Summary: | As global information consumption accelerates, the demand for concise, relevant, and trustworthy news summaries has grown significantly. Large Language Models (LLMs) have demonstrated powerful capabilities in text summarization, yet current methods struggle to balance two critical goals: personalizing content to reflect user preferences and ensuring neutrality by mitigating bias. This thesis proposes a novel framework for bias-aware, personalized summarization of global news using LLMs, aiming to generate user-aligned summaries without distorting factual accuracy or fairness. The study introduces a prompt-guided personalization strategy where user preferences, modeled as a set of attributes such as reading level, content structure, and abstractivity, are embedded into dynamically generated prompts. We evaluate three summarization paradigms: generic, attribute-guided, and prompt-based. Summaries are assessed using an LLM-as-a-judge approach across four dimensions: factual consistency, bias neutrality, personalization fit, and readability. Our evaluation combines GPT-4o and Gemini scoring with selective human validation to ensure robust measurement. Results reveal that prompt-based summarization, particularly with neutrally phrased prompts, achieves the highest alignment with user preferences while also improving or maintaining neutrality and factual integrity. Contrary to concerns that personalization may introduce bias, we observe positive correlations between personalization fit, factual accuracy, and fairness. Readability scores remain uniformly high across all approaches. This work contributes a scalable methodology for personalized summarization with bias control and demonstrates the potential of LLMs to deliver nuanced, context-aware summaries. By bridging personalization and neutrality, the framework sets a foundation for developing future AI systems that are both user-centric and ethically responsible.
|