CAREER: Learning Multi-Level Narrative Structure

Why do certain stories, but not others, resonate so powerfully with certain populations? Stories (a.k.a. narratives) are powerful: where rational argument fails, a single story can drive home a point, change a mind, and even change a life. What specific structures underlie the power of narrative, and what new artificial intelligence (AI) techniques are needed to learn these structures automatically so we can leverage them in applications? This project seeks to develop these new AI techniques to automatically uncover and confirm the fundamental structures underlying narrative, developing and testing with data drawn from the domains of education and culture. This work is of broad relevance to developing more intelligent machines, understanding the mind and brain, and improving education. It will produce fundamental insights into a universal form of communication (narrative), providing a potentially transformative new set of tools to researchers and educators. The project will develop new machine learning and natural language processing approaches to learning key aspects of narrative structure. The basic structure of a narrative involves the plot, a time-ordered sequence of important events, and the plot can be divided into three levels of structure: (1) plot pieces, (2) stereotypical characters, and (3) narrative arcs. The PI and his students will first learn to extract these three types of narrative structure, the third of which (narrative arcs) is as-yet untried, using novel combinations of existing grammar learning approaches and Bayesian approaches, specifically the PI’s Analogical Story Merging (ASM) algorithm, the Infinite Relational Model (IRM), and iterative learning. Second, the researchers will test hypotheses that reflect why specific stories are persuasive to specific cultures and apply these insights to improving minority engagement in STEM and computing in middle-school classrooms in Miami Dade County Public Schools. Third, the researchers will seek to uncover systematic regularities in professional education cases (such as business cases, or medical case reports) that will lead to the ability to make computational predictions as to which cases should be most effective in the classroom.

Dates Active: June 1, 2018 — May 31, 2023