Research Question
Do toxicity classification models disproportionately flag African American English (AAE) as toxic compared to Standard American English (SAE), and can this disparity be reduced without significantly harming overall model performance?
Motivation
Automated toxicity detection is increasingly used to moderate online content, but prior research has shown these systems can encode racial and dialectal bias. Misclassifying AAE as toxic at higher rates can silence and disproportionately penalize Black social media users. This project investigated the scope of that bias and evaluated whether established fairness interventions could meaningfully reduce it.
Dataset
We used the Davidson hate speech and offensive language dataset alongside the TwitterAAE corpus, which provides dialect-probability labels for tweets, to evaluate model behavior across AAE- and SAE-associated text.
My Contribution
This was a group research project. I implemented and evaluated the XGBoost and vector-scaling mitigation approaches, contributed to the adversarial-debiasing experiments, and analyzed subgroup fairness and performance tradeoffs across all methods.
Methods
- BERT embeddings for text representation
- XGBoost classification baseline
- Vector scaling — post-hoc calibration to equalize subgroup score distributions
- Reweighting — adjusting training sample weights across dialect subgroups
- Adversarial debiasing — an adversarial network trained to remove dialect signal from learned representations
Evaluation Framework
Model performance and fairness were evaluated using accuracy, F1 score, subgroup false-positive rate (FPR) and false-negative rate (FNR), and disparate impact, comparing outcomes for AAE-associated and SAE-associated text.
Results
Across the mitigation methods evaluated, vector scaling and adversarial debiasing produced the largest reductions in the AAE–SAE false-positive-rate gap, without a significant loss in overall classification accuracy or F1 score.
Fairness–Performance Tradeoffs
Reducing subgroup disparities in FPR required tradeoffs against raw classification accuracy in some configurations. The strongest bias-mitigation methods did not universally dominate on every metric, underscoring that fairness interventions typically involve balancing multiple, sometimes competing, objectives rather than a single optimal solution.
Limitations
This evaluation was limited to two datasets and a specific set of mitigation techniques, and dialect was approximated using probabilistic labels rather than self-reported speaker identity. Results may not generalize to other toxicity classifiers, languages, or dialect groups without further evaluation.
Future Work
Future work could extend this evaluation to additional model architectures, incorporate more representative dialect-labeled datasets, and explore fairness interventions that generalize across multiple protected attributes simultaneously.
Links
This was a team repository; a full written report is available upon request.