Reducing Gender Bias in Abusive Language Detection

TL;DR

Gender bias in abusive language detection arises from imbalanced identity terms in training data, causing false positives for certain identities. The authors measure biases with unbiased test sets using identity templates and compare CNN/GRU/alpha-GRU with various embeddings. They propose three mitigation methods—debiased embeddings, gender swap augmentation, and bias-aware fine-tuning—and show substantial reductions in bias (up to 90-98%) with some performance trade-offs. The work highlights the importance of evaluating NLP systems for fairness and offers extensible strategies for reducing bias in other tasks and identities.

Abstract

Abusive language detection models tend to have a problem of being biased toward identity words of a certain group of people because of imbalanced training datasets. For example, "You are a good woman" was considered "sexist" when trained on an existing dataset. Such model bias is an obstacle for models to be robust enough for practical use. In this work, we measure gender biases on models trained with different abusive language datasets, while analyzing the effect of different pre-trained word embeddings and model architectures. We also experiment with three bias mitigation methods: (1) debiased word embeddings, (2) gender swap data augmentation, and (3) fine-tuning with a larger corpus. These methods can effectively reduce gender bias by 90-98% and can be extended to correct model bias in other scenarios.