<p>This paper introduces a computational framework for classifying entity role framing in unstructured textual data, with a significant application in quantitatively measuring and understanding political bias in news media. Traditional approaches to political bias analysis have often relied on less interpretable quantitative metrics, creating a demand for more transparent and nuanced computational solutions. The interpretability of our approach stems from the explicit identification of entities and the specific roles (hero, villain, victim) assigned, offering a direct insight into the narrative strategies employed. Building on previous works on computational modelling of entity framing in multimodal content, we examine the efficacy of textual features for this task, using three experimental setups of increasing complexity and two language models. Crucially, when applied to a corpus of politically diverse news outlets in the US and UK, the framework successfully quantifies polarised patterns in how media sources frame political entities, demonstrating clear alignment with established media bias measurements. This methodological framework provides a systematic approach to mapping entity framing patterns to political orientation, thereby creating a more interpretable and nuanced measure of polarisation in digital news ecosystems. The framework can serve as a valuable tool for scholars and practitioners in political communication, journalism, and computational social science seeking to understand the dynamics of digital news ecosystems.</p>