Gender bias in machine learning: insights from official labour statistics and textual analysis

The interplay between technology and societal norms often reveals a troubling reality: machine learning systems not only reflect existing gender stereotypes but can also amplify and entrench them, making these biases harder to detect and address. This paper adopts an interdisciplinary approach, combining quantitative and qualitative methods with recent technological advancements, such as machine learning techniques for textual analysis and computational linguistics, to offer a new framework for understanding occupational gender bias in machine learning. The study is motivated by persistent gender inequalities in the labour market and rising concerns about gendered algorithmic bias, as outlined in the European Commission's Gender Equality Strategy 2020-2025. Focusing on language translation technologies, the research explores how machine learning may perpetuate or amplify gender stereotypes, aiming to foster more inclusive digital systems aligned with EU strategic goals. More specifically, it investigates occupational gender segregation and its manifestations in various forms of gender bias in machine learning across English, French, and Greek. The study introduces a classification of gender biases in machine learning, providing insights into professional areas needing intervention to address gender imbalances and identifying enduring stereotypical representations in textual data. To support this, statistical analysis is conducted to explore gender variations in occupations over the past thirteen years, using official data and international classifications such as the International Standard Classification of Occupations (ISCO-08). Moreover, gendered occupational distributions are extracted from 200,920 text instances in the three languages, revealing significant discrepancies between official labour statistics and the training data.