Debating Inequality: An LLM Approach to Analyzing Debates in Westminster

Economic inequality is rising in advanced economies (Piketty 2015). The widening disparities have a complex and negative impact, which touches many parts of society. Access to housing, education, healthcare, jobs, and other critical resources are affected. It is unsurprising then, to find that discussion of inequality in the UK has increased significantly in the 21st century (Hansard.gov.uk). Whether politicians seek to address inequality directly or seek to frame other issues in terms of their connection to inequality, they must choose what facet of inequality they will focus on. The consequences of inequality are simply too broad to be addressed holistically and politicians must select their focus strategically. As such, it is important to understand how politicians and political parties choose to conceptualize and frame inequality’s impact. We aim to do this by analyzing text from the last twenty-four years of debates in Westminster Hall. Westminster Hall debates are agenda setting moments for UK MPs, who are able to raise issues of their choice, and are guaranteed responses from the relevant minister, and the heads of major parties. We assume that MPs raising debates will act strategically and raise issues that are especially important to them, which makes the debates an excellent opportunity to learn about a diverse set of salient issues while also receiving a number of relevant perspectives. We evaluate the texts on a paragraph-by-paragraph basis using large language models (LLMs), which are tasked with answering a series of questions about the topic of the paragraph – specifically whether it addresses some facet of economic inequality. Having identified instances where politicians are discussing inequality, we analyze the text again. This time asking the LLM to identify demographic groups and resources discussed. This two-step process provides us with an acceptable set of F-scores and a dataset that provides semantic coding of all debates held in Westminster Hall since 2000. We pair this information with meta-data on the debates, the speakers themselves, and the districts they represent. We then employ produce a series of descriptive and inferential analyses to highlight how discussions of inequality differ between MPs, across parties, and through time. Our preliminary analyses indicate that the agenda setting decisions of politicians vary both within and across parties, and that (at least when speaking in Westminster Hall) politicians’ approaches to inequality are dependent upon the economic conditions in their home constituency. Ultimately, we hope to provide similar insights into how discussions of inequality have changed across the major political and economic inflection points of the last twenty years. Thus, our paper makes two important contributions, first by providing a semantic mapping of how inequality has been discussed by UK MPs and second by offering an example of how LLMs can substitute for largescale hand coding to augment legislative analyses.