Towards Understanding and Practices of Ethical Artificial Intelligence

The successes of machine learning (ML) and artificial intelligence (AI) models encourage their widespread deployments in high-stakes domains -- from public transportation to social decision-making such as autonomous driving, criminal justice, and company hiring. Such widespread deployments call for assessing and addressing the ethical concerns of AI systems.

The thesis aims to develop practical techniques and theoretical understanding for building ethical AI systems. We divide the thesis into two parts. The first part of the thesis focuses on automatic information extraction using natural language processing (NLP) from policy documents. Policy documents are natural language documents about how different stakeholders (e.g., users and ML services providers) in internet services agree on how the services providers commit to the ethical usage of users' data. Specifically, we develop NLP techniques and benchmarks for privacy policies, a type of policy document describing the practices of using, sharing, and protecting users' data. Such developed NLP techniques could be extended to other natural language law documents describing ethical AI principles and help improve mutual trust among different parties.

The second part of the thesis focuses on the theoretical understanding and development of algorithmic interventions for ethical artificial intelligence. In particular, we study the fairness problems for various machine learning tasks, such as classification, regression, and sequential decision-making: (1) we provide bias mitigation techniques for text classification using contrastive representation learning; (2) we provide the theoretical understanding and mitigation techniques for accuracy disparity problem in regression; (3) we propose a fairness notion that requires long-term equality on expected utility for different demographic groups for sequential decision-making and develop methods to achieve the proposed fairness notion. In addition, we also study adversarial representation learning, a technique that has been widely used for algorithmic fairness, and its implications for information obfuscation.

We hope the research presented in the thesis will facilitate the practices of building ethical machine learning systems and help increase the understanding and trust among stakeholders towards the machine learning systems.