Between the Lines: Contextual Understanding and Bias in LLMs

Dr Nafise Sadat Moosavi

School of Computer Science, Sheffield University

Abstract:

Contextual understanding is central to the performance of Large Language Models, yet it remains one of their most persistent challenges. In this talk, I will explore how LLMs navigate (and often fail to navigate) the complexities of context, illustrated through two case studies: idiomaticity detection and fairness in identifying dehumanizing language.

Idiomaticity detection requires models to distinguish between literal and figurative meanings, relying on nuanced contextual reasoning. Our experiments with a controlled contrastive dataset reveal that even state-of-the-art models struggle to resolve such ambiguities, often failing to attend to subtle contextual cues.

In the domain of fairness, we investigate how LLMs identify dehumanizing language, a critical task where understanding social and linguistic contexts is key. Biases in the models' sensitivity to different protected attributes highlight an incomplete grasp of the contexts in which dehumanization occurs, underscoring ethical implications for their deployment in sensitive applications. By reading between the lines, this talk shows the limitations of LLMs in contextual understanding across linguistic and ethical dimensions

Tehran Time: May 19, 2025 - 13:30 to 15:00

Link:

https://webinar.ihcs.ac.ir/rooms/8vr-b1h-dsj-1so/join