New Shared Topic Space Method Aligns Media Corpora with IPTC Taxonomy

Comparing topic attention across different media sources has been hindered by a fundamental modeling problem: topic models fitted separately to each corpus produce incompatible topic spaces. A new paper from researchers including Din Iskakov and Rodrigo Wilkens, published on arXiv, introduces a reproducible framework that solves this by placing multiple corpora into a single shared topic space defined by the IPTC Media Topics taxonomy. The method uses guided BERTopic to score discovered topics against 94 level-1 IPTC topics, then collapses them to 17 parent categories using a maximum-similarity rule.

Developed and validated on a controlled New York Times 2011 corpus, the framework underwent a five-stage narrowing sequence: broad model screen, focused mapping refinement, strict finalist comparison, target-construction ablation, and threshold calibration. Key results show the guided family retaining substantially stronger mapped coverage than a zero-shot benchmark under stricter assignment thresholds, while a parent-enriched target construction improved both coverage and parent consistency. Coverage declined gradually rather than collapsing as the assignment threshold tightened, demonstrating robustness. This externally anchored method opens the door for reproducible, cross-source topic comparison in media analysis.