In the realm of digital information, ensuring the fairness and neutrality of textual content, especially news, is paramount. This paper introduces FairFrame, a novel framework engineered to both detect and mitigate bias in textual data. By harnessing the capabilities of state-of-the-art transformer models, FairFrame excels in identifying bias, surpassing the performance of current benchmarks. Additionally, the framework incorporates an explainable artificial intelligence (XAI) module based on Local Interpretable Model-agnostic Explanations (LIME), which aids in interpreting the rationale behind bias detection, thus fostering greater transparency. Uniquely, FairFrame employs large language models (LLMs) to mitigate detected biases through sophisticated few-shot prompting, marking a pioneering approach in the use of LLMs for bias mitigation. We validate the effectiveness of FairFrame through extensive experimental comparisons with leading fairness methods and an in-depth analysis of its components in diverse settings. The results demonstrate that FairFrame not only improves the detection of bias but also effectively mitigates it, offering a significant advancement in the development of fair artificial intelligence (AI) systems.
@article{sallami2024fairframe, title={Fairframe: a fairness framework for bias detection and mitigation in news}, author={Sallami, Dorsaf and A{"\i}meur, Esma}, journal={AI and Ethics}, pages={1--17}, year={2024}, publisher={Springer} }