One major challenge in information understanding stems from the diverse nature of the audience, where individuals possess varying preferences, experiences, educational and cultural backgrounds. Consequently, adopting a one-size-fits-all approach to provide information may prove suboptimal. While prior research has predominantly focused on delivering pre-existing content to users with potential interests, this thesis explores generative language models for personalized information understanding. By harnessing the potential of generative language models, our objective is to generate novel personalized content for individual users. As a result, users from diverse backgrounds can be provided with content that is tailored for their needs and better aligns with their interests. Our research will encompass two main domains: the general media domain and the patient education domain. We seek to apply these technologies in both sectors to enhance information accessibility and comprehension for a broader audience. In the general media domain, we explore personalized news headline generation. We present a novel framework that identifies perspectives that users are interested in from news passages based on users reading histories, which are then used to personalize news headline generation. Compared with presenting a fixed headline to all users, personalized headlines generated by our framework have the potential to improve the efficacy of news recommendations and facilitate creation of personalized content. We also explore personalized reading assistive technology to assist users understand complex information in news article or academic documents at ease, we propose a novel method for users to interactively acquire information, i.e. a user does not gain information through reading a passage but through conversations with an AI-powered chatbot who reads the passages. The chatbot actively leads the conversation and timely addresses the user's questions, making the information acquisition process more engaging and effective. In the patient education domain, we propose a novel after-visit summaries (AVS) writing assistant. After-visit summaries notes are documents given to patients to help them understand their clinical visits and disease self-management. However, writing AVS notes is both time and labor-consuming for physicians. Our approach not only automatically generates AVS drafts, but also detects potential errors in the generated drafts, allowing physicians to revise and produce AVS notes with higher efficiency and accuracy, thus better guide patients' personalized self-management. Moreover, we present PaniniQA, a patient-centric interactive question answering system designed to help patients understand their discharge instructions. PaniniQA first identifies important clinical content from patients' discharge instructions and then formulates personalized educational questions for distinctive patients. In addition, PaniniQA is also equipped with answer verification functionality to provide timely feedback to correct patients' misunderstandings. Overall, we aspire to contribute to the advancement of information dissemination techniques, promoting a more inclusive and effective means of communication in our information-driven world.
Advisor: Hong Yu