Fairness in Opinion Summarisation in the Era of Large Language Models

Opinion summarisation is the computational task of automatically generating concise summaries from collections of subjective texts, such as product reviews, and social media posts. This process involves extracting and synthesising key opinions, sentiments, and perspectives from large volumes of user-generated content to provide readers with concise overviews of users' opinions.Fairness in opinion summarisation entails the equitable inclusion of diverse viewpoints and that the distribution of represented viewpoints reflects the underlying diversity of the source material rather than amplifying dominant voices at the expense of marginalised ones. While large language models have been increasingly used for text summarisation, bias can be introduced at various stages from pretraining to model adaptation and finetuning for downstream tasks such as summarisation. This thesis makes several key contributions to the field of fair opinion summarisation. Firstly, it provides a comprehensive examination of general fairness issues in opinion summarisation systems, establishing a theoretical framework for understanding bias manifestation in automated summarisation processes. Secondly, it introduces novel evaluation metrics specifically designed to measure fairness in opinion summarisation, offering quantitative tools for assessing bias. Thirdly, the work investigates various model editing techniques and their differential effects on fairness outcomes, providing empirical insights into how different approaches to model modification impact equitable representation. Finally, the thesis presents an innovative cognitive-inspired prompting method tailored specifically for debiasing opinion summarisation. This contributes a practical solution for mitigating unfairness in real-world applications whilst maintaining summarisation quality and coherence.