Scientific knowledge is only meaningful if people can access, understand, and make use of it. Open access has removed paywalls for many papers, but an OA paper may still be inaccessible to people using assistive reading technology, incomprehensible to non-specialists, or difficult to apply in practice. Emergent AI technologies create new opportunities to study the scope of these challenges, as well as introduce tooling to benchmark and improve access, for example by converting inaccessible PDFs to accessible formats, generating plain language content, allowing for personalized interactions, and more. But these transformations also introduce risks around information fidelity and the communication of uncertainty.
In this talk, I first discuss the problem of scientific PDF accessibility; studying accessibility failures at scale can reveal how access varies across venues, publishers, and publication models, including whether the rise of OA has actually been accompanied by more accessible documents. Some machine-learning-based solutions can help identify accessibility failures and support the conversion of PDFs into more accessible formats. Going beyond format, I then turn to plain language work, where AI can help transform complex scientific and medical evidence into more understandable forms for healthcare consumers, supporting better discoverability and comprehension. I connect this work to broader evaluation challenges for LLMs in scientific and consumer health information access. For medical text simplification and summarization, standard evaluation metrics can fail to capture uncertainty, and whether AI-generated transformations are faithful to the source and/or support users’ actual information needs.
While AI methods are promising for helping to quantify and address the accessibility challenges around scientific knowledge, the lack of faithful evaluation, the high prevalence of uncertainty, and high cost of errors in these domains should make us cautious. We need to support better access to scientific information, but also better evaluation methods that tell us when AI tools are most suitable and when they fail.
| 11:45am - 12:15pm: | Food and community socializing. |
| 12:15pm - 1:15pm: | Presentation with Q&A. Available hybrid via Zoom. |
| 1:30pm - 2:15pm: | Student meeting with speaker, held in the same location. |
Lucy Lu Wang is an Assistant Professor at the University of Washington Information School, where she leads the Language Accessibility Research (LARCH) lab. She holds adjunct appointments in the Paul G. Allen School of Computer Science & Engineering, Department of Biomedical Informatics & Medical Education, and Department of Human Centered Design & Engineering at the University of Washington, and is a Research Scientist at the Allen Institute for AI (Ai2). Her work spans scholarly document understanding, document accessibility, scientific evidence synthesis, and health communication. She focuses on developing language technologies to improve access to and understanding of information in high-expertise domains like science and healthcare, with an emphasis on dataset development and evaluation practices. Her work on supplement interaction detection, document accessibility, and academic publishing trends have been featured in media such as Geekwire, Boing Boing, Axios, VentureBeat, and the New York Times.