In recent years, the debate about the democratic health of the UK’s media landscape has intensified, with several commentators and scholars highlighting a rather distorted system, characterised by a mostly hyper-partisan press. These findings are concerning, given that the integrity of the democratic process requires a balanced and diverse media landscape, providing citizens with access to a variety of viewpoints about public policy choices. There is also evidence that these distortions can affect news coverage by 'trusted sources', such as the BBC, who have a statutory commitment to impartiality. Effective solutions to capturing media representations of the political discourse are therefore essential for evaluating the health of a democracy and informing regulatory policies. However, the analyses carried out in the media and political science literature have typically relied on manual investigations, a state of affairs that limits their scalability and applicability to large news corpora. In this talk we present our work on developing AI solutions for modelling and analysing media representations of the political discourse accurately and at scale. In particular, I will illustrate some fundamental work that has produced a formal model of the news classification task, as well a computational solution that, for a given topic (e.g., the debate about UK immigration), follows a hybrid human-AI pipeline able i) to identify individual statements ("claims") in the media by contributors to the political debate; ii) to identify the political viewpoints expressed in the news relevant to the topic in question — viewpoints are characterised as abstractions of individual positions that are semantically and ideologically coherent (e.g., all statements which characterise immigrants as a threat to the UK's national security); and iii) to classify all claims with respect to the applicable political viewpoints. Our results on a challenging dataset covering media representations of the immigration debate in the UK demonstrate that a neuro-symbolic architecture combining a state-of-the-art large language model (LLM) with a knowledge graph (KG), which captures background information about the actors taking part in the political debate, provides excellent performance on our testbed. From an application point of view, our results also appear to confirm the aforementioned concerns about lack of fairness and balance in the UK's media landscape, by indicating a significant lack of viewpoint diversity in media coverage of the immigration debate.

Watch the live webcast >>