Summarization of Political Documents with Varied Perspectives

This technical report presents a novel dataset and task aimed at summarizing diverse political perspectives from opinionated news articles. The study introduces POLISUM, a framework designed to evaluate the performance of models in generating independent summaries for both left and right-leaning viewpoints. The report details the challenges posed by increasing global polarization, particularly during presidential elections, and emphasizes the importance of exposing users to alternative perspectives to mitigate this issue. The authors benchmark ten models of varying architectures, revealing that while some recent models perform well, they often struggle to accurately reflect the intended perspectives in their summaries. The report also discusses the factors influencing model performance, such as input document features and biases, and highlights the need for improved evaluation metrics in future work. The dataset comprises nearly 1,000 samples, providing a rich resource for researchers interested in political discourse and summarization techniques.