Fair Summarization in the Age of Large Language Models: A Comprehensive Survey

Fairness in summarization has been getting increasing attention recently. Achieving fairness through Large Language Models (LLMs) is particularly of special interest as LLMs are considered state-of-the-art models in various Natural Language Tasks, especially considering that LLM generated summaries are largely preferred by humans. As these automatic summarizers are used more extensively, they propagate and amplify biases present in source material.Few papers have explored the entirety of the landscape on the types of biases, notions of fairness, and strategies of mitigation. In this paper, we attempt to bridge this gap and provide a comprehensive overview of the problem of increasing fairness in text summarization and strategies to overcome it. Specifically, we present a brief introduction into why fairness in summarization is crucial from a societal standpoint. We also thoroughly explore the various possible definitions of fairness and how bias is manifested in LLMs trained on skewed data. Additionally, we provide an examination of the various mitigation techniques that have been used to enhance fairness and the challenges that continue to impede progress.