Latent Topic Synthesis: Leveraging LLMs for Electoral Ad Analysis (arxiv.org)

Researchers introduce "Latent Topic Synthesis," an end-to-end framework that uses unsupervised clustering plus prompt-based LLM labeling to automatically construct an interpretable topic taxonomy from unlabeled text — no seed sets or domain expertise required. The method iteratively clusters documents, solicits semantic topic labels from a large language model, and augments topics with moral-framing annotations, producing both coarse and fine-grained, human-readable categories. The authors validate the approach with quantitative and qualitative analyses and release code/data and demos, showing the pipeline can scale to large corpora while preserving interpretability. Applied to a corpus of Meta political ads in the month before the 2024 U.S. Presidential election, the system uncovers latent discourse structures and actionable patterns: voting and immigration dominate spending and impressions, while abortion and election-integrity ads punch above their weight in reach. Funding and framing are polarized—economic appeals are driven by conservative PACs, abortion messaging splits between pro- and anti-rights coalitions, and crime-and-justice ads are locally fragmented. Moral-framing annotations reveal systematic correlations (e.g., abortion emphasizes liberty/oppression; economic messages mix care/harm, fairness/cheating, and liberty rhetoric), and demographic targeting signals emerge. The work demonstrates a scalable, interpretable ML tool for monitoring political messaging, with implications for media auditing, policy oversight, and ethical concerns around automated surveillance and labeling.