The Technical Setup of the BIAS Project: Detecting and Mitigating Biases Related to the Labour Market

Biases in applications of artificial intelligence are known to be a major issue. Historical biases in training data of machine learning software, or societal stereotypes reflected in the textual training data of large language models are posing a major risk of adverse impact. In this paper, preliminary results from the BIAS project investigate how biases are reflected in AI applications for the labour market, focusing on an interdisciplinary approach to ensure recruitment systems are trustworthy and socially responsible. This involves developing innovative, transparent, and fair methods using case-based reasoning, alongside novel techniques to detect and mitigate societal and intersectional biases in word embeddings and language models. The focus of the project is on European languages, advancing the field of non-English natural language processing research.