CAREER: Trustworthy Human-Centered Summarization

The rapid advancement of generative artificial intelligence and large language models is transforming how people process data, gather information, and acquire knowledge. Especially, in the era of information explosion, people are increasingly resorting to large language models to summarize documents to capture important information and make decisions. However, despite the prevalent usage of large language models, they have been criticized for inadequate alignments with humans by ignoring user needs of styles and contents, displaying bias by neglecting opinions from certain groups, and making factual mistakes on human knowledge and truth. The goal of this project is to advance trustworthy summarization by centering design, development, and deployment on humans in terms of user preferences for controllability, social perspectives for fairness, and human knowledge for factuality. It enlightens the research imperative of integrating human factors into large language models for trustworthy artificial intelligence and diverse societal impacts in domains such as scientific discovery, legal democracy, and public health. This project also initiates several aspiring education and outreach activities supported by project research outcomes to involve, mentor, and empower female, underrepresented, disabled, and interdisciplinary students. This research advances state-of-the-art controllable, fair, and factual summarization for efficient information gathering and knowledge acquisition by establishing a comprehensive set of infrastructure including algorithmic foundations, technical innovations, public benchmarks, and integrated platforms. The research plan consists of four thrusts to establish rigorous definitions of new tasks, reliable metrics, and novel models. First, the project incorporates fine-grained user preferences by customizing summaries based on their compositional requirements of contents and styles. Second, the research embraces voices from different social groups by generating fair and unbiased summaries to comprehensively cover diverse perspectives and conflicting opinions. Then, the research honors human knowledge by summarizing documents with multimodal information including unstructured text, structured knowledge of tables and citations, and visuals of figures and plots. Finally, the research deploys an integrated, interactive, and visualizable summarization platform to assist users with efficient and accessible document understanding. Through these innovations for trustworthy human-centered summarization, the research thrusts collectively advance usable, trustworthy, responsible, and safe artificial intelligence that operates under user control, reflects social norms, and honors human knowledge. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.