SS25: Pragmatic Processing in Humans Versus Language Models

Seminar Description

In this seminar, we will explore the differences and similarities in how humans and large language models (LLMs) process pragmatic information. Pragmatic processing involves interpreting language beyond its literal meaning, incorporating context, background knowledge, and social cues. While LLMs have demonstrated exceptional abilities in various language tasks, it remains uncertain whether their capabilities compare to human performance when it comes to more creative tasks, such as the interpretation and generation of novel metaphors, humour comprehension, or other complex aspects of language that may require additional cognitive abilities. 

The first two sessions will provide participants with an overall understanding of how human data is collected in psycholinguistics and how LLMs are being used to generate datasets, followed by how these models are fine-tuned to enhance their understanding of pragmatic reasoning. Each participant will present a paper on a relevant topic, such as comparing humans and LLMs in any language task that requires creativity or processing of pragmatic cues. 

The goal of the seminar is to assess the current state of LLMs' ability to handle such tasks and discuss the implications of these findings for improving LLMs' capacity to process language in a more human-like way, as well as the challenges in aligning their outputs with human-level pragmatic reasoning.