Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Algorithmic Fairness in NLP: Persona-Infused LLMs for Human-Centric Hate Speech Detection

Ewelina Gajewska, Arda Derbent, Jaroslaw A Chudziak, Katarzyna Budzynska

TL;DR

This work tackles biases in hate speech detection arising from annotator identities by introducing Persona-LLMs that simulate in-group and out-group perspectives. It compares shallow and deeply contextualised persona prompting using Gemini-2.0-Flash and GPT-4.1-mini on a Toxigen-derived dataset across six identity pairings, evaluating with macro-$F1$, $FPR$, and $FNR$. Findings indicate that in-group-oriented annotations improve detection accuracy, and that deeply contextualised prompting generally yields higher accuracy while reducing bias gaps, compared to shallow prompting and non-persona baselines. The study highlights the potential and challenges of identity-aware annotation for fairer content moderation and advocates human-in-the-loop oversight to mitigate risks of stereotype reinforcement and adversarial misuse.

Abstract

In this paper, we investigate how personalising Large Language Models (Persona-LLMs) with annotator personas affects their sensitivity to hate speech, particularly regarding biases linked to shared or differing identities between annotators and targets. To this end, we employ Google's Gemini and OpenAI's GPT-4.1-mini models and two persona-prompting methods: shallow persona prompting and a deeply contextualised persona development based on Retrieval-Augmented Generation (RAG) to incorporate richer persona profiles. We analyse the impact of using in-group and out-group annotator personas on the models' detection performance and fairness across diverse social groups. This work bridges psychological insights on group identity with advanced NLP techniques, demonstrating that incorporating socio-demographic attributes into LLMs can address bias in automated hate speech detection. Our results highlight both the potential and limitations of persona-based approaches in reducing bias, offering valuable insights for developing more equitable hate speech detection systems.

Algorithmic Fairness in NLP: Persona-Infused LLMs for Human-Centric Hate Speech Detection

TL;DR

This work tackles biases in hate speech detection arising from annotator identities by introducing Persona-LLMs that simulate in-group and out-group perspectives. It compares shallow and deeply contextualised persona prompting using Gemini-2.0-Flash and GPT-4.1-mini on a Toxigen-derived dataset across six identity pairings, evaluating with macro-, , and . Findings indicate that in-group-oriented annotations improve detection accuracy, and that deeply contextualised prompting generally yields higher accuracy while reducing bias gaps, compared to shallow prompting and non-persona baselines. The study highlights the potential and challenges of identity-aware annotation for fairer content moderation and advocates human-in-the-loop oversight to mitigate risks of stereotype reinforcement and adversarial misuse.

Abstract

In this paper, we investigate how personalising Large Language Models (Persona-LLMs) with annotator personas affects their sensitivity to hate speech, particularly regarding biases linked to shared or differing identities between annotators and targets. To this end, we employ Google's Gemini and OpenAI's GPT-4.1-mini models and two persona-prompting methods: shallow persona prompting and a deeply contextualised persona development based on Retrieval-Augmented Generation (RAG) to incorporate richer persona profiles. We analyse the impact of using in-group and out-group annotator personas on the models' detection performance and fairness across diverse social groups. This work bridges psychological insights on group identity with advanced NLP techniques, demonstrating that incorporating socio-demographic attributes into LLMs can address bias in automated hate speech detection. Our results highlight both the potential and limitations of persona-based approaches in reducing bias, offering valuable insights for developing more equitable hate speech detection systems.
Paper Structure (15 sections, 3 figures, 3 tables)

This paper contains 15 sections, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Methodology
  4. Data
  5. Experimental Design
  6. Evaluation metrics
  7. Annotation Platform
  8. Shallow prompting.
  9. Deeply contextualised persona development.
  10. Results
  11. Shallow Personas.
  12. Deeply Contextualised Personas.
  13. Error Analysis.
  14. Discussion
  15. Conclusions

Figures (3)

  • Figure 1: A sample versus a population - how sampling bias of the annotator pool can influence hate speech annotation.
  • Figure 2: A distribution of annotations by ingroup (outward categories in the chart) and outgroup (inward categories) LLM annotators.
  • Figure 3: The process of developing and annotating text with deeply contextualised persona prompting.