Machine Learning Seminar Series Fall 2026 | Measuring Cultural Variation: from Twitter to LLMs

Abstract: People's language use reflects their culture. As a result, LLMs should adapt to the cultures of their many users, but they struggle to do so. Today's LLMs are multilingual but not multicultural: they speak fluent Hindi or Mandarin yet reflect the (often stereotypical) American perspectives implicit in their training data.  We first demonstrate how language can be used to quantify cultural variation, analyzing Twitter texts to quantify variation in individualism and collectivism across US counties. We then present a benchmark to assess how well LLMs understand cultural variation—a set of questions that should be answered differently across countries—and demonstrate that LLMs struggle to respond appropriately. Achieving truly global generative AI requires not just linguistic diversity, but LLMs that adapt to their multicultural users.

Bio: Lyle Ungar is a Professor of Computer and Information Science at the University of Pennsylvania, where he holds secondary appointments in Psychology, Bioengineering, Genomics and Computational Biology, and Operations, Information and Decisions. His group builds chatbots and language analysis tools for psychological and medical research.  They study the drivers of physical and mental well-being and their cultural variation, and develop socially sensitive AI-based tutors and coaches.

For CODA guest access, please contact shatcher8@gatech.eduat least 2 business days prior to the event.

Zoom link: https://gatech.zoom.us/j/91607375625?pwd=ybCsFPzknA8VqYZaSqvenqlWbrvyzx.1

Meeting ID: 916 0737 5625
Passcode: 038828