Abstract
Xiaolei Huang, Linzi Xing, Franck Dernoncourt, Michael J. Paul. Proceedings of the 12th Language Resources and Evaluation Conference.
The authors present the first multilingual Twitter corpus for hate speech detection that includes inferred author demographics (age, gender, race, and country) across five languages. By evaluating standard classifiers on this data, they demonstrate that language variations across demographic groups lead to measurable algorithmic bias, providing a new benchmark for developing fairness-aware NLP models.
Introduction
Hate speech detection is a critical but sensitive task. Research has shown that machine learning models often "learn" human biases from their training data, leading to discriminatory outcomes—for example, a model might be more likely to flag posts from certain ethnic groups as "toxic" regardless of content.
The authors identify two major gaps in current fairness research: the lack of ground truth demographics in existing datasets and a heavy focus on English-only benchmarks. Linguistic variations, such as gendered nouns or regional dialects, mean that bias manifests differently across languages.
Methodology
The researchers assembled a corpus by combining existing hate speech datasets and augmenting them with author-level demographic information across five languages: English, Italian, Polish, Portuguese, and Spanish. The final corpus includes over 100,000 documents.
Demographic Inference and Validation
To ensure demographics were independent of tweet text, the authors used profile metadata. They employed Face++ (a computer vision API) to estimate age, race, and gender from profile pictures, and the Google Maps API to resolve location strings. To ensure the reliability of these automated tools, they validated the inferences against human annotations from a crowdsourcing platform.
Predictability Analysis
A key component of the study was testing how well author demographics could be predicted from the text itself using word-level (TF-IDF), POS-level (parts of speech), and topic-level (LDA) features. This analysis helps determine if text acts as a proxy for demographic identity, which can lead to indirect bias in classifiers.
Results and Insights
Text as a Demographic Proxy
The authors found that demographics are highly predictable from text alone, especially age and gender. This confirms that even if a model isn't "told" a user's race or age, it can pick up on linguistic cues that act as proxies, leading to indirect bias.
Model Performance and Fairness
The study evaluated four architectures: Logistic Regression (LR), CNN, RNN (bi-GRU), and BERT.
- Performance: The RNN (bi-GRU) generally achieved the best balance of accuracy and F1-score. BERT performed relatively poorly, likely due to the domain shift from its pre-training data (Wikipedia) to short-form tweets.
- Fairness Metric: Using Equality Difference (ED), the authors found no significant correlation between raw accuracy and fairness. Interestingly, while BERT had lower accuracy, it often exhibited the least amount of demographic bias (lowest SUM-ED scores), while Logistic Regression was frequently the most biased.
Language-Specific Bias
Bias patterns varied by language. For instance, models trained on Polish and Italian data were most biased by Age and Gender, while Spanish and Portuguese models showed higher bias regarding the author's Country.
Discussion and Limitations
While this corpus provides a valuable benchmark, the authors acknowledge several limitations. Binary gender and four-category race labels are oversimplifications that exclude non-binary individuals and smaller minority groups. Furthermore, using computer vision to "guess" demographics can introduce its own set of algorithmic biases, and original hate speech labels may already contain the biases of the initial annotators.
Future work should expand to more granular demographic categories and explore active "de-biasing" techniques using this multilingual framework.
Summary updated