RI: Small: DaRE: Detection and Recognition of Euphemisms

为了完全理解人类语言，机器需要能够识别和解释包含隐藏含义的表达。这个项目集中在委婉语，温和或间接的短语用来代替严厉或更冒犯的。委婉语通常用来掩饰亵渎，或礼貌地提及诸如死亡、性、宗教、残疾或个人关系等敏感话题。人们一直在使用委婉语，例如，“病人的消极结果”、“工作之间”、“经济上幸运”、“惩教设施”、“友军火力”或“阳光单位”。不同的文化/语言使用不同的委婉语。委婉语会随着时间而变化。处理人类语言的机器还不能理解委婉语。这个项目致力于让机器理解不同语言的委婉语，从而为提高人工智能的能力做出贡献。其他的好处还包括对委婉语的本质进行了有趣的新概括，并在一个困难的跨学科问题上训练了一支由本科生和研究生组成的多元化骨干队伍，他们在高度实际的工作中发挥了作用。蒙特克莱尔州立大学是一所西班牙裔服务机构，以其多样化的学生群体和很大比例的第一代大学生而闻名。蒙特克莱尔州立大学非常重视学术界的公正和包容性。这个项目也不例外。比喻语言的检测和解释是自然语言处理（NLP）中一个快速发展的领域。遗憾的是，目前在自然语言处理中还缺乏对委婉语的处理。该项目主要解决以下问题：1)委婉语检测和解释的算法设计；2)通过创建一系列新的数据集和任务，探索委婉语识别转换语言模型的嵌入空间，实现黑盒神经模型的可解释性。关键的见解是：1)委婉的表达方式及其释义对应的表达方式在表达情感的强度上有所不同；2)委婉语和非委婉语的解释具有语境敏感性；委婉语比它们所代替的禁忌语更模糊。这些实验测试了深度学习方法捕捉到委婉语的哪些语言特性以及原因。开发的算法可以在没有人为干预的情况下检测到以前没有记录在字典中的新委婉语。委婉语的计算工作对于进一步理解语言的策略性使用如何使人们对重要和极具争议的行为的看法产生偏见，并可能找到消除语言模型偏见的方法非常重要。这项关于委婉语的研究有助于理解在特定文化中哪些话题是有争议的或敏感的。将该算法应用于历时性数据，检测委婉语使用的变化，可以更好地理解文化的变化。所产生的语料库对于回答人工智能、自然语言处理、语言学、文化人类学和社会心理学交叉领域的问题非常有用。语言的范围提供了一种自然的方式来对委婉语进行有趣的语言观察。由于委婉语是语言行为的一种形式，找到一种自动检测和解释委婉语的方法可能会使我们更好地理解人类的一般行为。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

To fully understand human language, machines need to be able to recognize and interpret expressions that contain hidden meanings. This project concentrates on euphemisms, mild or indirect phrases used in place of harsher or more offensive ones. Euphemisms are often used to mask profanity or refer to sensitive topics such as death, sex, religion, disability, or personal relationships in a polite way. People use euphemisms all the time, e.g., 'negative patient outcome', 'between jobs', 'financially fortunate', 'correctional facility','friendly fire', or 'sunshine unit'. Different cultures/languages use different euphemisms. Euphemisms change over time. Machines that process human language do not understand euphemisms yet. This project is devoted to making machines understand euphemisms in different languages, and therefore contributing to improving the capabilities of artificial intelligence. Additional benefits include interesting new generalizations about the nature of euphemisms and the training of a diverse cadre of undergraduate and graduate students in highly practical work on a difficult interdisciplinary problem. Montclair State University, a Hispanic Serving Institution, is known for its diverse student population and a large proportion of first-generation college students. Montclair State University puts great emphasis on justice and inclusivity in academia. This project is not an exception.Detecting and interpreting figurative language is a rapidly growing area in Natural Language Processing (NLP). Unfortunately, the processing of euphemisms is lacking in NLP thus far. The project addresses the following: 1) algorithm design for detecting and interpreting euphemisms, and 2) interpretability of black-box neural models by creating a series of new datasets and tasks that explore the embedding space of transformer language models for euphemism recognition. The key insights are 1) euphemistic expressions and their paraphrased counterparts differ in the strength of the sentiment they convey; 2) euphemistic and non-euphemistic interpretation is context-sensitive; 3) euphemisms are vaguer than the taboo expressions they substitute. The experiments test what linguistic properties of euphemisms the deep learning approaches capture and why. The algorithm developed can detect new euphemisms, not previously recorded in dictionaries, without human intervention. The computational work on euphemisms is important to further the understanding of how strategic use of language can bias people's perceptions of important and highly contentious actions and perhaps find ways how to de-bias language models. This work on euphemisms helps understand what topics are controversial or sensitive in a specific culture. Applying the algorithm to diachronic data and detecting the change in euphemism usage leads to a better understanding of culture changes. The corpora produced are useful for answering questions at the intersection of AI, NLP, linguistics, cultural anthropology, and social psychology. The range of languages provides a natural way of making interesting linguistic observations about euphemisms. Since euphemisms are a form of verbal behavior, finding a way to detect and interpret euphemisms automatically may lead to a better understanding of human behavior in general.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.