Fairness in Natural Language Processing

近年来，自然语言处理（NLP）在商业上取得了巨大的成功，并已部署在越来越广泛的设备和应用领域。与此同时，有明显的证据表明，经过天真训练的模型会放大训练数据中的偏差，并且在与不同人口统计学群体相关的文本中表现不一致。该项目旨在系统地量化这种偏见的程度，并开发出既更具社会公平性，又不容易在学习表示中暴露私人数据的模型。通过这样做，它将使新的用户群体更容易获得NLP，并消除NLP吸收的社会技术障碍。

Natural language processing (NLP) has achieved spectacular commercial successes in recent years, and has been deployed across an ever-increasing breadth of devices and application areas. At the same time, there has been stark evidence to indicate that naively-trained models amplify biases in training data, and perform inconsistently across text relating to different demographic groupings of individuals. This project aims to systematically quantify the extent of such biases, and develop models that are both more socially equitable, as well as less prone to expose private data in the learned representations. In doing so, it will make NLP more accessible to new populations of users, and remove socio-technological barriers to NLP uptake.