Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Evaluating Simple Debiasing Techniques in RoBERTa-based Hate Speech Detection Models

Diana Iftimie, Erik Zinn

TL;DR

The paper tackles dialect-induced bias in hate speech detection by evaluating two simple debiasing techniques—alternating adversarial debiasing and gradient negation debiasing—applied to RoBERTa-based encoders. It leverages the Founta dataset augmented with African-American English (AAE) and White-Aligned English (WAE) labels to study how training-data construction, specifically representation bias, shapes debiasing effectiveness. The findings show that debiasing performance is highly sensitive to data construction: with representation bias accounted for, debiasing can reduce disparities, whereas in biased data it yields limited fairness gains and may slightly affect accuracy. The work offers practical guidance on dataset design and fair debiasing for hate speech systems, and suggests avenues for stronger evaluation and further refinement of debiasing strategies.

Abstract

The hate speech detection task is known to suffer from bias against African American English (AAE) dialect text, due to the annotation bias present in the underlying hate speech datasets used to train these models. This leads to a disparity where normal AAE text is more likely to be misclassified as abusive/hateful compared to non-AAE text. Simple debiasing techniques have been developed in the past to counter this sort of disparity, and in this work, we apply and evaluate these techniques in the scope of RoBERTa-based encoders. Experimental results suggest that the success of these techniques depends heavily on the methods used for training dataset construction, but with proper consideration of representation bias, they can reduce the disparity seen among dialect subgroups on the hate speech detection task.

Evaluating Simple Debiasing Techniques in RoBERTa-based Hate Speech Detection Models

TL;DR

The paper tackles dialect-induced bias in hate speech detection by evaluating two simple debiasing techniques—alternating adversarial debiasing and gradient negation debiasing—applied to RoBERTa-based encoders. It leverages the Founta dataset augmented with African-American English (AAE) and White-Aligned English (WAE) labels to study how training-data construction, specifically representation bias, shapes debiasing effectiveness. The findings show that debiasing performance is highly sensitive to data construction: with representation bias accounted for, debiasing can reduce disparities, whereas in biased data it yields limited fairness gains and may slightly affect accuracy. The work offers practical guidance on dataset design and fair debiasing for hate speech systems, and suggests avenues for stronger evaluation and further refinement of debiasing strategies.

Abstract

The hate speech detection task is known to suffer from bias against African American English (AAE) dialect text, due to the annotation bias present in the underlying hate speech datasets used to train these models. This leads to a disparity where normal AAE text is more likely to be misclassified as abusive/hateful compared to non-AAE text. Simple debiasing techniques have been developed in the past to counter this sort of disparity, and in this work, we apply and evaluate these techniques in the scope of RoBERTa-based encoders. Experimental results suggest that the success of these techniques depends heavily on the methods used for training dataset construction, but with proper consideration of representation bias, they can reduce the disparity seen among dialect subgroups on the hate speech detection task.
Paper Structure (16 sections, 2 equations, 12 figures)

This paper contains 16 sections, 2 equations, 12 figures.

Table of Contents

  1. Introduction and Background
  2. Data
  3. Methodology
  4. Encoder, Tokenizer, and Preprocessing
  5. Alternating Adversarial Debiasing Technique
  6. Gradient Negation Debiasing Technique
  7. Additional Training Specifications
  8. Evaluation
  9. Results and Discussion
  10. Conclusion
  11. Acknowledgments
  12. Appendix
  13. Detailed Model Architecture Diagrams
  14. Exploration of Model Prediction Errors
  15. Per-Class False Positive Rates for Two-Class Hate Speech Detection Models
  16. ...and 1 more sections

Figures (12)

  • Figure 1: Founta Dataset Tree Map, grouped by Hate Speech classes and Dialect classes
  • Figure 4: High-level model architecture for Alternating Adversarial Debiasing Technique
  • Figure 5: Example training progress for Alternating Adversarial Debiasing Technique
  • Figure 6: High-level training technique for Gradient Negation Debiasing Technique
  • Figure 7: Performance results (accuracies and FPRs) for four-class hate speech detection models
  • ...and 7 more figures