A Framework for the Unsupervised Modeling and Extraction of Polarization Knowledge from News Media

Polarization poses global concerns for social cohesion and stability, making its understanding crucial for effective mitigation measures. In this paper, we introduce an unsupervised, domain-agnostic framework for computationally modeling, extracting, and measuring polarization in digital media. By leveraging Natural Language Processing and Graph Analysis techniques, the proposed framework creates a Polarization Data Model (PDM) that encompasses key elements of Polarization Knowledge (PK), such as entities, fellowships, dipoles, and discussion topics. To evaluate the effectiveness of the framework, we propose a multi-level PK evaluation methodology that assesses its ability to: (i) capture entities’ attitudes toward various topics, (ii) align politically cohesive fellowships with their respective party manifestos, and (iii) identify domain-specific topics along with their degree of polarization. We applied this evaluation methodology to the use cases of Abortion, Immigration, and Gun Control. The results demonstrate our framework’s robust performance across these case studies, yielding promising outcomes compared to state-of-the-art and baseline methods.