Towards Globally Equitable Language Technologies (EQUATE)

语言技术现在可以为沟通、教育、医疗保健和人类生活的许多其他方面提供有效的支持。然而，这些技术并不是均匀分布的。它们只适用于世界79亿人口中的一小部分，主要是那些生活在全球北部的人。这是因为它们所需的资源对世界上7000多种活着的语言中的绝大多数来说是有限的或缺乏的。这种情况具有重大的科学和社会经济后果。我们项目的目标是调查在开发全球公平的语言技术方面的方法挑战，并设计克服这些挑战的变革性方法，总的目的是为多边公平的自然语言规划创造一个现实的方法基础。我们将首先发展对语言技术中的平等的理解，并制作一个新的指数，根据世界语言和语言人口对语言技术的准备情况来描述他们。然后，我们将为多语言NLP开发新的方法，以解决公平性的关键方面(从样本效率到模块化、模型紧凑性、透明度、公平性等)，以及一种新的统一方法，该方法集成了这些方法，以支持不同准备程度的NLP。我们还将与当地语言群体合作，编制新的、具有公平意识的评估资源，在地理区域、语言特征和NLP准备情况方面代表世界上资源较少的语言。我们的新方法将在下游的NLP任务上进行评估，并在它们目前提供的服务不足的语言的有用的现实生活应用程序的背景下进行评估。该项目可以改变我们处理多语言NLP的方式，并极大地提高我们对如何使语言技术在全球范围内公平和包容的理解。

Language technologies can now offer effective support to communication, education, healthcare, and many other aspects of human life. Yet, these technologies are not distributed equally. They are only available for a small part of the world's 7.9 billion population, mainly those living in the Global North. This is because the resources needed for them are limited or lacking for the vast majority of the world's over 7,000 living languages. This situation has significant scientific and socioeconomic consequences. The goal of our project is to investigate the methodological challenges in the development of globally equitable language technologies and to design transformative approaches to overcome them, with the overall aim of creating a realistic methodological basis for multilingually equitable NLP. We will first develop an understanding of the (in)equalities in language technologies, and produce a novel index that profiles the world's languages and language populations in terms of their readiness for language technologies. We will then develop new methods for multilingual NLP that address critical aspects of equity (ranging from sample efficiency to modularity, model compactness, transparency, fairness and others), along with a novel unified approach that integrates such methods to support NLP at different levels of readiness. Working with local language populations, we will also produce new, equity-aware evaluation resources that are representative of the world's low-resource languages in terms of geographic regions, linguistic characteristics, and NLP readiness. Our novel methodology will be evaluated on downstream NLP tasks as well as in the context of useful real-life applications in languages that are currently under-served by them. This project can transform the way we approach multilingual NLP, and substantially improve our understanding of how language technologies can be made fair and inclusive at a global level.