DETECTION AND MITIGATION OF SOCIAL BIASES IN NATURAL LANGUAGE PROCESSING SYSTEMS

Abstract

There are now NLP systems at work in many fields, including virtual assistants, chatbots, legal document study and choosing candidates during recruiting. However, more and more, evidence suggests that these systems regularly display social biases by reflecting and even boosting stereotypes about gender, race, religion and other sensitive topics. As a result, biases can result in people being treated unfairly, discriminated against and losing their trust in automated systems, so they should be dealt with at the technical and ethical level. This thesis investigates whether bias appears in NLP models and suggests methods to find and reduce such bias. Starting with existing research, the study uncovers that bias can appear from skewed training data, not having the right model architecture and unbalanced pre-trained embeddings. To support detailed studies, a custom set of text with samples that are biased or unbiased was formed, carefully annotated by bias category and target groups. Assessment of bias detection approaches involved statistical tests, embedding association measures and transformer-based classification models. In order to address mitigation, the thesis explores adversarial debiasing, creating biased data to replace real data and refining models with fairness-based loss functions. Experimentally, it is evident that while no approach alone can solve this, mixing detection with mitigation strategies greatly lowers bias without affecting model quality a lot. The work adds to the research encouraging ethical AI and responsible NLP, helping provide practical advice on creating more equitable language technologies. In the end, the discussion covers the constraints, ethical points and future approaches to reach equity in NLP.