Reproduce “Entropy-based Attention Regularization Frees Unintended Bias Mitigation from Lists” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Giuseppe Attanasio, Debora Nozza, Dirk Hovy, Elena Baralis
Researchers from Bocconi University and Politecnico di Torino introduced Entropy-based Attention Regularization (EAR), a knowledge-free method to mitigate unintended bias in hate speech detection models. This approach, which maximizes attention entropy, consistently improved bias metrics and reduced false negative rates by an average of 15.04% compared to baselines, without relying on predefined identity term lists.
Key insights
- Maximizing attention entropy during training encourages models to build robust token representations from a wider context, reducing over-reliance on individual, potentially biased, terms.
- Low attention entropy for a token indicates that its representation is formed by overly concentrating attention on a narrow context, suggesting lexical overfitting.
- Attention entropy can serve as a diagnostic tool to automatically identify terms that are most likely inducing unintended bias due to overfitting within the model.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.