Fairness in Language Models (en Inglés) - Wenbin Zhang

As language models increasingly influence critical decisions in healthcare, hiring, and criminal justice, their capacity to perpetuate and amplify societal biases poses significant risks to marginalized communities. Although awareness of these fairness issues is growing, practitioners still face many barriers. The proliferation of competing fairness definitions leads to conceptual confusion and the lack of systematic guidance on how to select appropriate evaluation methods and mitigation strategies, whereas bias metrics are scattered across disconnected sources. This lack of structure has hindered progress in building fair and trustworthy language models. Motivated by these challenges, this book provides the first systematic, architecture-aware guide to bias in modern language models, offering a unified framework that synthesizes theory, measurement, and practical solutions. Covering models from BERT to GPT and beyond, the book gives readers the tools they need to understand and address bias effectively. 

Bridging theory and practice, this book takes readers through the entire fairness process step by step. It starts with the history of language models, from basic statistical models to transformers. It explains how bias appears in training data, embeddings, and annotation processes. Next, the book introduces a novel two-tiered framework for bias quantification that organizes metrics according to model architecture, including encoder-only, decoder-only, and encoder-decoder models. This framework resolves confusion around competing fairness definitions that have fragmented the field. Building on this foundation, the book introduces a comprehensive taxonomy of mitigation techniques across pre-processing, in-processing, intra-processing, and post-processing approaches. The book also provides an in-depth analysis of evaluation datasets and a decision-tree selection framework. The final chapter explores emerging challenges, including intersectional fairness, adversarial robustness, and human-AI fairness comparisons.

This book is written for AI researchers, machine learning engineers, and policymakers. It brings together scattered research into one clear resource that balances practical advice with solid theory. By the end, readers will have the knowledge and tools they need to check, measure, and mitigate bias in language models used at scale. A basic understanding of machine learning and natural language processing is recommended.