Towards a Robust Gender Bias Evaluation in NLP

Abstract

With the advent of deep learning technology, Natural Language Processing (NLP) has made remarkable progress. Deep learning models have improved the performance of many NLP tasks such as text summarization, translation, and sentiment analysis. However, NLP models have been shown to present gender biases, which can be detrimental to decision-making. As a result, assessing the gender bias of these models before deploying them is a must. We develop the Gender Bias Evaluation Framework (GBEF), a framework that measures gender bias in Masked Language Models (MLMs). The GBEF consists of two approaches. Each approach uses preconstructed data and a gender bias metric. The first evaluation approach is called the Sentence-Based Evaluation (SBE) and it can assess gender bias in three different categories: occupation bias, benevolent sexism, and hostile sexism. The second approach is called the Template-Based Evaluation (TBE) and will be used for a more accurate assessment of the counterfactual data substitution debiasing technique, a technique that relies on balancing female-related words and male-related words in the training corpus. We first use the SBE to quantify gender bias in different BERT models and show that BERTlarge is the most biased model while RoBERTalarge is the least gender biased one. The SBE was also used to quantify gender bias in corpora. We develop a new method for this task that relies on fine-tuning BERT for the masked language model task on the corpus on which we want to measure the bias. We compare Jigsaw’s toxic comments with Jigsaw’s severe toxic comments and reveal that the latter presents a higher degree of gender bias. Finally, the TBE was able to shed light on the issues of the debiasing technique that relies on fine-tuning BERT on a counterfactual data substituted corpus. While this technique was able to reduce gender bias in BERT at a high level, we show with the TBE that the model is simply treating male and female-related pronouns as equal, which may be problematic when it comes to gender-related words (e.g., pregnant). We propose a solution to this problem by including sentences with gender-related words in the training corpus. The inclusion of these sentences in the training corpus allowed the debiased version of BERT to associate gender-related words with the right gender. We believe that our proposed evaluation framework will aid in a more accurate assessment of the gender bias in different MLMs improving fairness in artificial intelligence.