentially biased framings, across languages, data domains, and social contexts.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.