Challenges in sarcasm handling by language models (and humans)

• Event: Seminar
• Presented by: Hyewon Jang from the University of Konstanz
• Date: 21 May 2025
• Time: 13:15-15:00
• Venue: Gothenburg University, Humanisten and online
• Address: Renströmsgatan 6, 412 55 Göteborg
• Room: J411
• Zoom link: https://gu-se.zoom.us/j/69780476534?pwd=Q9Uw2lu0zda8MsXkL08eGrqU64DMpp.1

Abstract

This talk addresses the challenges that LLMs face in sarcasm detection, for reasons that relate to the way humans use sarcasm. I provide empirical evidence about the factors that influence the production and interpretation of sarcasm by humans. Based on such evidence, I focus on two key challenges for LLMs in sarcasm detection: 1) generalizability and 2) instances where sarcasm fails to be communicated. Through modeling experiments, I show whether these models process sarcasm in a way that aligns with human interpretation, and whether they face similar challenges to humans. Finally, I discuss how the methodological framework proposed in this talk–connecting human data with modeling experiments–can provide a more comprehensive approach to understanding natural language in general.