Exploration and mitigation of gender bias in word embeddings from transformer-based language models

Abstract

Machine learning has the potential to uncover data biases resulting from human error when it’s implemented without proper restraint. However, this complexity arises from word embedding, which is a prominent technique for capturing textual input as vectors applied in different machine learning and natural language processing tasks. Word embeddings are biased because they are trained on text data, which frequently incorporates prejudice and bias from society. These biases may become deeply established in the embeddings, producing unfair or biased results in AI applications. There are efforts made to recognise and lessen certain prejudices, but comprehensive bias elimination is still a difficult task. In Natural Language Processing (NLP) systems, contextualized word embeddings have taken the place of traditional embeddings as the preferred source of representational knowledge. It is critical to evaluate biases contained in their replacements as well since biases of various kinds have already been discovered in standard word embeddings. Our focus is on transformer-based language models, primarily BERT, which produce contextual word embeddings. To measure the extent to which gender biases exist, we apply various methods like cosine similarity test, direct bias test and ultimately detect bias through probability of filling MASK by the models. Based on this probability, we develop a novel metric called MALoR to observe bias. Finally, to mitigate the bias, we continue pretraining these models on a gender balanced dataset. Gender balanced dataset is created by applying Counterfactual Data Augmentation (CDA). To ensure consistency, we perform our experiments on different gender pronouns and nouns - “he-she”, “his-her” and “male names-female names”. These debiased models can then be used across several applications.