This paper presents a transparent and reproducible pipeline for modelling online political discourse through clustered opinion spaces derived from real Reddit post–comment pairs. Addressing the challenge of mapping increasingly fragmented online ideologies, we use large-scale orchestration of LLMs for ordinal feature scoring and k-medoids clustering with Gower distance to construct interpretable archetypes that capture ideological, moral, and stylistic variation in public debate. These clusters act as filters within retrieval-augmented generation (RAG), enabling synthetic comments grounded in authentic discourse rather than arbitrary persona assumptions. Evaluation using semantic similarity, linguistic naturalness, and a human line-up test demonstrates that generated comments are often indistinguishable from real ones while maintaining cluster-consistent rhetorical characteristics. Case studies on tariff debates and stock-market discussion illustrate the method’s flexibility in identifying specific discursive tropes within polarised communities. The pipeline offers a scalable approach for studying online discourse, providing interpretable synthetic data and a controlled framework for examining how discursive styles contribute to digital polarisation.

Authors:
Ian Drumm, The University of Salford, United Kingdom
Ser-Huang Poon, University of Manchester, United Kingdom

About the Presenter(s)
Dr Ian Drumm is Senior Lecturer at University of Salford specialising in AI, XR and software engineering, and has been PI and CI on Innovate UK, EPSRC and EU projects. Current projects include technology transfer and batch LLM orchestration.

Connect on Linkedin
https://www.linkedin.com/in/ian-drumm-4288302a/

Connect on ResearchGate
https://www.researchgate.net/profile/Ian-Drumm

See this presentation on the full schedule – Monday Schedule

Bookmark