This paper reviews the use of word embeddings in investigating bias. Word embeddings have emerged as a powerful technique for representing words as numeric vectors. This allows for a powerful way of analysing text data. By integrating human understanding of words into machine processing, word embeddings offer a valuable tool for studying human behaviour and social phenomena. Specifically, word embeddings have shown promise in capturing biases in human communication that may not be communicated explicitly. This paper explores, through a systematic literature review, the current challenges of using word embeddings to investigate bias and the strategies that can be employed to ensure the effective utilisation of word embeddings for bias investigation. In doing so, the paper examines the current state of research, emerging trends, gaps, and opportunities in this area. Preliminary results reveal, inter alia, that word embeddings predominantly capture gender bias, while other biases such as race, class, ethnicity, and religion remain largely underexplored. Additionally, we observed a significant challenge in this research domain, as it predominantly relies on US English text, neglecting the incorporation of other linguistic contexts. This review extends scholarship on prejudice and stereotyping by providing valuable resource for harnessing the power of word embeddings in understanding various forms of bias in public opinion and effective bias mitigation strategies.

Authors: Nnaemeka Ohamadike, Kevin Durrheim, Mpho Raborife

Published in: International Conference on Information Society (i-Society-2023)

• Date of Conference: 24-26 October 2023
• DOI: 10.20533/iSociety.2023.0003
• ISBN: 978-1-913572-69-3
• Conference Location: Dún Laoghaire, Ireland