A Shared IPTC Topic Space for Comparable Topic Modelling of Social and Mainstream Media

Comparing topic attention between mainstream news and social media is often hindered by the fact that topic models fitted to different corpora produce incompatible, corpus-specific topic spaces. The authors seek to solve this by creating a reproducible, externally anchored framework that allows researchers to map disparate data sources into a single, standardized taxonomy for direct comparison.

The researchers utilize guided BERTopic, which allows for unsupervised topic discovery while using seed words to shape topic boundaries. They map discovered topics to the IPTC Media Topics taxonomy—an international standard for news classification. The mapping process is hierarchical: discovered topics are first scored against ninety-four level-1 IPTC topics using weighted keyword centroids and then collapsed into seventeen parent topics using a maximum-similarity rule. The framework was rigorously developed and selected using a New York Times 2011 corpus, employing a sequence of model screening, target-construction ablation, and threshold calibration to ensure the final configuration was both robust and not overly sensitive to specific parameters.

The study demonstrates that a guided BERTopic approach significantly outperforms zero-shot benchmarks in maintaining mapped coverage under strict assignment thresholds. Furthermore, the authors found that a parent-enriched target construction—which includes parent labels in the centroid calculation—improves both the coverage of mapped documents and the consistency of parent-level assignments. Finally, the researchers show that their mapping rule is not brittle; as assignment thresholds are tightened, coverage declines gradually rather than collapsing, indicating a stable and defensible configuration for cross-source analysis.

This framework provides a principled way to decouple the topic discovery process from the comparison space. By anchoring analysis in an external, validated taxonomy, researchers can compare topic attention across structurally different media (e.g., decentralized social media vs. editorially structured news) without the risk of manufacturing alignment through corpus-specific tuning. This approach is portable and can be adapted to any domain where sources can agree on a shared set of categories.