Machine Learning Framework Automatically Maps Political Ad Topics Using AI

A new computational approach allows researchers to automatically organize large collections of political advertisements into meaningful topic categories without requiring manual labeling or predefined subjects.

How the System Works#

The method combines two key techniques: embedding-based clustering—which groups similar documents together in a high-dimensional space—and iterative large language model inference. The framework first identifies natural clusters within a text corpus, then uses LLMs to synthesize candidate topic labels for those clusters. These labels are subsequently refined through an iterative process to ensure consistency and accuracy across the entire taxonomy.

Application to 2024 U.S. Election Advertising#

Researchers tested the approach on political advertisements from the lead-up to the 2024 U.S. presidential election. Using the automatically generated topic taxonomy, they conducted downstream analyses examining issue prevalence, how moral framing appears in ads, spending patterns, and exposure differences across demographic groups.

Why This Matters#

Social media platforms have become central to modern political communication, but the sheer volume and constant flux of online advertising make traditional human-annotated analysis impractical at scale. This framework provides a scalable and interpretable alternative for researchers seeking to systematically examine unlabeled text corpora while maintaining transparency about topic categorization.

Research Foundation#

The work represents part of broader efforts in natural language processing to extract structured insights from unorganized data. By eliminating the need for seed topics or pre-existing labels, the method could extend to other domains beyond electoral advertising where rapid topic discovery is valuable.