UC San Diego

In the era of generative artificial intelligence (GenAI), people increasingly rely on these AI-powered systems for daily tasks, ranging from conversational chatbots to healthcare assistance and beyond. Despite their widespread adoption, AI or natural language processing (NLP) models face critical challenges, such as their vulnerabilities to robustness issues across diverse groups, biased behaviors, and harmful outputs. These challenges raise significant concerns about their reliability and real-world applicability, emphasizing the urgent need for human-centered NLP -- systems designed to prioritize human values, trust, and socially beneficial outcomes.

This dissertation explores three core aspects of human-centered NLP. First, it addresses the trustworthiness of NLP systems, especially the large language models (LLMs), by examining critical concerns about their reliability and presenting strategies to enhance their robustness and trustworthiness. Second, it introduces a novel perspective on learning from humans, emphasizing the importance of understanding, modeling, and drawing inspiration from human cognition to align NLP systems more closely with human reasoning and behavior. Third, it highlights the impact of human-centered NLP in socially beneficial applications, such as improving patient care and outcomes in healthcare. By addressing critical challenges and integrating insights from interdisciplinary fields, this dissertation aims to pave a path toward NLP systems that not only perform effectively but also respect human values and advance social good, thereby laying the groundwork for the next generation of responsible, human-centered NLP technologies.