The Effects of Demographic Instructions on LLM Personas

TL;DR

The paper tackles the subjectivity of sexism detection on social media by applying a perspectivist labeling approach using the EXIST dataset and a range of LLMs. It investigates whether demographic prompts can induce LLMs to adopt personas and align with different human annotator groups, quantifying agreement with Krippendorff's $\alpha$ and bootstrap CIs. Results show that LLMs consistently align more with female annotators than male, with age effects varying by model, and that demographic persona prompting yields inconsistent, model-dependent effects. The findings suggest that demographic-based prompting is not a robust method for mitigating bias in sexism classification and highlight the need for more nuanced demographic modeling and evaluation across tasks and models.

Abstract

Social media platforms must filter sexist content in compliance with governmental regulations. Current machine learning approaches can reliably detect sexism based on standardized definitions, but often neglect the subjective nature of sexist language and fail to consider individual users' perspectives. To address this gap, we adopt a perspectivist approach, retaining diverse annotations rather than enforcing gold-standard labels or their aggregations, allowing models to account for personal or group-specific views of sexism. Using demographic data from Twitter, we employ large language models (LLMs) to personalize the identification of sexism.

Figures (1)

• Figure 1: Prompt structure. Bold text indicates the parts that were added/varied in the various experiments, as extensions beyond the "baseline" prompt.