8th September 2023
AI applications can be beneficial in cutting costs in HR departments, leading to their increasing use in practice. The range of potential applications is vast, covering aspects from recruiting to employee management. In recruitment, technologies can rank applicants, pre-select them, or even conduct job interviews with chatbots or avatars. In some cases, especially when processing or generating text, word vectors or state-of-the-art language models are employed.
Language models are trained using extensive text data, allowing for mathematical computations to handle word relationships. In the simpler case, each word is assigned a mathematical vector. These vectors enable computations between words, such as deriving “Estonia” from the vectors of “Tallinn,” “Paris,” and “France.” This capability is invaluable for various applications, but these models, trained on society’s texts, also encode its stereotypes, potentially leading to unwanted associations. For instance, vectors for gender-specific names may be closer to family-related topics than career-related ones, impacting decisions and generated texts.
One component of the Debiaser, developed within the BIAS project, focuses on identifying and mitigating such bias in word vectors and language models. While much research has concentrated on the English language, this EU-funded project takes into account European languages, recognizing the language and cultural variations that make addressing bias a challenging and region-specific endeavour.