Researchers build open-source pipeline to map European political networks from news text

Scholars have long struggled to understand whether political elites operate as self-interested coalitions or collaborative governance networks—but tracking these informal relationships at scale has typically required painstaking manual work. A new research effort from computational social scientists offers a potential solution: an open-source system that automatically identifies political figures and their relationships by analyzing news articles in multiple languages.

The Technical Approach#

The pipeline combines several machine-learning techniques to transform raw text into structured knowledge. It starts with named-entity recognition to identify people, organizations, and institutions in news reports, then uses a three-step process to link these mentions to their standardized identifiers in Wikidata—ensuring the same person isn't counted multiple times under different name variations.

A specialized AI model, trained to understand political relationships, then extracts the nature of connections between figures: alliances, conflicts, or economic ties. The system uses guided decoding, a technique that constrains the model's output to follow a predefined ontology of valid relationship types. This keeps the model grounded in domain expertise rather than allowing it to invent relationships.

Testing and Validation#

When checked against 3,491 manually verified relationships, the pipeline achieved strong accuracy: 68.2% precision under strict evaluation, rising to 93.7% under more lenient metrics that allow minor variations in how relationships are described.

The researchers tested the system on two real-world cases. In Austria, it successfully reconstructed a political party's internal divisions, tracking when factions split off and following where key members ended up—including legal consequences. In Poland, the system mapped overlapping networks of state-enterprise patronage and identified the structural balance of conflicts between the Civic Platform and Law and Justice parties that have dominated recent Polish politics.

Why It Matters#

The framework is fully open-weight and multilingual, avoiding the limitations of proprietary commercial AI services that often work only in English and raise questions about data privacy. By moving beyond simple mention co-occurrence to identify actual, directed relationships—including whether ties are cooperative or antagonistic—researchers gain a tool for comparative politics across borders. The modular design means scholars can replicate and extend the work without reinventing core components.