Gender Bias in Neural Natural Language Processing

TL;DR

This paper investigates gender bias in neural NLP systems, introduces a benchmark for quantifying bias, and proposes a causal intervention method (CDA) that effectively reduces bias without sacrificing model accuracy.

Contribution

It presents a new benchmark for measuring gender bias in neural NLP and introduces CDA, a novel corpus augmentation technique that mitigates bias more effectively than existing methods.

Findings

CDA significantly reduces gender bias in neural NLP models.

CDA outperforms traditional word embedding debiasing methods.

Bias increases during training as the model's loss decreases, but CDA mitigates this effect.

Abstract

We examine whether neural natural language processing (NLP) systems reflect historical biases in training data. We define a general benchmark to quantify gender bias in a variety of neural NLP tasks. Our empirical evaluation with state-of-the-art neural coreference resolution and textbook RNN-based language models trained on benchmark datasets finds significant gender bias in how models view occupations. We then mitigate bias with CDA: a generic methodology for corpus augmentation via causal interventions that breaks associations between gendered and gender-neutral words. We empirically show that CDA effectively decreases gender bias while preserving accuracy. We also explore the space of mitigation strategies with CDA, a prior approach to word embedding debiasing (WED), and their compositions. We show that CDA outperforms WED, drastically so when word embeddings are trained. For pre-trained embeddings, the two methods can be effectively composed. We also find that as training proceeds on the original data set with gradient descent the gender bias grows as the loss reduces, indicating that the optimization encourages bias; CDA mitigates this behavior.