UC San Diego

Artificial Intelligence (AI) systems that use language generation models hold incredible promise to assist humans to perform complex decision-making tasks. State-of-the-art language generation models can produce engaging content, reason about the world, and retrieve relevant information for an information-seeking task. However, these models often ignore sparse, long-tail knowledge about individual users, cultural subtleties, and domain-specific knowledge, preventing end users from reaping the full benefit of the scale. In this dissertation, we redesign AI systems to start with individual needs.

Ideally, a user-centric AI system must be grounded in the real-world, produce faithful chains of reasoning to explain its prediction, and align with the user's preferences. We elevate existing AI systems with knowledge, explanations, and interactions and develop both training-time and post-hoc techniques to make these systems user-centric.We show additional knowledge-grounding promotes user success in achieving conversational goals while using a conversational AI system. We demonstrate that AI explanations, when attributed to world knowledge, render them to be faithful and consistent. Finally, we discover that user-centric interventionist approach can help users obtain more equitable predictions backed by faithful explanations as compared to a black-box counterpart. In summary, our research establishes that increased effectiveness, explainability, and equitability can be achieved through knowledge-grounding and user-centric approaches to personalize AI models---a long-standing goal of artificial general intelligence.