Are Stereotypes Leading LLMs' Zero-Shot Stance Detection ?

TL;DR

This paper investigates whether large language models exhibit stereotypes in zero-shot stance detection by annotating stance datasets with African American English versus Standard American English and text readability using the Flesch–Kincaid measure. It evaluates GPT-3.5-turbo and five open models with a Context Analyze prompting approach, measuring performance via weighted F1 and fairness via Equal Opportunity. Key contributions include revealing systematic stereotype-driven biases, releasing enriched datasets, and proposing fairness-aware prompting or causal debiasing as paths forward. The findings underscore the need for robust, equitable stance detection and caution against deploying bias-prone LLMs in politically sensitive contexts, while offering practical avenues to mitigate such biases.

Abstract

Large Language Models inherit stereotypes from their pretraining data, leading to biased behavior toward certain social groups in many Natural Language Processing tasks, such as hateful speech detection or sentiment analysis. Surprisingly, the evaluation of this kind of bias in stance detection methods has been largely overlooked by the community. Stance Detection involves labeling a statement as being against, in favor, or neutral towards a specific target and is among the most sensitive NLP tasks, as it often relates to political leanings. In this paper, we focus on the bias of Large Language Models when performing stance detection in a zero-shot setting. We automatically annotate posts in pre-existing stance detection datasets with two attributes: dialect or vernacular of a specific group and text complexity/readability, to investigate whether these attributes influence the model's stance detection decisions. Our results show that LLMs exhibit significant stereotypes in stance detection tasks, such as incorrectly associating pro-marijuana views with low text complexity and African American dialect with opposition to Donald Trump.

Figures (30)

• Figure 1: Proportion of SAE and AAE tweets in the PStance dataset for each political figure
• Figure 2: Proportion of readability classes in the SCD dataset for each topic
• Figure 3: Proportion of readability classes in the KE-MLM dataset for each political figure
• Figure 4: Equality of Opportunity on Joe Biden on the PStance dataset. SAE stands Standard American English, AAE for African American English. In green, EO for the label "favor", in red for "against".
• Figure 5: Equality of Opportunity on Bernie Sanders on the PStance dataset. SAE stands Standard American English, AAE for African American English. In green, EO for the label "favor", in red for "against".