An Argumentation Framework for Graph RAG with UK Political Personas

ACL ARR 2025 May Submission3726 Authors

Abstract: We present an argumentation framework that was instantiated using argumentative data from 30 debates that aired on the BBC television politics programme \emph{Question Time} throughout 2020 and 2021. We then tasked 13 generative models with predicting the political position of the dialogue locution and proposition stored within each node of the argumentation graph. From this, we were able to compute an ensemble average political position and show how the variance in those predictions was reduced by removing smaller large language models (LLMs). Results demonstrate that the utterances and resolved propositions were, on average, estimated to be left of centre, with the average political position per episode changing, possibly reflecting different locations where the television programme took place within the UK. The argumentation framework is stored within an open graph database management system so that it can be used for graph retrieval-augmented generation (RAG) of UK political personas.

Paper Type: Short

Research Area: NLP Applications

Research Area Keywords: knowledge graphs, retrieval-augmented generation, LLM/AI agents, neurosymbolic approaches

Contribution Types: Data resources

Languages Studied: English

Submission Number: 3726