Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Bias in, Bias out: Annotation Bias in Multilingual Large Language Models

Xia Cui, Ziyi Huang, Naeemeh Adel

TL;DR

This paper addresses annotation bias in multilingual LLMs, showing how task framing, annotator subjectivity, and cultural mismatches distort data and model behavior. It introduces a three-part bias taxonomy (instruction, annotator, contextual/cultural), surveys detection metrics (IAA, model disagreement, metadata, multilingual divergence, cultural inference), and proposes an ensemble-based reactive mitigation via Weak Ensemble Learning (WEL) tailored for multilingual settings. It also presents an ethical analysis of annotation labour and outlines both proactive (diverse annotators, iterative guidelines, culturally grounded taxonomies) and reactive (post-hoc debiasing, fine-tuning, in-context debiasing, multi-objective ensembles) strategies. Collectively, these contributions aim to enable more equitable, culturally grounded annotation pipelines and improve the reliability of multilingual NLP systems.

Abstract

Annotation bias in NLP datasets remains a major challenge for developing multilingual Large Language Models (LLMs), particularly in culturally diverse settings. Bias from task framing, annotator subjectivity, and cultural mismatches can distort model outputs and exacerbate social harms. We propose a comprehensive framework for understanding annotation bias, distinguishing among instruction bias, annotator bias, and contextual and cultural bias. We review detection methods (including inter-annotator agreement, model disagreement, and metadata analysis) and highlight emerging techniques such as multilingual model divergence and cultural inference. We further outline proactive and reactive mitigation strategies, including diverse annotator recruitment, iterative guideline refinement, and post-hoc model adjustments. Our contributions include: (1) a typology of annotation bias; (2) a synthesis of detection metrics; (3) an ensemble-based bias mitigation approach adapted for multilingual settings, and (4) an ethical analysis of annotation processes. Together, these insights aim to inform more equitable and culturally grounded annotation pipelines for LLMs.

Bias in, Bias out: Annotation Bias in Multilingual Large Language Models

TL;DR

This paper addresses annotation bias in multilingual LLMs, showing how task framing, annotator subjectivity, and cultural mismatches distort data and model behavior. It introduces a three-part bias taxonomy (instruction, annotator, contextual/cultural), surveys detection metrics (IAA, model disagreement, metadata, multilingual divergence, cultural inference), and proposes an ensemble-based reactive mitigation via Weak Ensemble Learning (WEL) tailored for multilingual settings. It also presents an ethical analysis of annotation labour and outlines both proactive (diverse annotators, iterative guidelines, culturally grounded taxonomies) and reactive (post-hoc debiasing, fine-tuning, in-context debiasing, multi-objective ensembles) strategies. Collectively, these contributions aim to enable more equitable, culturally grounded annotation pipelines and improve the reliability of multilingual NLP systems.

Abstract

Annotation bias in NLP datasets remains a major challenge for developing multilingual Large Language Models (LLMs), particularly in culturally diverse settings. Bias from task framing, annotator subjectivity, and cultural mismatches can distort model outputs and exacerbate social harms. We propose a comprehensive framework for understanding annotation bias, distinguishing among instruction bias, annotator bias, and contextual and cultural bias. We review detection methods (including inter-annotator agreement, model disagreement, and metadata analysis) and highlight emerging techniques such as multilingual model divergence and cultural inference. We further outline proactive and reactive mitigation strategies, including diverse annotator recruitment, iterative guideline refinement, and post-hoc model adjustments. Our contributions include: (1) a typology of annotation bias; (2) a synthesis of detection metrics; (3) an ensemble-based bias mitigation approach adapted for multilingual settings, and (4) an ethical analysis of annotation processes. Together, these insights aim to inform more equitable and culturally grounded annotation pipelines for LLMs.

Paper Structure

This paper contains 32 sections, 15 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Background and Motivation
  3. Types of Annotation Bias
  4. Instruction Bias
  5. Annotator Bias
  6. Contextual and Cultural Bias
  7. Impact on Model Behaviour
  8. Case Studies and Empirical Evidence
  9. Case Study: Cross-Cultural Hate Speech Detection
  10. Case Study: Multimodal Emotion Recognition
  11. Detecting Annotation Bias
  12. Mitigation Strategies
  13. Proactive Strategies
  14. Diverse Annotator Pools
  15. Dynamic Annotation Guidelines
  16. ...and 17 more sections

Figures (2)

  • Figure 1: Taxonomy of annotation bias types observed in multilingual LLMs, showing three primary categories with distinct colouring: Instruction Bias, Annotator Bias, and Contextual & Cultural Bias.
  • Figure 2: Relationships between annotation bias sources, detection methods, and mitigation strategies. Solid lines indicate primary correlations; dashed lines (purple) show secondary cross-connections.