Narrative Models for Interpretable LLM Analysis of Document Corpora
Abstract:
Large language models (LLMs) are powerful tools for corpus analysis, but they are opaque, hiding how they reason, and concealing potentially systematic interpretation biases. In this talk we will develop the concept of a narrative model, an interpretable and auditable framework in which LLMs generate structured question-answer data that statistical models analyze transparently. Each narrative is a collection of probability distributions over categorical answers, enabling direct semantic interpretation. We establish identifiability by extending anchor-word methods with reproducing kernel Hilbert space (RKHS) smoothness constraints, whose functional gradient descent mirrors Transformer attention for few-shot learning. The narrative model paradigm is demonstrated by considering LLM analysis of multiple real-world corpora.
Bio:
Lawrence Carin is the Director of AI for Health at A*STAR. Prior to that he spent nearly 30 years at Duke University, where he performed research in AI, and held two Distinguished Professor positions. He undertook leadership roles at Duke, including ECE Department Chair, and the Vice President for Research. He was also the Provost at KAUST. Further, he was the founder of two AI companies, both of which were subsequently acquired. He has published widely in the leading AI forums.